A knowledge structure enhanced subject graph embedding system and method

Abstract

This invention relates to a knowledge structure-enhanced subject graph embedding system and method. The system includes a processor, which comprises a data processing unit and a first computing unit. The data processing unit performs serialization representation on entities and relationships in the knowledge graph to generate an initialization vector sequence. The first computing unit receives the initialization vector sequence and performs a fusion operation of position embedding, word segmentation embedding, and structural information embedding in the semantic space to generate a knowledge chain embedding sequence. A dynamic weight sequence is calculated based on the causal logic of subject relationships to construct an information function representing the importance of the links. Through the fusion of serialization representation and multi-dimensional embedding, the system achieves accurate conversion of unstructured subject data into high-fidelity vector sequences. The dynamic weight mechanism based on causal logic enables the system to automatically capture the logical priority between subject knowledge, fundamentally solving the problems of chaotic system construction and inaccurate recommendations caused by neglecting knowledge dependencies in traditional models.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and artificial intelligence, and in particular to a subject graph embedding system and method for knowledge structure enhancement. Background Technology

[0002] Artificial intelligence (AI) technology is leading a profound transformation in the field of digital education, with knowledge graphs, as a crucial tool, gaining significant attention in areas such as personalized educational resource recommendation, knowledge tracking, and learning path generation. A knowledge graph is a structured representation of facts, composed of entities, relations, and semantic descriptions, expressed as knowledge triplets (head entity / subject, relation, tail entity / object). Subject-specific knowledge graphs integrate factual and conceptual knowledge from textbooks, learning materials, and expert advice into the knowledge graph, representing it in a structured form. This facilitates students' mastery of knowledge points, independent learning, and collaborative improvement of teaching and learning efficiency between teachers and students, making it indispensable in current AI-enabled education applications. With the rapid development of educational informatization and digitalization, the requirements for subject knowledge graphs are becoming increasingly stringent, while also exposing the following problems: Subject knowledge graphs contain many complex relationships (such as prerequisite / successor, sibling, hierarchical, and combinatorial relationships), whose dependencies are easily overlooked, leading to a chaotic knowledge structure and inaccurate knowledge recommendation. Furthermore, subject knowledge graphs contain knowledge chains composed of multiple triples, making it difficult to mine and identify positional and structural information, affecting the efficiency, completeness, and interpretability of learning path generation. Therefore, this paper proposes a subject graph representation learning method and system based on knowledge structure enhancement. This method fully utilizes the rich semantic information between subject knowledge, embedding positional and structural information of entities and complex relationships to improve the accuracy, completeness, and interpretability of the model.

[0003] Subject-specific knowledge graph representation learning achieves distributed embedding of entities and relations through dimensionality reduction vector spaces, thereby enabling reasoning prediction and completion of knowledge triples for missing elements. This significantly improves the computational accuracy, application completeness, and interpretability of subject-specific knowledge graph reasoning prediction. In recent years, research methods based on translation / topological geometry / neural network models, counterfactual reasoning methods, and attention mechanisms for knowledge recommendation have gained considerable attention. Knowledge graph embedding aims to learn entities and relations in a knowledge graph and transform them into vector form for easier semantic computation. It uses relational networks to perceive topological connections, embedding semantic relationships more clearly in the space. Counterfactual reasoning creates counterfactual links, learning node and edge representations from observed graph data, minimizing the variance of counterfactual predictions to improve reasoning accuracy. Attention mechanisms use bidirectional long short-term memory networks and attention mechanisms to extract entities and relations, with recommendation models evaluating the importance between triples to obtain prediction values.

[0004] However, the complex relationships between subject knowledge are crucial in knowledge graph representation learning. Using complex relationships as the core, combined with the embedding and encoding / decoding of structural and positional information, makes the results of knowledge chain embedding sequence prediction, personalized knowledge recommendation, and learning path generation more reasonable. However, research on this issue is still in its early stages and faces several challenges: 1) Existing subject knowledge graph representation learning methods simply represent knowledge as entities and relationships, making it difficult to learn complex relationships (such as: syntagmatic relationships, hierarchical relationships, acquisition / application relationships, and combination relationships), which are essential for building a knowledge system. 2) Multiple complex relationships also exist between triples. Ordinary methods struggle to identify knowledge chains composed of multiple triples and their interrelationships, leading to a lack of knowledge connectivity and system construction. In other words, locating knowledge along the relationship chain can be incorrect, and missing representations can result in the loss of key information, making it impossible to generate knowledge learning paths or leading to reasoning illusions, thus limiting the accuracy and completeness of learning path generation.

[0005] CN113568987A discloses a training method, apparatus, computer device, and storage medium for a knowledge graph embedding model. The method includes: acquiring a knowledge graph input graph as training samples; performing preliminary partitioning of the knowledge graph input graph to obtain two partitions; determining the cutting edge vertices based on the partitioning results; calculating the gain value of each cutting edge vertex based on key edges and general edges; when the gain value of a cutting edge vertex is greater than a preset value, determining that the cutting edge vertex is located on a critical path and is not in the same partition as the critical path, moving the cutting edge vertex from the original partition to the other partition to obtain the final partitioning result; parameterizing the final partitioning result to obtain a training sample set of entity embedding parameters and relation embedding parameters of the knowledge graph; and enabling worker nodes to train the knowledge graph embedding model based on the training sample set to obtain the trained knowledge graph embedding model. This technical solution focuses on graph partitioning optimization in a distributed training environment, aiming to improve training efficiency by calculating the gain value of cutting edge vertices. This strategy, which focuses on hardware resource scheduling and physical partitioning, does not endow the model with the ability to understand hierarchical or deductive logic. Moreover, this technical solution only treats the structure as partition boundaries and reduces communication overhead by moving vertices. This approach lacks awareness of the subject logic at the genetic level of embedding vector generation and cannot generate standardized inputs with topological fidelity.

[0006] CN115935968A discloses a knowledge graph embedding method based on semantic and relational structure fusion embedding. Step 1: Extract the "entity description dataset" and "relational structure dataset" from the knowledge graph; the "entity description dataset" originates from the descriptive attributes of entities; the "relational structure dataset" originates from entity relation triples (h, r, t), where h and t represent the head entity and tail entity, and r represents the relation type; Step 2: Train a word embedding model; perform word embedding training based on the "entity description dataset" to construct a word embedding model; the word embedding model stores the embedding vectors of words in the "entity description dataset"; Step 3: Entity pre-vector embedding; randomly select training data from the "relational structure dataset"; for each triple (h, r, t), obtain the embedding vectors from the "entity description dataset"... The steps are as follows: Step 4: Semantic Embedding; The head entity prevector pre_H_vector and the tail entity pre_T_vector are embedded into a head entity vector H_vector and a tail entity vector T_vector after passing through a semantic embedding network with the same structure and parameters; Step 5: Relational Structure Embedding; The head entity vector H_vector and the tail entity vector T_vector are input together with the relation vector R_vector into the relational structure model for optimization training, while optimizing the parameters of the head and tail entity vectors, relation vectors, and semantic embedding network to achieve joint training and fusion embedding of semantics and relational structure. This technical solution adopts a primary fusion of semantics (word2vec) and relational structure (TransE), and the core is still based on the triple translation model. It can only process single-step triples and cannot capture the features of complex knowledge chains composed of multiple triples. CN110837567A discloses a method and system for realizing knowledge graph embedding. This method includes: establishing a unified representation of a knowledge graph embedding model; constructing a structure search space for the unified representation; searching for a corresponding structure in the structure search space for a specific knowledge graph; training a knowledge graph embedding model based on the corresponding structure based on the specific knowledge graph; and obtaining the embedding representation of the specific knowledge graph using the trained knowledge graph embedding model. This technical solution focuses on automatically optimizing the general embedding model structure using search algorithms; its essence is architecture search. However, it lacks the ability to perceive the logic specific to a subject domain (causality, deduction) and does not involve the joint representation of multi-hop knowledge chains, thus failing to solve the reasoning illusion problem caused by long-distance dependencies.

[0007] This invention proposes a subject graph embedding system and method with enhanced knowledge structure, aiming to accurately represent subject knowledge and complex relationships. It also proposes a counterfactual link generation method to form a knowledge learning path based on students' weaknesses, ensuring the accuracy of the knowledge system while enhancing its completeness and interpretability.

[0008] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0009] This invention targets the field of basic education, addressing issues such as unclear identification of complex relationships, incomplete learning paths generated through reasoning, and sparse knowledge structure in subject knowledge graphs. It researches a representation learning method for subject knowledge graphs. Existing representation learning methods mostly only simply represent knowledge as entities and relationships, neglecting the learning of knowledge chains containing diverse and complex relationships such as hierarchical relationships, derivations, and prerequisites. This makes it difficult to accurately construct the subject knowledge system and affects the generation of learning paths when processing subject knowledge graph link prediction, thus hindering the realization of personalized basic education.

[0010] To address the shortcomings of existing technologies, this invention provides a knowledge structure-enhanced subject graph embedding system from a first aspect. The system includes a processor, which comprises a data processing unit and a first computing unit. The data processing unit performs serialization representation on entities and relationships in the knowledge graph to generate an initialization vector sequence. The first computing unit receives the initialization vector sequence and performs a fusion operation of position embedding, word segmentation embedding, and structural information embedding in the semantic space to generate a knowledge chain embedding sequence. Based on the causal logic of subject relationships, a dynamic weight sequence is calculated to construct an information function representing the importance of the links.

[0011] By fusing serialization representation with multidimensional embedding, this system achieves accurate transformation of unstructured subject data into high-fidelity vector sequences. Based on a dynamic weighting mechanism rooted in causal logic, the system automatically captures the logical priorities between subject knowledge, fundamentally solving the problems of chaotic system construction and inaccurate recommendations caused by traditional models neglecting knowledge dependencies.

[0012] According to a preferred embodiment, the processor further includes a second computing unit and a sampling training unit. The second computing unit is used to construct a structured recurrent neural network model based on the knowledge chain embedding sequence, generate aggregated endogenous variables and captured exogenous variables, and map the extracted feature vectors into an adjacency matrix with topological fidelity to determine fact links and counterfactual links; the sampling training unit is used to generate counterfactual links for fact links using a negative sampling strategy for iterative optimization, and combine aggregated endogenous variables and captured exogenous variables to construct a reliability function to quantify the interpretability of the model's inference results.

[0013] This technical solution introduces counterfactual reasoning and structured recursive networks to solve the black-box problem in knowledge path generation in deep learning. Through matrix mapping with topological fidelity and negative sampling iterative optimization, this structured recursive neural network can not only accurately determine factual links, but also quantify the interpretability of reasoning through a reliability function, providing a credible basis for personalized education decisions.

[0014] According to a preferred embodiment, the specific method by which the data processing unit performs serialization representation of entities and relations in the knowledge graph includes: recursively exploring the dependencies between subject knowledge entities using a depth-first search algorithm to complete the structured representation; converting the triples in the generated knowledge graph into an ordered knowledge sequence containing entities and relations, and mapping it to an initialization vector sequence.

[0015] By recursively exploring entity dependencies using the Depth-First Search (DFS) algorithm, a static graph is transformed into an ordered sequence containing logical evolutionary paths. The aim is to capture deep temporal relationships between knowledge points, providing standardized inputs with topological logic awareness for subsequent models, effectively improving the depth, efficiency, and completeness of data processing in complex disciplines.

[0016] According to a preferred embodiment, the specific method by which the first computing unit performs the fusion operation includes: performing a direct summation operation on the position embedding vector and the word segmentation embedding vector, and simultaneously concatenating the structural information embedding vector to generate a complex relationship representation that can characterize the relationship containing upper and lower positions or derivation logic; and realizing a refined representation from sequence to sequence through sequence decoding operation to provide semantic support for constructing link information.

[0017] By employing direct summation operations and synchronous concatenation, position, word segmentation, and topological structure are forcibly fused in the semantic space, accurately depicting abstract logic such as hyponyms and prepositions, and derivations within the discipline. The bidirectional encoder's sequence decoding operation supports refined representation from sequence to sequence, providing deep semantic support for the construction of multi-hop knowledge links and solving the problem of missing position awareness.

[0018] According to a preferred embodiment, the specific steps of the first computing unit in constructing the information function representing the importance of the link include: performing a product summation operation on the knowledge chain embedding sequence and the dynamic weight sequence calculated according to the causal logic of the subject relationship through linear combination to obtain the information function.

[0019] By constructing an information function through linear combination and product summation mechanisms, the problems of insufficient dependency and feature decay in long-range knowledge paths during transmission are solved. This technique can accurately locate key knowledge combinations and generate high-quality paths based on weighted aggregation of existing nodes, even in cold-start scenarios lacking historical data, significantly enhancing the system's robustness.

[0020] According to a preferred embodiment, the second computing unit in the processor runs the structured recurrent neural network model in the following ways: aggregating the semantic features of entities and the topological features of subject relationships layer by layer according to the arrangement order of each element in the knowledge chain to generate aggregated endogenous variables; using dynamic coupling mapping to filter covariate information outside the system boundary, and establishing a correlation model between external context vectors and endogenous representations to obtain captured exogenous variables.

[0021] This approach, by aggregating endogenous variables layer by layer, strengthens the native connections and semantic continuity between nodes, while the dynamic coupling mapping captures exogenous variables, effectively filtering out interfering information outside the system boundary. This collaborative modeling mechanism significantly improves the anti-interference ability of structured recurrent neural network models in complex teaching scenarios, ensuring the stability of the knowledge reasoning process and the accuracy of background dependency modeling.

[0022] According to a preferred embodiment, the second computing unit in the processor performs matrix mapping in the following ways: performing an extraction operation to extract feature vectors and combining the feature vectors to form a feature matrix; performing a computation operation to map the elements of the feature matrix to the adjacency matrix space using coupling relationships and outputting it to a multilayer perceptron to perform nonlinear mapping to complete link determination.

[0023] This method performs a mapping from the feature matrix to the adjacency matrix space, leveraging the coupling relationship and the nonlinear mapping capability of the multilayer perceptron to achieve accurate regression from the feature space to the graph structure. Its advantage lies in maintaining the topological fidelity of the hierarchical structure, ensuring the uniqueness and determinism of the output links, and providing strong mathematical support for complex link determination.

[0024] According to a preferred embodiment, the sampling training unit in the processor performs negative sampling in the following ways: generating corresponding virtual knowledge combinations for known fact links; modifying the logical features of subject relationship paths to simulate counterfactual links that do not conform to subject logic, for comparison and training with fact links.

[0025] This method generates counterfactual links through negative sampling for comparative training, simulating virtual combinations that do not conform to disciplinary logic. The aim is to enhance the model's ability to distinguish correct disciplinary knowledge paths. This technique continuously reduces loss through iterative optimization, effectively suppressing the inference illusion caused by long-distance dependencies and improving the model's prediction accuracy from a logical perspective.

[0026] The present invention provides a subject graph embedding method for knowledge structure enhancement from a second aspect. The method includes: performing serialization representation on entities and relations in the knowledge graph to generate an initialization vector sequence; receiving the initialization vector sequence and performing a fusion operation of position embedding, word segmentation embedding and structural information embedding in the semantic space to generate a knowledge chain embedding sequence; and calculating a dynamic weight sequence based on the causal logic of subject relations to construct an information function representing the importance of the link.

[0027] This method establishes a complete digital system from serialization representation and fusion embedding to causal weight calculation, explicitly transforming disciplinary logic into a computable mathematical model. Its role is to eliminate the problem of unclear identification of complex relationships in disciplinary maps, and to significantly improve the accuracy and completeness of learning path generation through structural enhancement techniques.

[0028] According to a preferred embodiment, the method further includes: constructing a structured recurrent neural network model based on the knowledge chain embedding sequence, generating aggregated endogenous variables and captured exogenous variables, mapping the extracted feature vectors into an adjacency matrix with topological fidelity to determine fact links and counterfactual links; using a negative sampling strategy to generate counterfactual links for iterative optimization of fact links, and combining aggregated endogenous variables and captured exogenous variables to construct a reliability function to quantify the interpretability of the model's reasoning results.

[0029] This method integrates structured neural network feature extraction with reliability function quantization, forming a closed-loop optimization structure with parameter feedback. Its technical advantage lies in its ability to continuously approximate the optimal parameter configuration through negative sampling iterations, while simultaneously quantifying the interpretability of the model's inference results, thus solving the problem of traditional search algorithm architectures lacking a measure of prediction reliability.

[0030] The present invention provides an electronic device from a third aspect, including a processor and a storage component, the storage component storing a computer program, wherein the processor implements the method of the present invention when running the computer program.

[0031] The present invention provides, from a fourth aspect, a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method of the present invention. Attached Figure Description

[0032] Figure 1 This is a logical diagram illustrating part of the information processing in the subject graph embedding system with enhanced knowledge structure provided by the present invention. Figure 2 This is a logical diagram illustrating another part of the information processing in the knowledge structure-enhanced subject graph embedding system provided by the present invention; Figure 3 This is a complete logical schematic diagram of the subject graph embedding system with enhanced knowledge structure provided by the present invention; Figure 4 This is a flowchart illustrating the subject graph embedding method for knowledge structure enhancement provided by the present invention.

[0033] List of reference numerals 101: Text Data; 102: Knowledge Graph; 103: Knowledge Sequence; 104: Structured Representation; 105: Serialization Representation; 106: Knowledge Input; 107: Position Embedding; 108: Word Segmentation Embedding; 109: Complex Relation Representation; 110: Structural Information Embedding; 111: Sequence Decoding Operation; 112: Linear Combination; 113: Knowledge Chain Embedding Sequence; 114: Weight Sequence; 115: Information Function; 116: Structured Recurrent Neural Network Model; 117: Aggregating Endogenous Variables; 118: Capturing Exogenous Variables; 119: Feature Vector; 120: Feature Matrix; 121: Adjacency Matrix; 122: Extraction operation; 123: Computation operation; 124: Multilayer perceptron; 125: Decoding operation; 126: Determining link operation; 127: Fact link; 128: Counterfact link; 129: Scoring function; 130: Electronic device; 131: Processor; 132: Storage component; 133: Data processing unit; 134: First computing unit; 135: Second computing unit; 136: Sampling training unit; 137: First data input port; 138: First data transmission port; 139: Second data transmission port; 140: Third data transmission port; 141: Second data input port. Detailed Implementation

[0034] The following is a detailed explanation with reference to the accompanying drawings.

[0035] This invention provides explanations and clarifications for some terms and concepts.

[0036] Text Data 101: The original form of subject data is presented as text, so the initial subject data is collectively referred to as Text Data 101, which is the input source of the entire system.

[0037] Knowledge Graph 102: refers to the use of a graph structure containing nodes (entities) and edges (relationships) to represent the entities and relationships in text data 101, as a structured form of knowledge representation.

[0038] Knowledge Sequence 103: Arrange knowledge such as entities (nodes) and relations (edges) into an ordered set, that is, represent it as a sequence, such as entity 1, entity 2, ..., entity n; relation 1, relation 2, ..., relation n.

[0039] Structured representation 104: The process of transforming messy text data 101 into a graph structure with logical topology (including nodes and edges).

[0040] Serialization Representation 105: The process of converting triples in graph data into a sequence of initialization vectors using the depth-first search (DFS) algorithm, so as to facilitate the input of subsequent deep learning models.

[0041] Knowledge Input 106: Connection section, used to input the initialization vector sequence generated by S100 into the bidirectional encoder.

[0042] Position Embedding 107: Injects the order information of a word in a sentence, encodes the position information of words in a sentence, that is, injects the absolute or relative order information of a word in a sequence, and solves the problem that the Transformer architecture itself cannot perceive position.

[0043] Word segmentation embedding 108: captures the semantic meaning of words themselves and transforms subject knowledge points into vectors in the semantic space.

[0044] The direct concatenation operation represents a vector concatenation operation used to fuse position vectors, word segmentation vectors, and structural information.

[0045] Complex Relationship Representation 109: In the semantic space, vectors that can accurately characterize complex disciplinary relationships such as "hyperposition" and "inference" are generated by combining embedded position, word segmentation and structural information.

[0046] Structural information embedding 110: The topological structure of knowledge is represented in the form of vectors, along with positional information, word segmentation, etc., in a semantic space with reduced dimensionality, so that the structural and semantic associations can be effectively computed in the numerical space.

[0047] Sequence decoding operation 111: The decoder of the BERT model is used to refine the encoded sequence representation, providing support for the generation task.

[0048] Linear combination 112: Set a linear weighted combination of the knowledge chain embedding sequence 113 to characterize the overall features of multiple links.

[0049] Knowledge chain embedding sequence 113: refers to a complete link vector sequence containing multiple triples and the associations (multi-hop relationships) between triples.

[0050] Weight Sequence 114: A dynamic set of weights calculated based on the categorical characteristics of complex relationships (such as importance and causal logic). .

[0051] Information function 115: A function obtained by multiplying and summing the linear combination 112 represented by the knowledge chain and the weight sequence 114, used to accurately locate the most critical knowledge combination.

[0052] Structured recurrent neural network model 116: By recursively computing the entity and relation features in the knowledge chain layer by layer, and explicitly embedding the knowledge chain embedding sequence 113, context and topology based on the weight sequence 114, the semantic continuity and structured features of the knowledge chain representation are systematically enhanced.

[0053] Aggregate endogenous variables 117: This aims to integrate explicit embeddings and topological features in the recursive process, strengthen the relationship characteristics between nodes, generate hierarchical semantic association entity vector representations, and provide structured semantic support for subsequent operations.

[0054] Capturing exogenous variables 118: Filtering and simulating covariate information outside the system boundary, using the correlation modeling between external context vectors and endogenous representations, and enhancing the model's accuracy in modeling knowledge background dependencies and improving its resistance to interference from external information through dynamic coupling mapping.

[0055] Feature matrix 120: A matrix composed of feature vectors 119 extracted from the link.

[0056] Adjacency matrix 121: Maps the elements of feature matrix 120 to a more knowledge-based (relational) computable space, ensuring that matrix computation has gradient optimization capabilities and maintaining the topological fidelity of the hierarchical structure.

[0057] Multilayer Perceptron 124: Through continuous mapping of high-dimensional feature space, it constructs an encoding-hidden-decoding pipeline by utilizing the gradient properties of activation functions, and completes the classification and embedding of links by using nonlinear mapping.

[0058] Link determinism: In this invention, it is used to determine the existence of a link and to classify factual links 127 and counterfactual links 128.

[0059] Fact Link 127: A knowledge combination that conforms to the correctness of data facts, including correct entities and relationships.

[0060] Counterfactual Link 128: Based on known links, counterfactual questions are proposed, and virtual / error combinations generated by negative sampling are used for comparative training to enhance the determinism of model inference.

[0061] Score function 129: The final quantification function that measures the accuracy and reliability of the output (such as the predicted learning path).

[0062] Negative sampling training strategy: This is a method for approximating gradient optimization by randomly sampling with an adaptive logarithmic frequency distribution. It uses adaptive sampling weights to approximately maximize the cross-entropy likelihood function, thereby alleviating training imbalance and finding the optimal model configuration.

[0063] Reliability function (Lc): A function constructed by combining endogenous variables, exogenous variables, and loss intensity, used to quantify the interpretability of model inference results.

[0064] Unidirectionality and Reversibility: In this invention, the unidirectionality of a relation can be represented as a directed relation, such as the relation... Its reversibility can be specifically expressed as .

[0065] Loss function: It is the basis for updating model parameters. By adaptively sampling weights to optimize parameters, it ensures that gradient descent achieves the best model configuration.

[0066] Example 1 This invention can be applied in the education industry, specifically in areas such as subject data management, subject knowledge reasoning and prediction, and optimization of subject knowledge systems.

[0067] This invention provides a knowledge structure-enhanced subject graph embedding method, system, and reasoning model. This invention also provides an electronic device 130 that enables the knowledge structure-enhanced subject graph embedding method of this invention.

[0068] The electronic device 130 of the present invention refers to an electronic device 130 capable of running a subject graph embedding program that enhances knowledge structure, such as a server, computer, mobile computer, mobile tablet computer, dedicated processor, etc.

[0069] The electronic device 130 of the present invention includes a processor 131 and a storage component 132. The processor 131 can run encoded information of a knowledge structure-enhanced subject graph embedding method stored by the storage component 132.

[0070] The processor 131 in this invention can also be a microprocessor, a dedicated integrated chip, other processing elements with processing capabilities, or other electronic components capable of running the encoded information of the subject graph embedding model.

[0071] The subject map embedding system of the present invention may further include a data processing unit 133, a first calculation unit 134, a second calculation unit 135, and a sampling training unit 136.

[0072] The data processing unit 133 is a processor or dedicated integrated chip capable of constructing knowledge extraction and deep search models for raw subject data.

[0073] The first computing unit 134 is a structure-enhanced embedded processor, which is a processor or dedicated integrated chip capable of performing complex relational, structural, and positional information representations of the processor and realizing knowledge sequence encoding information.

[0074] The second computing unit 135 is a counterfactual link learner, which is a processor or dedicated chip capable of building counterfactual reasoning and structured recurrent neural network models 116 based on vector sequences of knowledge chains.

[0075] The sampling training unit 136 is a processor or dedicated integrated chip capable of performing a simulated subject map embedding training process and executing a negative sampling strategy.

[0076] The data processing unit 133 and the first computing unit 134 establish a data transmission relationship through at least one first data transmission port 138. The data processing unit 133 is also provided with a first data input port 137 for receiving text data 101, structured data, etc.

[0077] The first computing unit 134 and the second computing unit 135 establish a data transmission relationship through at least one second data transmission port 139 to send data such as structural location information embedding, hyperparameters, and knowledge chain embedding sequence 113 to the second computing unit 135. Both computing units have the ability to transmit vector sequences. Vector sequences include, for example, entity vectors, relation vectors, location vectors, and knowledge sequence 103, etc.

[0078] The second computing unit 135 and the sampling training unit 136 establish a data transmission relationship through at least one third data transmission port 140. The sampling training unit 136 is also provided with a second data input port 141 to receive sampling data, scoring function hyperparameters, etc. from the second computing unit 135.

[0079] The subject graph embedding system also includes a storage component 132, which is used to store at least the data, knowledge, models, functions, and other information related to the subject graph embedding process. The storage module may be a storage medium such as DRAM, SRAM, disk, or Flash chip. The storage component 132 establishes a data transmission relationship with the data processing unit 133 through at least one first data transmission port 138, enabling the data processing unit 133 to retrieve the required text data 101 and structured data from the storage component 132. The storage component 132 establishes a data transmission relationship with the first computing unit 134 through at least one first data transmission port 138, enabling the first computing unit 134 to retrieve the BERT model, information function 115, etc., from the storage component 132 for actual subject graph embedding. The storage component 132 establishes a data transmission relationship with the second computing unit 135 through at least one second data transmission port 139, enabling the second computing unit 135 to retrieve the neural network model, scoring function 129, and reliability function, etc., from the storage component 132 for model training, optimization, and prediction.

[0080] The stored information in storage component 132 includes functions such as a subject graph embedding score function 129, a loss function, an information function 115, a state function, and a reliability function. Data includes text data 101, structured data, and hyperparameters. Knowledge includes subject knowledge, entities, and relationships. Models include BERT models and neural network models.

[0081] The data processing unit 133 performs structured knowledge extraction on text data 101 and connects to the first computing unit 134 with structure-enhanced embedding capabilities through at least one first data transmission port 138. This enables the first computing unit 134 to establish data transmission coupling with the data processing unit 133 when performing serialization transformation on entity and relation representations. The data processing unit 133 has knowledge extraction and deep search models. When it receives the input text data 101 through the first data input port 137, it is in the logical flow of constructing a knowledge graph 102. This allows the first computing unit 134 to receive the initialization vector sequence generated by the depth-first search algorithm and embed the location and structural information together with the BERT model called from the storage component 132. This establishes a cascaded communication connection with the second computing unit 135, which has counterfactual link learning capabilities, based on the knowledge chain embedding sequence 113. This achieves refined sequence-to-sequence representation and supports the calculation of the scoring function 129.

[0082] The second computing unit 135 has a structured recurrent neural network model 116, which is connected to the sampling training unit 136 by receiving the knowledge chain embedding sequence 113 output by the first computing unit 134 through the second data transmission port 139. This allows the second computing unit 135 to establish a dynamic coupling mapping relationship with the captured exogenous variable 118 when the aggregated endogenous variable 117 is in the state of estimation threshold effect. This allows the second computing unit 135 to be fixed to the logical part of generating fact link 127 and counterfactual link 128. The second computing unit 135, the sampling training unit 136 and the storage component 132 constitute a closed-loop optimization structure. This closed-loop optimization structure has the characteristic of dynamically adjusting the parameter strategy through the negative sampling model training method. When performing link prediction for the subject knowledge graph 102, the sampling training unit 136 communicates with the second computing unit 135 according to the inverse correlation between the reliability function and the loss function. This allows the optimal model parameters to be found while continuously reducing the loss through iterative training, providing support for the subsequent reasoning tasks of the subject knowledge graph 102.

[0083] Example 2 This embodiment is a further improvement on embodiment 1, and repeated content will not be described again.

[0084] The knowledge structure-enhanced subject graph embedding method in this invention, such as... Figure 4 As shown.

[0085] S100: By using named entity recognition and depth-first search (DFS) algorithm, the original subject text is mapped into an initialized structured vector sequence with topological logic.

[0086] S200: Based on the BERT architecture, it performs direct summation operations on position, word segmentation and structural information in the semantic space to generate complex relation representation vectors that accurately characterize the logic of disciplines.

[0087] S300: Information function 115 is constructed by using a linear combination 112 of multi-hop knowledge chain and a dynamic weight sequence 114 to solve the problem of insufficient dependency of long-range knowledge path in the embedding process.

[0088] S400: Constructs a structured recurrent neural network, which enhances the semantic continuity of knowledge links and the ability to resist background interference through dynamic coupling mapping of endogenous and exogenous variables.

[0089] S500: Extract the feature matrix 120 and map it into an adjacency matrix 121 with topological fidelity. The classification and embedding of complex links are completed through the nonlinear mapping pipeline of the multilayer perceptron 124.

[0090] S600: Based on counterfactual reasoning, it generates comparative links, combines negative sampling strategies and reliability functions, and iteratively optimizes model parameters to achieve interpretable subject knowledge prediction.

[0091] The knowledge structure-enhanced subject graph embedding method in this invention, such as... Figure 1 and Figure 3 As shown, the first part of the processing steps includes: S100: Graph data augmentation is performed based on subject text data 101 and structured data to obtain an initial vector sequence of vector sequences such as entities, relations, and knowledge chains; S200: The bidirectional encoder and decoder based on the BERT model analyze the initialization vector sequence, and combine position and structure embedding to enable the complex relationships between knowledge entities and multiple triples to be structurally represented 104, and obtain refined sequence representations such as complex relationship weights and knowledge structure positions. S300: Based on the knowledge sequence 103 obtained by the sequence decoding operation 111, the knowledge chain embedding sequence 113 is integrated, the linear combination of knowledge chain representations 112 is set, and the weight sequence 114 is calculated according to the characteristics of complex relationships to enhance the link-level information, thus obtaining the information function 115.

[0092] The present invention provides a detailed description of each of the above steps.

[0093] The detailed steps of step S100 are as follows.

[0094] For subject-specific text data 101 and structured data, data cleaning and multi-source heterogeneous alignment techniques are used to enhance all initial data through feature completion and semantic calibration. Entities and relations are extracted from text data 101 using techniques such as named entity recognition and relation extraction, and stored as graph data in the form of a knowledge graph 102 containing nodes and edges.

[0095] Specifically, named entity recognition adopts a joint architecture based on pre-trained language models (such as BERT) and bidirectional long short-term memory network-conditional random field (BiLSTM-CRF). It aims to provide basic nodes for the construction of knowledge graph 102 by accurately locating professional entities such as concept words, terminology words and formula symbols in subject texts, and combining subject field encyclopedias to perform entity disambiguation and normalization mapping.

[0096] Based on entity recognition, semantic role labeling or attention mechanisms are used to extract deep semantic relationships or attribute features between entities from text data 101. Structured data is introduced as a priori constraints for alignment verification. The data is expressed in the form of standardized triples (head entity, relation, tail entity) and finally forms graph data stored in the form of a knowledge graph 102 containing nodes and edges.

[0097] The knowledge extraction process involves extracting semantic relationships or attribute features between entities from text data 101 and expressing them in a structured form (such as triples) to form a knowledge graph 102. Each triple is defined as... ,in and These represent the head entity and the tail entity, respectively.

[0098] To capture the temporal and logical evolution paths between knowledge points, the triples in knowledge graph 102 are converted into knowledge chains. ,in, Each triple can be represented as a pair of triplets, and the relationships between them can be expressed as follows: , representing the first in the chain The triplet and the first The connection between the three triplets.

[0099] Based on the link representation results, a knowledge chain embedding sequence 113 is generated using a depth-first search algorithm, and the generated initialization vector sequence is as follows: This serves as the input for the next module. Specifically, this part transforms the text data 101 into a structured sequence of initialized text vectors.

[0100] Depth-first search algorithms refer to the recursive exploration of relationships between knowledge entities in a subject, prioritizing the exploration of deeper connections along the current path to construct knowledge chains or knowledge chain embedding sequences from the starting point to the target node.

[0101] The detailed steps of step S200 are as follows.

[0102] First, based on the encoding and decoding parts of the BERT model, the structured and serialized vectors obtained in step S100 are used as knowledge input 106 to represent the structural and positional information of the knowledge sequence 103.

[0103] Suppose that the given knowledge chain to be completed is represented as The knowledge pairs to be completed are represented as follows: The input to the BERT model is the previous one. The vector sequence corresponding to each link and By embedding corresponding contextual position information and combining it with word segmentation embedding 108, structural information can be represented.

[0104] For example This corresponds to the position embedding 107 of the knowledge sequence vector 103. The knowledge sequence vector 103 is represented as... , It corresponds to the word segmentation embedding sequence. One of the elements, It represents the complex relationship between corresponding knowledge triples 109.

[0105] Set the first Three pairs The weighted terms are , It is its modulo operation. These are trainable parameters. It's a sigmoid activation function, and the hyperparameter bias term settings are... . above satisfy .

[0106] The above process can represent the structured embedding of a knowledge chain. For example, selecting a triple in the knowledge chain... Its structured embedding can be represented as: .

[0107] The control coefficient for the range of values ​​in a continuous vector space is expressed as follows: , which are also trainable parameters. The probability of the relation in each triplet is expressed as: The state function is expressed as To enhance weight adaptability, and This is a hyperparameter.

[0108] To characterize the importance of complex relationships during model inference, relation weights are used. The expression is: .

[0109] The detailed steps of step S300 are as follows.

[0110] set up Linear combination 112 of knowledge chain embedding sequence 113, and The weight sequence is 114. The two sequences represent the common supporting information function 115.

[0111] By combining the weight sequence 114 to enhance link-level information, an information function 115 for complex relationships is constructed.

[0112] Information function 115 is represented as: .

[0113] In the above formula, This indicates the knowledge chain embedding sequence 113; It can be used as an index for summation, and also as a positional coefficient; This represents the i-th element in the weight sequence, used to characterize the importance or category characteristics of a specific relationship; Represents the knowledge chain embedding sequence The transpose of . It refers to a complete link vector sequence that contains multiple triples and the associations between triples (multi-hop relationships).

[0114] This differs from previous knowledge graph embedding methods. Specifically, traditional knowledge graph embedding methods (such as TransE / DistMult / BERT) focus on representing single-hop relationships or local subgraphs. While some methods (such as RotatE) focus on embedding complex relationships, they do not address the specific scenarios in subject-specific graphs, where complex relationships vary significantly. This invention, however, mines explicit linear combinations 112 of knowledge chains (i.e., multi-hop links), representing knowledge sequences 103 such as entities / relationships on the links. It introduces dynamic weight sequences 114 to assign importance to different knowledge chains, thereby decomposing the completeness of the information function 115 into an adjustable calculation of the "product of knowledge chain structure and weight sequence 114." This solves the problems of insufficient reliance on long-range knowledge chains and incomplete links in traditional methods. In scenarios requiring multi-step logical coherence, such as learning path generation and knowledge recommendation in the education field, the system can accurately locate the most critical knowledge combinations by displaying path weights. Furthermore, since the linear combination of knowledge chains 112 can decompose the representation of long-tail / sparse paths into a weighted aggregation of existing nodes, rather than relying entirely on the historical data of new paths, the resilience of cold start scenarios is improved, and the emergence of paths can enhance the interpretability of model decisions.

[0115] The counterfactual reasoning link generation module in this invention, namely the second computing unit 135, is as follows: Figure 2 and Figure 4 As shown, the subsequent implementation steps include: S400: Construct a structured recurrent neural network model 116, input a knowledge chain embedding sequence 113 with structural information, further generate aggregated endogenous variables 117, and obtain captured exogenous variables 118 based on this, so as to extract information from the link context more accurately and completely.

[0116] S500: Extract feature vector 119, then construct feature matrix 120 through extraction operation 122, perform calculation operation 123 on feature matrix 120 to obtain adjacency matrix 121, input matrix calculation, sequence and other results into multilayer perceptron 124, and then encode and decode by multilayer perceptron 124 to complete link classification.

[0117] S600: After decoding operation 125, perform link determination operation 126 to determine the existence of the link and classify it into fact links 127 and counterfactual links 128. Then, derive the scoring function 129 to measure the accuracy of the output result.

[0118] The detailed steps of step S400 are as follows.

[0119] S410: Based on the refined knowledge sequence 103 representation, a structured recurrent neural network model 116 is constructed to generate aggregated endogenous variables 117.

[0120] By using a recursive computation mechanism to aggregate entity and relation features in the knowledge chain layer by layer, the computation process is completed in the first... The weights of each embedding are represented as , It is the first 107 are embedded in each position. It is the first Each word segmentation embedding has 108 elements. It is the first A relational embedding, It is the first One bias term.

[0121] This step can strengthen the tightness of neural network connections and systematically enhance the semantic continuity and structured features of knowledge chain representation.

[0122] The aggregate endogenous variable 117 is represented as: .

[0123] S420: Based on the aggregation of endogenous variables 117, the estimated threshold is aggregated to generate captured exogenous variables 118, which can accurately extract the contextual information of the link and increase completeness.

[0124] By using the association model between external context vectors and aggregated endogenous variables 117, and through dynamic coupling mapping, the model's accuracy in modeling knowledge context dependencies is enhanced and its ability to resist interference from external information is improved.

[0125] Preferably, set For trainable parameters, It is the first and the The bias difference of the term.

[0126] The specific representation of the captured exogenous variable 118 is as follows: .

[0127] The detailed steps of step S500 are as follows.

[0128] S510: Extract the feature matrix 120 and adjacency matrix 121 from the input knowledge chain and its sequence, denoted as X and A, respectively. Preferably, extract feature vector 119 from the links. The feature vectors 119 are combined to form the feature matrix 120, denoted as X.

[0129] The adjacency matrix 121 maps the elements of the feature matrix 120 to a more complex, knowledge-based (relational) computational space, specifically used for inputting topological information, node attributes, and category labels. To ensure gradient optimization capabilities in matrix computation and maintain the topological fidelity of the hierarchical structure, the calculated adjacency matrix 121 is denoted as A. The adjacency matrix 121 provides embedding support for the input to the multilayer perceptron 124.

[0130] The detailed steps of step S600 are as follows.

[0131] S610: The features of node pairs in the knowledge chain are jointly embedded by the multilayer perceptron 124. Using calculation methods such as feature concatenation, inner product and nonlinear mapping, a tensor representation of the feature combination of all node pairs is formed, and an adjacency matrix 121 corresponding to the result is formed.

[0132] For example, the adjacency matrix 121 pointing to the factual result and the adjacency matrix 121 pointing to the counterfactual result are respectively: ; .

[0133] In the above formula, and They are respectively The first of the codes row and number Column vectors For trainable parameters, To control for the error coefficient, F represents the counterfactual.

[0134] S620: Using the softmax activation function in conjunction with the information function Ternary structured characterization Trainable parameters Control error coefficient Fact Link Representation and counterfactual link representation Derive the scoring function 129.

[0135] The scoring function 129 is expressed as: .

[0136] S630: Employs a negative sampling model training method to find the optimal parameters to optimize the parameters of the subject knowledge graph 102 representation learning model, thereby improving the performance and interpretability of tasks such as link prediction, learning path generation, and knowledge recommendation.

[0137] After a random sampling process, the original rules (correct rules) for the path are set as follows: Leveraging the unidirectional and reversible nature of paths, the negative sampling logic is modified to... .

[0138] The scoring function loss for quantifying differences in knowledge context information and the loss for counterfactual outcome estimation.

[0139] and The loss expression: ; .

[0140] In the above formula, To capture the linear combination 112 of the knowledge chain (and) (The elements in the middle represent consistency), and the aggregated endogenous variable 117 is... The exogenous variable 118 is , and For trainable parameters, and For bias terms, This is the path-aware weight vector.

[0141] Preferably, set To estimate the loss function for the counterfactual results of the control link prediction score, To control the loss coefficient of reliability error, To provide a trainable, dynamically supplementary loss to balance the reliability of the loss function.

[0142] loss function The expression is: .

[0143] S640: To enhance the interpretability of the scoring function 129 and the model inference results, the complex relationship scoring loss based on differences in episodic and structural information is combined with the embedding strength loss, along with endogenous variable aggregation and exogenous variable capture sequence representation, and trainable parameters are set. and bias terms .

[0144] In the training loss function As the value of the reliability function decreases, the reliability function becomes increasingly larger. That is, higher interpretability.

[0145] The reliability function expression is: .

[0146] Example 3 like Figure 3 As shown, the hardware foundation of this invention consists of a data processing unit 133 and a first computing unit 134. The data processing unit 133 serves as the execution entity for data entry, and its output is connected to the logical architecture of the knowledge graph 102 formed by internal storage. The first computing unit 134 serves as the core processing hub, and it integrates and connects a bidirectional encoder based on the BERT model. This bidirectional encoder couples the computing modules of position embedding 107, word segmentation embedding 108, and structural information embedding 110 at the circuit and logic levels.

[0147] like Figure 3As shown, the signal output port of the first computing unit 134 is connected to the input port of the second computing unit 135, forming a cascaded relationship. The second computing unit 135 internally houses a structured recurrent neural network model 116, and is further connected to the hardware path of the multilayer perceptron 124 via a data bus. This perceptron employs an encoding-hidden layer-decoding pipeline structure, ultimately connecting to the computing circuit responsible for the decoding operation 125, and forming a feedback loop with the computing register of the scoring function 129.

[0148] S101: Data preprocessing and knowledge sequence generation.

[0149] like Figure 1 and Figure 3 As shown, the data processing unit 133 in the physical hardware device 130 receives the raw text data 101 through the first data input port 137. The data processing unit 133 first performs structured representation 104 using Named Entity Recognition (NER) and relation extraction methods, transforming the unstructured subject text into a knowledge graph 102 containing nodes (entities) and edges (relationships). Subsequently, the data processing unit 133 performs serialization representation 105 using a depth-first search (DFS) algorithm, recursively exploring the deep dependencies between subject knowledge entities, transforming the triples in the graph into an ordered knowledge sequence 103 containing entities and relations, and mapping it as an initialization vector sequence stored in the storage component 132. Through this connection, the efficient computation of the hardware unit transforms the messy data into a standardized input with topological logic, significantly improving the accuracy of the conversion from raw data to serialized data.

[0150] Specifically, the data processing unit 133 first performs feature completion and semantic calibration based on the subject text data 101 and the structured data in the storage component 132 using data cleaning and multi-source heterogeneous alignment techniques. The data processing unit 133 utilizes an integrated Named Entity Recognition (NER) module, employing a joint architecture based on a pre-trained language model (BERT) and a bidirectional long short-term memory network-conditional random field (BiLSTM-CRF), to accurately locate concept words, terminology, and formula symbols in the subject text. It then combines this with subject-specific encyclopedia entries for entity disambiguation and normalization mapping, constructing a knowledge graph 102 in the form of standardized triples (h, r, t).

[0151] Subsequently, the data processing unit 133 extracts deep semantic relationships between entities from the text using semantic role labeling or attention mechanisms, and executes a depth-first search (DFS) algorithm to recursively mine deep dependencies between subject knowledge entities. This process represents the logical connection between the i-th triplet and the (i+1)-th triplet in the chain as r. i, i+1 Finally, the data is merged to generate an initialization vector sequence containing entities, relationships, and knowledge chains. .

[0152] This process transforms unstructured text into standardized input with topological fidelity through efficient logic processing of hardware units, providing a solid input foundation for subsequent deep learning models.

[0153] S102: Structured bidirectional coding and feature fusion.

[0154] like Figure 1 and Figure 3 As shown, after receiving the initialization vector sequence, the first computing unit 134 imports it as knowledge input 106 into a bidirectional encoder based on the BERT architecture. The first computing unit 134 performs multi-dimensional embedding in the semantic space: it performs a direct summation operation on the positional embedding 107 (injected with sequence information) and the word segmentation embedding 108 (capturing lexical semantics), and simultaneously concatenates the structural information embedding 110 to accurately generate a complex relational representation 109R capable of depicting hierarchical relationships, inferences, and other logical connections. m .

[0155] Specifically, the single triplet generated by the first computing unit 134 The structured embedding representation is as follows: .

[0156] Among them, E m For position embedding, E m 'For word segmentation embedding, R m For relational representation, μ is a trainable parameter. m This is the bias term. Subsequently, the BERT decoder performs sequence decoding operation 111 to achieve a refined sequence-to-sequence representation.

[0157] This forces the integration of subject-specific topology into the numerical space, enabling the hardware model to handle abstract logic such as hierarchical and superordinate elements, and significantly improving the accuracy of subject knowledge representation.

[0158] S103: Link-level enhancement and weighted calculation.

[0159] like Figure 1 and Figure 3 As shown, the first computing unit 134 integrates the decoded knowledge sequence to generate a knowledge chain embedding sequence 113 containing multiple triples and multi-hop relationships. Simultaneously, the system calculates a dynamic weight sequence 114 using a state function based on the category characteristics of complex relationships (such as causal logic or importance). The first computing unit 134 sets a linear combination 112 to perform product summation on the knowledge chain embedding sequence 113 and the weight sequence 114, constructing a link-level information function 115.

[0160] Relationship weight The expression is: .

[0161] In the above formula, K(·) is the state function used to enhance the adaptability of the weights. The system sets up a linear combination 112, which constructs the information function 115 by performing a product summation between the knowledge chain embedding sequence 113 and the weight sequence 114.

[0162] Information function 115 is represented as: .

[0163] In the above formula, This indicates the knowledge chain embedding sequence 113; It can be used as an index for summation, and also as a positional coefficient; This represents the i-th element in the weight sequence, used to characterize the importance or category characteristics of a specific relationship; Represents the knowledge chain embedding sequence The transpose of . It refers to a complete link vector sequence that contains multiple triples and the associations between triples (multi-hop relationships).

[0164] This step accurately locates key knowledge combinations through explicit path weights, effectively compensating for the insufficient dependence on long-range knowledge chains in traditional embedding methods. By using linear combination 112 and a product summation mechanism, the decay problem of long-range knowledge chains is compensated for, enhancing the system's ability to identify key knowledge paths.

[0165] S104: Feature extraction from recurrent neural networks.

[0166] like Figure 2 and Figure 4 As shown, the second computational unit 135, acting as a counterfactual link learner, inputs the received link sequence into the structured recurrent neural network model 116. This structured recurrent neural network model 116 aggregates features layer by layer through a recursive computation mechanism, generating aggregated endogenous variables 117 to enhance the inter-node relationship characteristics.

[0167] .

[0168] Simultaneously, the second computing unit 135 uses dynamic coupling mapping to capture exogenous variables 118, filters covariate information outside the system boundary, and establishes dynamic coupling mapping.

[0169] The specific representation of the captured exogenous variable 118 is as follows: .

[0170] The recursive mechanism strengthens the original connections and semantic continuity between nodes, and the collaborative modeling of endogenous and exogenous variables ensures the modeling accuracy of knowledge background information and improves the model's ability to resist interference from external information.

[0171] S105: Matrix mapping and nonlinear classification.

[0172] like Figure 2 and Figure 4 As shown, the second computation unit 135 performs extraction operation 122 to extract feature vector 119 and construct feature matrix 120. Subsequently, through computation operation 123, a dynamic evolution mechanism is simulated using coupled differential equations to map feature matrix 120 into adjacency matrix 121 with topological fidelity. This feature matrix 120 is then fed into a multilayer perceptron 124, where its encoding-hidden-decoding pipeline structure performs nonlinear mapping to complete the classification of complex links.

[0173] For example, the adjacency matrix 121 pointing to the factual result and the adjacency matrix 121 pointing to the counterfactual result are respectively: ; .

[0174] In the above formula, and They are respectively The first of the codes row and number Column vectors For trainable parameters, To control for the error coefficient, F represents the counterfactual.

[0175] This process ensures the determinism of the output link and maintains the topological fidelity of the hierarchical structure by utilizing the high-dimensional mapping capability of the perceptron.

[0176] S106: Counterfactual Reasoning and Scoring.

[0177] like Figure 2 and Figure 4 As shown, the second computing unit 135 performs a link determination operation 126 via a decoding operation 125, and outputs a comparison result between the factual link 127 and the counterfactual link 128 by proposing a counterfactual question based on the known link. Finally, the second computing unit 135 jointly derives a scoring function 129 to quantify the accuracy of the path.

[0178] The scoring function 129 is expressed as: .

[0179] Meanwhile, the sampling training unit 136 applies a negative sampling strategy to calculate the loss function based on the unidirectionality and reversal of complex relationship paths.

[0180] loss function The expression is: .

[0181] More preferably, the loss function The expression is: .

[0182] in, and Control the score loss and reliability error loss separately. To dynamically replenish losses.

[0183] The model parameters are iteratively optimized by calculating the loss from the path score function and the loss generated by the counterfactual links. Through counterfactual logic comparison and negative sampling training, the second computation unit 135 finds the optimal parameters while continuously reducing the loss. To enhance the interpretability of the results, the system constructs a reliability function to quantify the reliability of the model's inference.

[0184] The reliability function expression is: .

[0185] This invention achieves the transformation from unstructured text to high-fidelity structured data through data processing unit 133. The system-integrated named entity recognition architecture combines a pre-trained language model (BERT) with a bidirectional long short-term memory network-conditional random field (BiLSTM-CRF), which can accurately extract subject-specific terms, conceptual words, and formula symbols, and perform entity disambiguation by combining encyclopedic knowledge, ensuring the accuracy of the basic nodes of the knowledge graph. The depth-first search (DFS) algorithm is used to recursively mine the deep dependencies between subject-specific knowledge entities, transforming the messy text into a knowledge chain embedding sequence 113 containing evolutionary paths. This serialization representation 105 provides standardized input with topological fidelity for subsequent deep learning, significantly improving the efficiency and quality of data transformation.

[0186] In this invention, the first computing unit 134 achieves a deep characterization of subject logic within the numerical space, solving the problem of traditional methods struggling to handle abstract relationships. By deeply integrating word positional information, word segmentation semantics, and the topological structure information of subject knowledge, the system can accurately generate complex relation vectors representing hierarchical, deductive, and other subject logic. Simultaneously, the system introduces a dynamic weight-aware mechanism, automatically adjusting weights based on causal logic or the importance of knowledge points using a state function. This ensures that the representation of subject knowledge is no longer a static stack of features, but rather a dynamic semantic association with logical priority, greatly improving the accuracy of subject knowledge modeling.

[0187] To address the shortcomings of traditional embedding methods in their insufficient reliance on long-range knowledge, this invention ensures logical coherence through a link-level information enhancement mechanism. The first computational unit 134 effectively compensates for the decay problem of long-distance knowledge paths by constructing a linear combination 112 of knowledge chain representations and combining it with a weight sequence 114. This linear combination mechanism allows the system to decompose the representations of long-tail or sparse paths into weighted aggregations of existing nodes, enabling the system to generate high-quality knowledge path recommendations even in "cold start" scenarios lacking historical data, thus enhancing the system's robustness in practical educational applications.

[0188] The second computational unit 135 introduces a collaborative modeling mechanism for endogenous and exogenous variables, significantly enhancing the system's robustness in complex contexts. The structured recurrent neural network model strengthens the native connections and semantic continuity between nodes by aggregating entity and relational features layer by layer, generating stable endogenous representations. Simultaneously captured information on exogenous variables at system boundaries enhances the modeling accuracy of the external context through dynamic coupling mapping. This collaborative mechanism improves the model's resistance to external noise, ensuring that the knowledge reasoning process maintains extremely high stability in diverse teaching application scenarios.

[0189] This invention not only improves prediction accuracy but also addresses the black-box problem of deep learning through a counterfactual reasoning mechanism. The system trains by comparing generated factual and counterfactual links, utilizing the nonlinear mapping capability of a multilayer perceptron to classify the links, ensuring the uniqueness and determinism of the output results. Finally, by constructing a reliability function that combines endogenous, exogenous, and embedding strength, the system quantifies the interpretability of the model's reasoning results. This makes reasoning tasks in subject graphs (such as learning path recommendation) not only more numerically accurate but also provides teachers and students with trustworthy and understandable decision-making basis at the logical level.

[0190] As described above, the data processing unit 133, through the first data transmission port 138, connects to the first computing unit 134, which has structure-enhanced embedding capabilities, to extract structured knowledge from the original subject-specific text data 101. This allows the first computing unit 134 to form a cascaded communication connection with the data processing unit 133 when performing serialization representation 105 on entity and relation representations. The data processing unit 133 has a knowledge extraction and deep search model. When it receives text data 101 input through the first data input port 137, it is in the logical flow of constructing a knowledge graph 102, so that the first computing unit 134 can receive the initialization vector sequence generated by the depth-first search algorithm. When performing structured representation 104, the data processing unit 133 is in the logical position of recursively exploring entity dependencies, so that the text data 101 is transformed into an ordered knowledge sequence 103 containing entities and relations.

[0191] Compared to CN115935968A, which simply uses the average of word vectors as prevectors and does not consider the topological hierarchy, the first computing unit 134 of this invention performs the steps of position embedding 107, word segmentation embedding 108, and structural information embedding 110. In this case, the position embedding vector and the word segmentation embedding vector are in a direct summation operation with the structural information embedding vector. In this state, a fusion operation relationship is established, fixing the first computation unit 134 to the semantic computation part of generating the complex relation representation 109. The first computation unit 134 is used to guide the embedding requirements issued by the initialization vector sequence to a specific level in the continuous semantic space. The dynamic weight sequence 114 is used to evaluate the importance of the link based on causal logic, and the information function 115 is used to aggregate link-level features. Wherein, when it is necessary to represent the hierarchical logic of the discipline, the first computation unit 134 performs a linear combination 112 with the weight sequence 114 under the condition of performing the sequence decoding operation 111. When it is necessary to enhance the long-range knowledge path dependency, the weight sequence 114 performs a product summation relationship with the knowledge chain embedding sequence 113 under the condition of dynamic product, so that the information function 115 has a logical support relationship with the scoring function 129 according to the causal mapping relationship.

[0192] The second computing unit 135 is used to perform counterfactual reasoning and link feature extraction. The structured recurrent neural network model 116 is used to generate aggregated endogenous variables 117 and captured exogenous variables 118. When the knowledge chain embedding sequence 113 is input, the second computing unit 135 performs a layer-by-layer aggregation relationship with the aggregated endogenous variables 117 under the condition of recursive computation. When it is necessary to perceive background environmental interference, the aggregated endogenous variables 117 performs an association modeling relationship with the captured exogenous variables 118 under the condition of dynamic coupling mapping. This allows the second computing unit 135 to be connected to the multilayer perceptron 124 in the case of the first computing unit 134 being in a cascaded connection relationship. This allows the multilayer perceptron 124 to perform a classification mapping relationship with the adjacency matrix 121.

[0193] The second computing unit 135 obtains feature vector 119, feature matrix 120, and adjacency matrix 121 through extraction operation 122. The extracted feature vector 119 and feature matrix 120 can establish a topological fidelity relationship when the second computing unit 135 performs computation operation 123, thus fixing the second computing unit 135 to the location of link determination operation 126. The second computing unit 135 is connected to the processor 131, which has a sampling training unit 136 and is in a data reading relationship with the storage component 132, by receiving the knowledge chain embedding sequence 113 output by the first computing unit 134 through the second data transmission port 139, enabling the sampling training unit 136 to establish a parameter feedback relationship with the second computing unit 135.

[0194] Compared to CN110837567A, which only establishes a static structural search space and lacks a measure of the interpretability of the prediction results, the sampling training unit 136 of this invention executes a negative sampling strategy and a reliability function. The fact link 127 and the counterfactual link 128, together with the sampling training unit 136, form a closed-loop optimization structure, enabling the closed-loop optimization structure to possess the cooperative iterative characteristics of calculating the loss of the path score function 129 and the estimated loss of the counterfactual result. The sampling training unit 136 is used to perform adaptive optimization of model parameters, the fact link 127 is used to provide correct subject logic references, and the reliability function is used to quantify the interpretability of reasoning. Specifically, when performing iterative optimization, the sampling training unit 136, under the condition of applying a negative sampling strategy to the fact link 127, engages in a comparative training relationship with the counterfactual link 128. When the model output meets a preset threshold, the counterfactual link 128, under the condition of loss reduction, engages in a causal quantification relationship with the reliability function, and / or, when calculating the loss function L, the reliability function, according to a weighted compensation relationship, engages in a performance support relationship with the score function 129.

[0195] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A subject graph embedding system for knowledge structure enhancement, characterized in that, The system includes a processor (131), the processor (131) comprising: The data processing unit (133) performs serialization representation (105) on entities and relations in the knowledge graph (102) to generate an initialization vector sequence; The first computing unit (134) receives the initialization vector sequence, performs a fusion operation of position embedding (107), word segmentation embedding (108) and structural information embedding (110) in the semantic space to generate a knowledge chain embedding sequence (113); and calculates a dynamic weight sequence (114) based on the causal logic of subject relationship to construct an information function (115) representing the importance of the link.

2. The system according to claim 1, characterized in that, The processor (131) also includes: The second computing unit (135) constructs a structured recurrent neural network model (116) based on the knowledge chain embedding sequence (113), generates aggregated endogenous variables (117) and captured exogenous variables (118), and maps the extracted feature vector (119) into an adjacency matrix (121) with topological fidelity to determine fact links (127) and counterfactual links (128). The sampling training unit (136) uses a negative sampling strategy to generate a counterfactual link (128) for the fact link (127) and performs iterative optimization. It also combines the aggregated endogenous variable (117) and the captured exogenous variable (118) to construct a reliability function to quantify the interpretability of the model inference results.

3. The system according to claim 1 or 2, characterized in that, The specific methods by which the data processing unit (133) performs serialization representation (105) on entities and relations in the knowledge graph (102) include: The depth-first search algorithm is used to recursively explore the dependencies between subject knowledge entities to complete the structured representation (104). The triples in the generated knowledge graph (102) are transformed into an ordered knowledge sequence (103) containing entities and relations, and mapped to the initialization vector sequence.

4. The system according to any one of claims 1 to 3, characterized in that, The specific methods by which the first computing unit (134) performs the fusion operation include: Perform a direct summation operation on the position embedding vector and the word segmentation embedding vector, and simultaneously concatenate the structural information embedding vector to generate a complex relation representation that can characterize the relationship containing the upper and lower positions or the derivation logic (109). The sequence decoding operation (111) enables a refined representation of the sequence to the sequence, providing semantic support for constructing link information.

5. The system according to any one of claims 1 to 4, characterized in that, The specific steps of the first computing unit (134) in constructing the information function (115) representing the importance of the link include: The information function (115) is obtained by performing a product summation operation on the knowledge chain embedding sequence (113) and the dynamic weight sequence (114) calculated according to the causal logic of subject relationship through linear combination (112).

6. The system according to any one of claims 1 to 5, characterized in that, The specific manner in which the second computing unit (135) in the processor (131) runs the structured recurrent neural network model (116) includes: The semantic features of entities and the topological features of subject relationships are aggregated layer by layer according to the arrangement order of each element in the knowledge chain to generate aggregated endogenous variables (117). By using dynamic coupling mapping to filter covariate information outside the system boundary, we can establish a correlation model between external context vectors and endogenous representations to capture exogenous variables (118).

7. The system according to any one of claims 1 to 6, characterized in that, The specific method by which the second computing unit (135) in the processor (131) performs matrix mapping includes: Perform extraction operation (122) to extract feature vectors (119), and combine the feature vectors (119) to form feature matrix (120). Perform computational operation (123), using the coupling relationship to map the elements of the feature matrix (120) to the adjacency matrix (121) space, and output to the multilayer perceptron (124) to perform nonlinear mapping to complete the link determination.

8. The system according to any one of claims 1 to 7, characterized in that, The specific methods by which the sampling training unit (136) in the processor (131) performs negative sampling include: Generate corresponding virtual knowledge combinations for the known fact links (127); Modify the logical features of the subject relationship path to simulate a counterfactual link (128) that does not conform to the subject logic, and use it for comparative training with the factual link (127).

9. A method for embedding subject graphs to enhance knowledge structure, characterized in that, The method includes: Perform serialization representation (105) on entities and relations in the knowledge graph (102) to generate an initialization vector sequence; The initialization vector sequence is received, and the fusion operation of position embedding (107), word segmentation embedding (108) and structural information embedding (110) is performed in the semantic space to generate a knowledge chain embedding sequence (113); the dynamic weight sequence (114) is calculated based on the causal logic of subject relationship to construct an information function (115) representing the importance of the link.

10. The method according to claim 9, characterized in that, The method further includes: Based on the knowledge chain embedding sequence (113), a structured recurrent neural network model (116) is constructed to generate aggregated endogenous variables (117) and captured exogenous variables (118). The extracted feature vector (119) is mapped to an adjacency matrix (121) with topological fidelity to determine fact links (127) and counterfactual links (128). For the fact link (127), a negative sampling strategy is used to generate a counterfactual link (128) for iterative optimization. The reliability function is constructed by combining the aggregated endogenous variable (117) and the captured exogenous variable (118) to quantify the interpretability of the model inference results.

Images