Policy update-oriented human social knowledge graph construction and induction characterization method and policy update-oriented human social knowledge graph construction and induction characterization system

Abstract

The invention provides a policy update-oriented human social knowledge graph construction and induction characterization method and system, and the method comprises the steps: obtaining multi-source data in the human social field, evaluating whether the information needs to be updated or not through triggering a decision mechanism, precisely extracting newly-added information from the multi-source data, converting the newly-added information into a triple form, and carrying out the calculation of the triple form. And calculating the semantic similarity between the new triad and the related triad in the existing human and social domain multi-mode knowledge graph, evaluating the confidence coefficient of newly added information, performing detection by using a knowledge conflict detection and resolution algorithm, and performing dynamic sampling of the sub-graph by using a feedback-driven dynamic sampling mechanism, thereby obtaining the multi-mode knowledge graph of the human and social domain. Adopting a multi-level neighborhood aggregation enhancement strategy to carry out aggregation representation on the sub-graphs obtained by sampling; and carrying out fine tuning on part of the original entities influenced by the diffusion of the newly added information to obtain an updated human and social domain multi-modal knowledge graph. According to the invention, the efficiency of updating related data of human resources and social security can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of information processing, and particularly relates to a human resource and social security knowledge graph construction and induction representation method and system for policy updating. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] As a core representation method of structured knowledge, knowledge graph has become a key to promoting the development of artificial intelligence due to its powerful semantic expression and reasoning ability. It effectively organizes and represents complex knowledge in the real world through a topological structure composed of nodes (entities) and edges (relationships), and is widely used in intelligent search, personalized recommendation, decision support and question-answering systems and other important fields.

[0005] At present, when performing representation learning after updating the knowledge graph, the mainstream method still mainly relies on full retraining. That is, whenever new knowledge (entity or relationship) is added, the old and new knowledge is combined, and representation learning is performed on all triples of the entire knowledge graph to generate new embedding vectors. Although this method can ensure the accuracy of the model on the latest data, its inherent limitations are also very obvious: (1) huge consumption of computing resources: as the scale of the knowledge graph continues to expand, the computing overhead and time cost required for full training continue to grow, posing a serious challenge to hardware resources; (2) low updating efficiency: the long training period cannot meet the real-time requirements of knowledge in the fast iteration scenario of policies, resulting in that the business system cannot obtain the latest knowledge in time; (3) high economic cost: frequent full training requires continuous investment in a large amount of computing power, making the maintenance cost of the knowledge graph high. SUMMARY

[0006] In order to solve the above problems, the present application proposes a human resource and social security knowledge graph construction and induction representation method and system for policy updating, which can accurately locate key training data through a dynamic and adaptive sampling mechanism, and deeply mine neighborhood information using enhanced aggregation technology, thereby realizing efficient updating of human resource and social security related data while achieving high-quality fusion and representation of new knowledge.

[0007] According to some embodiments, the present application adopts the following technical solutions: A human resource and social security knowledge graph construction and induction representation method for policy updating, comprising the following steps: Obtaining multi-source data including structured databases, semi-structured documents and unstructured texts in the human social field, evaluating the absolute quantity of information needing to be updated, its proportion relative to the existing human social field multi-modal knowledge graph, and the potential impact degree through triggering a decision mechanism, and when the evaluation result exceeds a preset threshold, triggering the update of the human social field multi-modal knowledge graph; Precisely extracting the newly added information from the multi-source data, and converting it into a standardized triple form including a head entity, a relationship and a tail entity according to the data type of the data; Calculating the semantic similarity between the new triple and the related triple in the existing human social field multi-modal knowledge graph, and evaluating the confidence of the newly added information, detecting by using a knowledge conflict detection and resolution algorithm, and directly inserting the newly added information into the existing human social field multi-modal knowledge graph when the newly added information is highly consistent and the confidence is higher than a set value; Using a feedback-driven dynamic sampling mechanism to dynamically sample subgraphs, and dynamically adjusting the sampling probability according to the real-time feedback signal of the last round of training in each round of training iteration to focus on the neighborhood structure with the highest training value at present; Using a multi-level neighborhood aggregation enhancement strategy to aggregate the subgraphs sampled to obtain an aggregated representation; Fine-tuning part of the original entities affected by the diffusion of the newly added information to obtain an updated human social field multi-modal knowledge graph.

[0008] As an optional implementation, the process of evaluating the absolute quantity of information needing to be updated, its proportion relative to the existing human social field multi-modal knowledge graph, and the potential impact degree through triggering a decision mechanism includes using composite indicators to evaluate:

[0009] Wherein , respectively represent the number of newly added entities and relationships; , represent the total number of entities and relationships in the knowledge graph; Calculate the importance score of the new entity in the knowledge graph; Score the authority of the data source; , , are weight coefficients and satisfy , comprehensively consider three dimensions of update size, entity importance and data source reliability, and when the preset threshold, automatically trigger the update process.

[0010] As an optional implementation, the process of detecting by using a knowledge conflict detection and resolution algorithm includes a conflict score:

[0011] wherein is a normalization factor; is a BERT-based semantic similarity calculation function; is a confidence score for evaluating new knowledge triples, which is a comprehensive evaluation index for data source confidence, extraction method confidence, and triple consistency confidence; denotes the semantic neighbor set of According to the conflict score, a three-level processing mechanism is set: when , the new knowledge triples are directly inserted; when , based on and the score of each item to select whether to insert; when , it is marked for manual review.

[0012] As an optional implementation, in the process of converting data into a standardized triple form according to the data type to which the data belongs, a hybrid extraction strategy is adopted for different types of data sources to balance efficiency and accuracy: for structured data, direct parsing and mapping are performed; for semi-structured data, information is extracted using predefined rules and templates; for unstructured text, entity recognition and relationship joint extraction are performed using a deep semantic parsing model based on natural language processing.

[0013] As an optional implementation, in the process of converting data into a standardized triple form according to the data type to which the data belongs, an entity recognition and relationship extraction parallel execution strategy is adopted, and its loss function is:

[0014] wherein is the input text; , is the head and tail entity, is the relationship; this loss function jointly optimizes entity recognition and relationship extraction tasks to ensure the accuracy and consistency of knowledge extraction.

[0015] As an optional implementation, in the process of dynamic sampling of subgraphs, the sampling probability calculation formula is:

[0016] wherein denotes the gradient norm of the node embedding vector, reflecting the contribution of the node to the current loss; is the semantic similarity calculated based on the knowledge graph embedding; denotes the k-hop neighborhood of . denotes a training round; The neighborhood nodes are given more exploration opportunities at the beginning of training, and gradually focus on high-value nodes as training progresses; and during the training process, redundant or irrelevant nodes are automatically ignored.

[0017] As an optional implementation, the process of aggregating the subgraphs sampled by the multi-level neighborhood aggregation enhancement strategy includes: using a multi-head graph attention mechanism to distinguish the importance of different neighborhood nodes to capture the complex semantic relationships between entities in multiple hops, and at each layer of information transmission, each entity aggregates information from its adjacent entities, and the aggregation weight is determined by the representation of the entity itself, the representation of the neighbor and the relationship vector connecting them.

[0018] As an optional implementation, the process of fine-tuning part of the original entities affected by the diffusion of new information includes constructing a fine-tuning loss function, which includes two key parts: one is the representation smoothing constraint, and the other is the triple component constraint, which is specifically:

[0019] wherein 、 denote the entity embedding vectors before and after fine-tuning; calculate the score of the triple . is a marginal hyperparameter; is a balance hyperparameter.

[0020] A human resource knowledge graph construction and induction representation system for policy updates includes: An update perception and triggering module is configured to obtain multi-source data including structured databases, semi-structured documents and unstructured texts in the human resource field, evaluate the absolute quantity of information that needs to be updated, its proportion relative to the existing human resource field multi-modal knowledge graph, and the potential impact degree through a triggering decision mechanism, and trigger the update of the human resource field multi-modal knowledge graph when the evaluation result exceeds the preset threshold; A triple extraction module is configured to accurately extract new information from multi-source data, and convert it into a standardized triple form including a head entity, a relationship and a tail entity according to the data type of the data; A fusion and consistency checking module is configured to calculate the semantic similarity between the new triple and the related triple in the existing human resource field multi-modal knowledge graph, and evaluate the confidence of the new information, detect using a knowledge conflict detection and resolution algorithm, and directly insert the new information into the existing human resource field multi-modal knowledge graph if the new information is highly consistent and the confidence is higher than a set value. The subgraph dynamic sampling module is configured to utilize a feedback-driven dynamic sampling mechanism to perform dynamic sampling of the subgraph, and in each round of training iteration, dynamically adjusts the sampling probability according to the real-time feedback signal of the last round of training, so as to focus on the current most training valuable neighborhood structure. The multi-level neighborhood aggregation module is configured to adopt a multi-level neighborhood aggregation enhancement strategy to perform aggregated representation of the sampled subgraph. The spatial consistency fine-tuning module is configured to fine-tune part of the original entity affected by the diffusion of the new information, and obtain an updated human resource and social security field multi-modal knowledge graph.

[0021] A computer device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the policy-oriented human resource and social security knowledge graph construction and induction representation method.

[0022] Compared with the prior art, the beneficial effects of the present application are: The present application can improve the efficiency and accuracy of human resource and social security related data updating.

[0023] The sampling precision and efficiency of the present application are high: through the introduction of a feedback mechanism, the dynamic sampling strategy can focus on the neighborhood with the largest amount of information in real time, avoiding redundant calculation and improving training efficiency and stability.

[0024] The representation learning accuracy of the present application is high: the multi-level neighborhood aggregation enhancement mechanism can capture more complex semantic relationships and generate more discriminative entity and relationship representations.

[0025] The system of the present application is highly practical: the whole method forms a complete closed loop from update perception to service provision, taking into account the update efficiency and knowledge graph representation accuracy, and has strong practical value.

[0026] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are described in detail below, and the accompanying drawings are used for explanation. BRIEF DESCRIPTION OF DRAWINGS

[0027] The drawings accompanying the specification of the present application form part of the present application and are used to provide a further understanding of the present application, the illustrative embodiments of the present application and the description thereof serve to explain the present application, and do not constitute an improper limitation of the present application.

[0028] Figure 1 A policy-oriented human resource and social security knowledge graph construction and induction representation method flowchart of an embodiment. DETAILED DESCRIPTION

[0029] The present application will be further described below in conjunction with the drawings and embodiments.

[0030] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0031] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0032] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0033] Example 1 As described in the background section, existing methods often employ static or simple random sampling strategies when selecting neighborhood contexts for new entities to construct training samples. For example, they may fixate on selecting all nodes within a first- or second-order neighborhood, or perform random sampling. This "one-size-fits-all" approach ignores the specificity of different entities and their neighborhood structures, failing to dynamically select the neighborhood information with the most information and best suited to the current representation learning task based on the entity's semantic characteristics, the importance of relationships, and the model's real-time training state. This can easily lead to the sampled subgraph either containing too many redundant nodes, increasing unnecessary computational burden, or omitting key related nodes, preventing the model from learning accurate semantic representations, ultimately affecting training efficiency and the stability of the training results.

[0034] Even after obtaining suitable neighboring nodes, effectively aggregating this information into the representation of the central node is equally crucial. Many existing models employ relatively simple aggregation functions (such as mean pooling or basic graph convolution), which struggle to capture the complex, multi-layered relational semantics and long path dependencies in knowledge graphs. For example, in a multimodal knowledge graph in the human resources and social security domain, a "college graduate" entity may be associated with multiple concepts at different levels, such as "employment policy," "social security regulations," and "local talent introduction plan." Simple aggregation methods cannot distinguish the different contributions of these relational paths, resulting in insufficient discriminative power in the generated feature representations, failing to accurately reflect the semantic position of the entity in the global graph.

[0035] Current technology lacks a comprehensive multimodal knowledge graph construction scheme for the human and social domain that can intelligently optimize the sampling process and deepen the semantic fusion of the neighborhood. This severely restricts the practical performance of knowledge graphs in frequently updated scenarios. Therefore, there is an urgent need for an innovative method that can accurately locate key training data through a dynamic and adaptive sampling mechanism and use enhanced aggregation techniques to deeply mine neighborhood information, thereby achieving high-quality fusion and representation of new knowledge while ensuring high efficiency.

[0036] like Figure 1 As shown, the present invention will be described in detail below with reference to embodiments. This embodiment takes the update of human resources and social security industry policies as an example to illustrate the method for constructing and summarizing human resources and social security knowledge graphs for policy updates, including the following steps: S1. Knowledge Update Perception and Triggering This step establishes an intelligent update triggering mechanism. The system continuously monitors multi-source data in the human resources and social security field, including policy databases (structured), policy documents in XML / JSON format (semi-structured), and policy press releases (unstructured). The core of this mechanism is to accurately identify meaningful update events and avoid overreacting to minor changes. To achieve this goal, we designed a multi-dimensional evaluation system that considers not only the quantity of updates but also the quality impact of updates. Update triggering is evaluated using a composite indicator system:

[0037] in , These represent the number of newly added entities and relations, respectively. , This represents the total number of entities and relations in the knowledge graph; Calculate the importance score of new entities in the knowledge graph; Rate the authority of the data source; , , The weighting coefficients are and satisfy the following conditions: This formula comprehensively considers three dimensions: update scale, entity importance, and data source reliability. When a preset threshold is reached, the system automatically triggers the update process. In actual deployment, we adjust the weight coefficients according to specific business scenarios. For example, for the human resources and social security policy knowledge graph, we assign higher weights to the data source's authority to ensure the accuracy and authority of policy documents. This approach effectively avoids the problem of frequently activating computing resources due to minor changes, while ensuring that important policy updates can be captured and processed in a timely manner.

[0038] S2. New Knowledge Triple Extraction After the update process is triggered, this step is responsible for accurately extracting structured knowledge from heterogeneous data sources. This is a crucial data preprocessing step, and its quality directly affects the effectiveness of all subsequent steps. We adopt differentiated extraction strategies for different data sources: for structured data, it is directly converted into triples based on predefined schema mapping rules, which is efficient and accurate; for semi-structured data, we use template matching technology based on XPath / JSONPath for information extraction, and we use flexible extraction templates to adapt to policy documents of different formats; for unstructured text, we use a joint extraction model based on BERT, which can deeply understand the semantic information of natural language text.

[0039] Specifically, a strategy of parallel execution of entity recognition and relation extraction is adopted, and its loss function is:

[0040] in Enter the text; , It is a head and tail entity. The loss function, by jointly optimizing the two tasks of entity recognition and relation extraction, ensures the accuracy and consistency of knowledge extraction.

[0041] S3. New Knowledge Integration and Consistency Verification This step aims to integrate the extracted new knowledge into the existing knowledge graph, a complex process involving semantic understanding and logical reasoning. The fusion process employs a confidence-based conflict resolution strategy to ensure both the timely integration of new knowledge and the maintenance of the knowledge graph's logical consistency. We first calculate the conflict score:

[0042] in Normalization factor; This is a semantic similarity calculation function based on BERT; It is used to evaluate the confidence score of new knowledge triples. It is a comprehensive evaluation index for multi-dimensional evaluation (data source confidence, extraction method confidence, triple consistency confidence). express The semantic neighbor set.

[0043] A three-tiered handling mechanism is established based on conflict scores: when When, directly insert the new knowledge triple; when At that time, based on and Choose whether to insert the score for each sub-item; when The time stamp indicates that manual review is required. This hierarchical processing mechanism ensures both the efficiency of knowledge integration and the quality of knowledge.

[0044] S4. Subgraph Dynamic Sampling This step is the core innovation of this method. Traditional fixed sampling methods cannot adapt to the characteristics of different entities, while random sampling lacks specificity. Therefore, when dealing with new entities... When constructing the training subgraph, dynamic sampling of the subgraph utilizes a gradient-feedback-based dynamic sampling mechanism, which can adjust the sampling points in a timely manner based on real-time feedback during training. The sampling probability calculation formula is as follows:

[0045] in Represents a node The gradient norm of the embedding vector reflects the degree to which the node contributes to the current loss; The semantic similarity is calculated based on knowledge graph embedding. express k-jump neighborhood; This indicates the training rounds. This mechanism ensures that the model dynamically adjusts the sampling probability in each training round, always focusing on the neighborhood nodes with the most information and the greatest benefit to representation learning. Furthermore, in the early stages of training, we give neighborhood nodes more opportunities to explore, gradually focusing on high-value nodes as training progresses. This mechanism also has good scalability, adapting to knowledge graphs of different sizes and complexities.

[0046] S5. Multi-level neighborhood aggregation After obtaining high-quality training subgraphs, this step employs an improved graph attention network for multi-level neighborhood aggregation. Traditional graph neural network methods often neglect relational semantics and the differences in importance among different neighbors, resulting in poor aggregation performance. By introducing a relation-aware attention mechanism, more refined information aggregation can be achieved. The aggregation formula for the layer is designed as follows:

[0047] Attention coefficient Calculated through a relation-aware attention mechanism:

[0048] in This represents a vector concatenation operation; For the number of attention heads; For relationship The embedding vector; For the first Layer The learnable weight matrix for each size; These are learnable parameter vectors for the attention mechanism. This design enables the model to distinguish the importance of different neighborhood nodes and relationship paths, generating high-quality entity representations through multi-layered information transfer.

[0049] S6. Fine-tuning of global representation space consistency After introducing new knowledge and updating its representation, this step aims to maintain the stability of the entire knowledge graph representation space. This is a crucial step in preventing catastrophic forgetting and requires finding the optimal balance between learning new knowledge and retaining old knowledge. Specifically, a loss function based on dual constraints is used to achieve fine-tuning of the global representation space consistency.

[0050] in 、 Representing entities respectively Embedding vectors before and after fine-tuning; Calculate triples The score; These are marginal hyperparameters; To balance hyperparameters, the "representation smoothing constraint" ensures that the old entity representation does not drift drastically, and the "trinity component constraint" guarantees that the score of the original correct triples will not decrease due to the addition of new knowledge.

[0051] Example 2 A knowledge graph construction and inductive representation system for human resources and social security oriented towards policy updates, including; The knowledge update perception and triggering module is responsible for continuously monitoring changes in multi-source knowledge data. Its built-in triggering decision mechanism can calculate composite triggering indicators in real time, which comprehensively evaluate the number, proportion, and potential impact of newly added entities and relationships. When the indicators exceed preset thresholds, the module automatically sends a start signal to the system, thereby intelligently triggering the entire update process and ensuring that the system only allocates computing resources when there are substantial knowledge updates.

[0052] The new knowledge triple extraction module is responsible for accurately extracting structured knowledge from heterogeneous data sources. It integrates a hybrid extraction strategy, employing direct parsing, rule-based template matching, and deep semantic parsing based on natural language processing for structured, semi-structured, and unstructured data, respectively, to extract knowledge triples. The module's output is cleaned and standardized (head entity, relation, tail entity) triples, providing high-quality knowledge for subsequent processing.

[0053] The new knowledge fusion and consistency verification module ensures that the introduction of new knowledge does not disrupt the consistency of the existing knowledge system. It implements knowledge conflict detection and resolution algorithms by calculating semantic similarity and confidence levels. This module can automatically handle most consistency conflicts (such as prioritizing based on data source authority). For complex conflicts that cannot be resolved automatically, it marks them and submits them to domain experts for decision-making, thereby ensuring the logical rigor and reliability of the knowledge graph.

[0054] The subgraph dynamic sampling module dynamically constructs the optimal training subgraph for each new entity whose representation needs to be learned. It abandons fixed or random sampling strategies, employing a feedback-driven dynamic sampling mechanism. This mechanism dynamically adjusts the sampling probability of neighboring nodes based on real-time gradient information during model training, ensuring that each iteration focuses on the most valuable neighborhood structure for training, greatly improving training efficiency and the quality of representation learning.

[0055] Multi-level Neighborhood Aggregation Module. This module employs a multi-level neighborhood aggregation enhancement strategy to generate high-quality embedding representations for new entities in the subgraph. At its core is a graph neural network integrating a multi-head graph attention mechanism and relational path encoding. This network can distinguish the importance of different neighbor nodes and relational paths with fine granularity. Through multi-level information transmission and aggregation, it ultimately outputs a highly discriminative vector representation that accurately captures the complex semantic relationships of entities in the global knowledge graph.

[0056] The Global Representation Space Consistency Fine-Tuning Module. This module is dedicated to performing constrained, lightweight fine-tuning of affected existing entities, maintaining the balance of the entire representation space while introducing new knowledge. Through a built-in fine-tuning loss function, this module can effectively prevent the "catastrophic forgetting" of existing knowledge while absorbing new knowledge, ensuring the coexistence of new and old knowledge in a unified vector space and maintaining the coherence and stability of the overall knowledge representation.

[0057] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0058] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0059] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0060] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0061] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing and representing human resource knowledge graph oriented to policy update, characterized in that, The method comprises the following steps: Obtaining multi-source data including structured databases, semi-structured documents and unstructured texts in the human social field, evaluating the absolute quantity of information needing to be updated, the proportion of the information relative to the existing human social field multi-modal knowledge graph and the potential influence degree through a triggering decision mechanism, and triggering the update of the human social field multi-modal knowledge graph when the evaluation result exceeds a preset threshold; Precisely extracting the new information from the multi-source data, and converting the new information into a standardized triple form including a head entity, a relation and a tail entity according to the data type of the new information; Calculating the semantic similarity between the new triple and the related triple in the existing human social field multi-modal knowledge graph, evaluating the confidence of the new information, detecting the new information by using a knowledge conflict detection and resolution algorithm, and directly inserting the new information into the existing human social field multi-modal knowledge graph when the new information is highly consistent and the confidence of the new information is higher than a set value; Performing dynamic sampling of subgraphs by using a feedback-driven dynamic sampling mechanism, dynamically adjusting the sampling probability according to the real-time feedback signal of the last training iteration in each training iteration, and focusing on the neighborhood structure with the highest training value at present; Performing aggregated representation of the sampled subgraphs by using a multi-level neighborhood aggregation enhancement strategy; Fine-tuning part of the original entities affected by the diffusion of the new information, and obtaining the updated human social field multi-modal knowledge graph.

2. The method of claim 1, wherein the method is a policy-oriented update method for constructing and representing human resource knowledge graph. The process of evaluating the absolute quantity of information needing to be updated, the proportion of the information relative to the existing human social field multi-modal knowledge graph and the potential influence degree through a triggering decision mechanism comprises the following steps of using a composite index to evaluate: wherein , represent the number of new entities and relationships, respectively; , represent the total number of entities and relationships in the knowledge graph; calculate the importance score of new entities in the knowledge graph; score the authority of data sources; , , is a weight coefficient and satisfies , considering the three dimensions of update size, entity importance and data source reliability, when the preset threshold, automatically trigger the update process.

3. The method of claim 1, wherein the method is characterized by, The process of detecting the new information by using a knowledge conflict detection and resolution algorithm comprises the following steps of calculating a conflict score: wherein is a normalization factor; is a BERT-based semantic similarity calculation function; is a confidence score for evaluating new knowledge triples, which is a comprehensive evaluation index for data source confidence, extraction method confidence, and triple consistency confidence; represents a semantic neighbor set of According to the conflict score, three-level processing mechanisms are set: when , directly insert new knowledge triples; when , based on and , select whether to insert each sub-item score; when , mark for manual review.

4. The method of claim 1, wherein the method is characterized by, In the process of converting the new information into a standardized triple form according to the data type of the new information, a hybrid extraction strategy is adopted to balance efficiency and accuracy for different types of data sources: for structured data, direct parsing and mapping are performed; for semi-structured data, information is extracted by using predefined rules and templates; and for unstructured texts, entity recognition and relation joint extraction are performed by using a deep semantic parsing model based on natural language processing.

5. The policy update-oriented human resource knowledge graph construction and induction representation method according to claim 1, characterized in that, In the process of converting the new information into a standardized triple form according to the data type of the new information, an entity recognition and relation extraction parallel execution strategy is adopted, and a loss function is as follows: wherein is the input text; , is the head-tail entity, is the relation; the loss function ensures the accuracy and consistency of knowledge extraction by jointly optimizing the entity recognition and relation extraction tasks.

6. The method for constructing and representing a human resources and social security knowledge graph oriented towards policy updates as described in claim 1, characterized in that, In the process of dynamically sampling subgraphs, the sampling probability calculation formula is as follows: wherein representing nodes a gradient norm of the embedding vector, reflecting a degree of contribution of the node to a current loss; a semantic similarity calculated based on the knowledge graph embedding; representing k-hop neighborhood of representing a training round; More exploration opportunities are given to neighborhood nodes in the early stage of training, and the training gradually focuses on high-value nodes; and during the training process, redundant or irrelevant nodes are automatically ignored.

7. The policy update-oriented human resource knowledge graph construction and induction representation method according to claim 1, characterized in that, The process of performing aggregated representation of the sampled subgraphs by using a multi-level neighborhood aggregation enhancement strategy comprises the following steps of using a multi-head graph attention mechanism to distinguish the importance of different neighborhood nodes, capturing the complex semantic relationship between entities in multiple hops, and aggregating information from adjacent entities at each layer of information transmission, and the aggregation weight is determined by the representation of the entity itself, the representation of the neighbor and the relation vector connecting them.

8. The policy update-oriented human resource knowledge graph construction and induction representation method according to claim 1, characterized in that, The process of fine-tuning part of the original entity affected by the diffusion of new information includes constructing a fine-tuning loss function, which includes two key parts: one is a smooth constraint, and the other is a triple component constraint, which is: wherein 、 respectively represent entities before and after fine-tuning the embedding vectors; computing scores for triples is a marginal hyperparameter; is a balancing hyperparameter.​ 9. A policy-oriented update human resource knowledge graph construction and induction representation system, characterized in that, Comprise: The update perception and triggering module is configured to obtain multi-source data including structured databases, semi-structured documents and unstructured texts in the human resource field, evaluate the absolute quantity of information that needs to be updated, its proportion relative to the existing human resource field multi-modal knowledge graph, and the potential impact through a triggering decision mechanism, and when the evaluation result exceeds the preset threshold, trigger the update of the human resource field multi-modal knowledge graph; The triple extraction module is configured to accurately extract new information from multi-source data, and convert it into a standardized triple form including head entity, relationship and tail entity according to the data type of the data; The fusion and consistency checking module is configured to calculate the semantic similarity between the new triple and the related triple in the existing human resource field multi-modal knowledge graph, and evaluate the confidence of the new information, and use knowledge conflict detection and resolution algorithm for detection, for the new information with high consistency and confidence higher than the set value, directly insert into the existing human resource field multi-modal knowledge graph; The subgraph dynamic sampling module is configured to use a feedback-driven dynamic sampling mechanism to dynamically sample subgraphs, and in each training iteration, dynamically adjust the sampling probability according to the real-time feedback signal of the last training to focus on the neighborhood structure with the most training value at present. The multi-level neighborhood aggregation module is configured to use a multi-level neighborhood aggregation enhancement strategy to aggregate the sampled subgraphs; The spatial consistency fine-tuning module is configured to fine-tune part of the original entity affected by the diffusion of new information, and obtain the updated human resource field multi-modal knowledge graph.

10. A computer device, characterized by A computer program product comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the policy-oriented human resource knowledge graph construction and induction method for updating according to any one of claims 1-8 when executing the program.

Images