Multi-level construction and intelligent recall strategy implementation method, system and equipment of energy policy mapping knowledge domain and medium

Abstract

The invention discloses a multi-level construction and intelligent recall strategy implementation method, system and device for an energy policy knowledge graph and a medium, and belongs to the technical field of energy policy monitoring, and the method comprises the steps: obtaining energy policy text data, carrying out energy policy correlation analysis, and extracting entity information, relation clues and text vector representation from an energy policy text; carrying out knowledge graph construction by utilizing the extracted entity information, relation clues and vector representation, and generating a knowledge graph structure comprising an entity type, an entity attribute, a relation type and confidence; updating the knowledge graph according to the operations of newly adding, revising and revoking the policy document; and executing a multi-path recall strategy including longitudinal traceability recall and transverse comparison recall by using the knowledge graph structure, and generating a relevance recall result related to the target policy. According to the invention, efficient traceability, accurate comparison and full-link intelligent management of energy policies are realized, and the policy retrieval efficiency and accuracy are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of energy policy monitoring technology, specifically to methods, systems, devices, and media for implementing multi-level construction and intelligent recall strategies of energy policy knowledge graphs. Background Technology

[0002] The energy policy system, as a core tool guiding the development direction of energy enterprises and directing energy production and consumption at all levels, exhibits typical multi-level transmission characteristics, forming a complex policy network from macro-level strategies to micro-level local implementation details. Currently, this system faces multiple management challenges in its operation: The transmission of policy effectiveness often involves implicit connections that are difficult to trace. A typical example is that the technical parameters required in local new energy implementation plans often originate from early guiding documents from national ministries. However, due to the lack of a clear document numbering mechanism, staff must manually backtrack and cross-verify through cross-departmental historical document databases, a process whose complexity severely restricts policy implementation efficiency. Simultaneously, the implementation of policies across administrative regions highlights the comparative dilemma of differentiated execution. Taking air pollution prevention and control as an example, the same action plan presents technical indicators with gradient differences in supporting documents from different regions. While this regionally specific adjustment reflects policy flexibility, it also requires significant time and manual comparison for cross-regional policy collaborative analysis. More noteworthy is the version management risk brought about by the high frequency of iterations of technical specifications. Cases of multiple revisions in the field of energy storage safety standards within three years show that construction projects at different times may involve cross-references of multiple valid version clauses, which significantly increases the complexity of compliance review.

[0003] Existing technologies have revealed significant limitations in addressing these challenges. Traditional retrieval systems struggle to penetrate the surface information of policy texts to identify deep logical connections; static knowledge graphs cannot effectively capture the dynamic characteristics of policy provisions as they evolve over time; and general natural language processing models often suffer from conceptual confusion when faced with energy-related technical terms, particularly errors in distinguishing core concepts such as green certificate trading and carbon quotas, which directly impact the accuracy of policy analysis. These technological bottlenecks severely restrict the practical application value of intelligent policy analysis systems.

[0004] A consensus has been reached on a breakthrough direction, urgently requiring the construction of a multi-dimensional traceability system integrating document number tracking, semantic association, and implementation feedback verification. This includes developing a dynamic knowledge graph with time-lapse models and visualization capabilities for version evolution paths, and simultaneously improving the terminology recognition accuracy of natural language processing frameworks through specialized training on energy-related corpora. This intelligent upgrade path will effectively connect the transmission links across five policy levels, providing precise decision support capabilities for energy governance. Summary of the Invention

[0005] In view of the above-mentioned problems, the present invention is proposed.

[0006] Therefore, the technical problem solved by this invention is: how to address the difficulty of automatically identifying the "formulation basis" implicit in policy texts when traditional retrieval systems rely on technologies such as text similarity and keyword matching to find relevant policies, making it difficult to trace local energy policies back to their source. Furthermore, since there is no traditional knowledge graph or other structured data retrieval method for energy policies, when conducting horizontal comparisons of policies at the same level, it is necessary to manually search and compare policies issued by units at the same level, which is prone to omissions and errors.

[0007] To address the aforementioned technical problems, this invention provides the following technical solution: a method for multi-level construction and intelligent recall strategy implementation of an energy policy knowledge graph, comprising: acquiring energy policy text data; performing energy policy correlation analysis; extracting entity information, relationship clues, and text vector representations from the energy policy text; constructing a knowledge graph using the extracted entity information, relationship clues, and vector representations to generate a knowledge graph structure including entity type, entity attributes, relationship type, and confidence level; updating the knowledge graph for the addition, revision, and repeal of policy documents; and using the knowledge graph structure to execute a multi-path recall strategy including vertical source tracing recall and horizontal comparison recall to generate a correlation recall result related to the target policy.

[0008] As a preferred embodiment of the multi-level construction and intelligent recall strategy implementation method of the energy policy knowledge graph described in this invention, the energy policy correlation analysis includes processing energy policy texts using a combination of multiple analysis methods to improve the coverage of correlation judgment.

[0009] As a preferred embodiment of the multi-level construction and intelligent recall strategy for the energy policy knowledge graph described in this invention, the knowledge graph construction includes using a unified entity relationship modeling method based on text extraction results to complete the structured expression of energy policies.

[0010] As a preferred embodiment of the multi-level construction and intelligent recall strategy implementation method of the energy policy knowledge graph described in this invention, the updating of the knowledge graph includes the processing flow of policy documents, covering three scenarios: addition, revision and repeal, and adjusting the graph relationships based on interaction rules and confidence information.

[0011] As a preferred embodiment of the multi-level construction and intelligent recall strategy for the energy policy knowledge graph described in this invention, the knowledge graph construction further includes: constructing knowledge graph nodes with energy policy documents as entities based on the extracted entity information and relationship clues. Each node includes three types of attribute information: basic attributes, classification attributes, and derived attributes. The basic attributes include policy document number, issuing department, issuance time, and level of effectiveness; the classification attributes include energy type, jurisdiction, and policy type; and the derived attributes include a list of associated documents, revision history, and vectorized representation. Relationship edges are established between nodes, with relationship types including based on, detailed implementation, reference association, and clause conflict. A confidence value is assigned to each relationship edge for managing the relationships during the update process.

[0012] This preferred scheme divides the attributes of entities in energy policy documents into three categories: basic attributes, classification attributes, and derived attributes. It also constructs multiple types of relationship edges, including those based on and detailed implementations. This allows for the formation of a clearly structured and semantically rich entity relationship network during the construction process. Furthermore, by assigning a confidence value to each relationship, it provides a measurable and controllable basis for subsequent version evolution and relationship validity judgment, thereby enhancing the structural stability and evolvability of the graph during dynamic maintenance.

[0013] As a preferred embodiment of the multi-level construction and intelligent recall strategy implementation method of the energy policy knowledge graph described in this invention, the updating of the knowledge graph further includes: when adding a new policy document, extracting the document number, issuing department, and issuance time of the new policy document as basic entity attributes for initialization, performing energy policy correlation analysis, establishing a basis or detailed implementation relationship with existing policies, and storing the correlation relationship according to confidence level; when revising a policy, generating a new version node and recording revision history information, identifying relation edges with conflicting clauses by comparing the semantic vectors of the clauses of the new and old versions, and adjusting the confidence value; when a policy is repealed, setting the repealed policy to a historical state, retaining the correlation record of the policy before repeal, performing confidence decay on the downstream relation edges associated with the policy, and identifying the replacement basis document number through semantic analysis.

[0014] This preferred solution defines the processing flow for three policy change scenarios: addition, revision, and repeal. It also combines revision history, clause semantic vector comparison, and confidence value adjustment mechanisms to achieve differentiated management of policy relationships as they change throughout their lifecycle.

[0015] As a preferred embodiment of the multi-level construction and intelligent recall strategy for the energy policy knowledge graph described in this invention, the multi-path recall strategy includes: performing vertical tracing recall based on the relationship edges existing in the constructed knowledge graph, querying higher-level policies level by level along local, provincial, and national paths; performing horizontal comparison recall based on the refined implementation relationship edges, identifying the different implementation schemes among multiple local policies under the same higher-level policy; and combining the vectorized representation of each entity in the knowledge graph, selecting the entity most similar to the target policy text through Euclidean distance calculation to assist in completing and comparing related policy documents.

[0016] This preferred solution integrates two dimensions of recall strategy: structural path and semantic vector similarity. While supporting vertical policy tracing and horizontal difference comparison, it introduces a vectorized screening mechanism, which can supplement the discovery of potential policy relationships in scenarios without explicit references or with weak semantic associations, thereby improving the coverage and semantic accuracy of the recall strategy in complex reference chain structures.

[0017] Another objective of this invention is to provide a system for implementing a multi-level construction and intelligent recall strategy for energy policy knowledge graphs.

[0018] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-level construction and intelligent recall strategy implementation system for an energy policy knowledge graph, comprising: an energy policy correlation analysis module, an energy policy knowledge graph maintenance module, and an intelligent recall module; the energy policy correlation analysis module is used to acquire energy policy text data, perform energy policy correlation analysis, and extract entity information, relationship clues, and text vector representations from the energy policy text; the energy policy correlation analysis module is used to construct a knowledge graph using the extracted entity information, relationship clues, and vector representations, generating a knowledge graph structure including entity type, entity attributes, relationship type, and confidence level; updating the knowledge graph for the addition, revision, and repeal of policy documents; the intelligent recall module is used to utilize the knowledge graph structure to execute a multi-path recall strategy including vertical source tracing recall and horizontal comparison recall, generating correlation recall results related to the target policy.

[0019] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the multi-level construction and intelligent recall strategy implementation method of the energy policy knowledge graph.

[0020] The present invention provides a computer-readable storage medium storing a computer program thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method for implementing the multi-level construction and intelligent recall strategy of the energy policy knowledge graph.

[0021] The beneficial effects of this invention are as follows: This invention achieves automatic penetration of the policy hierarchy transmission link through a multi-level association parsing mechanism and spatiotemporal dimension indexing technology. The explicit relation edge traversal algorithm based on knowledge graphs can compress the vertical tracing response time from national policies to local regulations, thus improving efficiency. Simultaneously, the semantic similarity pre-screening module reduces the frequency of large language model calls, and the efficiency of candidate set generation for horizontal comparison tasks is improved compared to traditional methods.

[0022] This invention improves the accuracy of vertical source tracing through the synergistic effect of a multimodal association engine (explicit reference resolution + semantic vector matching + large language model inference). In horizontal comparison scenarios, a hybrid recall strategy based on knowledge graph attribute constraints (such as level of effectiveness and jurisdiction) enhances the identification accuracy of differentiated clauses in similar policies. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 The overall flowchart of the method for implementing a multi-level construction and intelligent recall strategy of energy policy knowledge graph provided in one embodiment of the present invention is shown.

[0025] Figure 2 The method for implementing a multi-level construction and intelligent recall strategy of an energy policy knowledge graph provided in one embodiment of the present invention is based on a relational edge graph.

[0026] Figure 3 This is a detailed implementation relationship side diagram of the multi-level construction and intelligent recall strategy implementation method of the energy policy knowledge graph provided in one embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of a scheme module diagram for a multi-level construction and intelligent recall strategy implementation system for an energy policy knowledge graph, provided as an embodiment of the present invention. Detailed Implementation

[0028] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0029] Example 1, referring to Figure 1This is one embodiment of the present invention, which provides a method for implementing a multi-level construction and intelligent recall strategy for an energy policy knowledge graph, including:

[0030] S1. Obtain energy policy text data, conduct energy policy correlation analysis, and extract entity information, relationship clues, and text vector representations from the energy policy text.

[0031] S2. Construct a knowledge graph using the extracted entity information, relational clues, and vector representations, generating a knowledge graph structure that includes entity type, entity attributes, relation type, and confidence level.

[0032] S3. Update the knowledge graph for the addition, revision and repeal of policy documents.

[0033] S4. Utilize the knowledge graph structure to execute a multi-path recall strategy, including vertical source tracing recall and horizontal comparison recall, to generate relevant recall results related to the target policy.

[0034] It should be noted that with the increased frequency of energy policy releases, explicit references, implicit dependencies, version replacements, and structural differences between policies have become increasingly complex. Traditional methods relying on manual analysis and static rules are no longer efficient in supporting tasks such as policy tracing, comparison, and compliance verification. Especially given the frequent policy version changes, inconsistent document formats, and diverse semantic expressions, the true semantic relationships between policies are easily overlooked or misjudged.

[0035] Therefore, to address the aforementioned issues of structural confusion, difficulty in tracing, and lagging maintenance, steps S1-S4 are used to complete the structured parsing and semantic modeling of policy texts, construct a dynamically evolving energy policy knowledge graph, and achieve high-precision extraction and maintenance of relevant information. Furthermore, a recall method is designed by combining structural paths and semantic vectors to effectively support typical application scenarios such as vertical tracing and horizontal difference comparison, thereby improving the scalability and accuracy of intelligent energy policy management.

[0036] Example 2, refer to Figure 2 and Figure 3 As an embodiment of the present invention, based on the previous embodiment, a method for implementing a multi-level construction and intelligent recall strategy for an energy policy knowledge graph is provided, including:

[0037] In this embodiment of the application, step S1 involves acquiring energy policy text data, performing energy policy correlation analysis, and extracting entity information, relationship clues, and text vector representations from the energy policy text.

[0038] The energy policy correlation analysis uses a combination of analytical methods to process energy policy texts, thereby expanding the scope of correlation judgment.

[0039] Based on keyword-based energy policy correlation technology, some energy policies contain texts that explicitly state the basis for their formulation, such as "according to [document name]" or "in accordance with [document name]". Regular expressions and information extraction NLP models are used to conduct preliminary analysis of the text content.

[0040] Using UIE model-based entity recognition technology, entity extraction is performed on energy policy texts using the UIE model, and classification information such as the issuing department and the types of energy involved is output.

[0041] Using semantic similarity correlation analysis based on text embedding models, policy documents with correlation often exhibit high semantic similarity. When energy policies are encoded using embedding models, the encoded vectors of correlated policies tend to have lower Euclidean distances. This method can be used to make preliminary correlation judgments on policy texts with the same classification and issuing departments with hierarchical relationships, reducing the cost of calling large language models.

[0042] This study employs a large language model-based correlation analysis technique to perform correlation analysis on policy texts with the same category and those issued by departments with hierarchical relationships. Cue word engineering and few-shot methods are used to enhance the performance of the general-purpose large language model in this task.

[0043] In one alternative implementation, energy policy correlation analysis can be achieved through knowledge graph rule matching, that is, based on existing policy citation rules in the field or a manually constructed template system, the policy content is matched and analyzed to identify citation paths and structural logic.

[0044] In another alternative implementation, energy policy correlation analysis can employ a graph neural network-based approach, combining existing policy graph structures with historical node semantic information to predict connections between new policy content nodes and existing nodes in the graph, thereby uncovering potential referencing or conflict relationships.

[0045] This application's implementation method combines keyword structure recognition, entity extraction, semantic vector modeling, and large language model reasoning analysis to achieve a composite recognition capability for explicit citations, semantic inheritance, version relationships, and clause conflicts among energy policy documents. Even when faced with energy policy texts that are inconsistent in format, expression, or citation, it can still construct multi-level association information with high confidence, providing a high-quality input foundation for subsequent policy map construction and evolution.

[0046] In this embodiment of the application, step S2 utilizes the extracted entity information, relational clues, and vector representations to construct a knowledge graph, generating a knowledge graph structure that includes entity type, entity attributes, relational type, and confidence level.

[0047] Knowledge graph construction involves using text extraction results as a basis and employing a unified entity relationship modeling approach to complete the structured expression of energy policies.

[0048] Based on the extracted entity information and relationship clues, a knowledge graph node is constructed with energy policy documents as the entity. The node includes three types of attribute information: basic attributes, classification attributes, and derived attributes.

[0049] The basic attributes include policy document number, issuing department, issuance time and level of effectiveness; the classification attributes include energy type, jurisdiction and policy type; and the derived attributes include a list of related documents, revision history and vectorized representation.

[0050] Establish relationship edges between nodes. Relationship types include based on, detailed implementation, reference association, and clause conflict.

[0051] And configure a confidence value for each relation edge to manage the relationships during the update process.

[0052] Specifically, the knowledge graph of this invention adopts a multi-dimensional attribute-enhanced entity relationship model, and its core elements are defined as follows: Entity type: energy policy document, entity attributes are shown in Table 1.

[0053] Table 1 Entity Attribute Table

[0054]

[0055]

[0056] The relationship type is a one-way relationship record, as detailed in Table 2.

[0057] Table 2 Relationship Type Table

[0058]

[0059] It is important to note that each relationship has an independent confidence level, which is used for updating and maintaining the graph.

[0060] The graph structure features spatiotemporal dimension indexing: timeline retrieval is constructed based on policy release time, and retrieval tree is constructed based on geographical hierarchy; version control mechanism: when policy documents are revised, the graph incorporates the old version of the document into the original node information, realizing version management of policy documents.

[0061] In one alternative implementation, the knowledge graph construction employs a predefined relationship type and confidence control mechanism. When establishing structural relationships between entities, the relationship edge types are set as based on, refined implementation, reference association, and clause conflict, and a confidence value is set for each edge for subsequent validity judgment, weight adjustment, and conflict marking processing in the graph evolution process.

[0062] In another optional implementation, the knowledge graph construction includes establishing a spatiotemporal index structure. That is, when generating each policy entity node, a joint index key is constructed based on the policy release time and geocoding information to form a spatiotemporal index tree that supports time-axis retrieval and regional hierarchical access, thereby improving the retrieval efficiency and horizontal comparison capability of the graph under multi-dimensional conditions.

[0063] This invention enables structured, highly controllable, and scalable modeling of energy policy documents, enhances the ability of knowledge graphs to express changes in policy attributes and relationships, and improves the retrieval efficiency and stability of the system in typical applications such as structural analysis, citation tracing, and semantic comparison.

[0064] In this embodiment of the application, step S3 updates the knowledge graph for the addition, revision and repeal of policy documents.

[0065] The process for handling policy documents covers three scenarios: addition, revision, and repeal, and adjusts the graph relationships based on interaction rules and confidence information.

[0066] When adding new policy documents, the document number, issuing department, and issuance time of the new policy document are extracted as basic attributes of the entity for initialization. Energy policy correlation analysis is then performed to establish the basis or detailed implementation relationship with existing policies, and the correlation relationship is stored hierarchically according to confidence level.

[0067] When revising policies, new version nodes are generated and revision history information is recorded. By comparing the semantic vectors of the clauses in the new and old versions, conflicting relationship edges are identified and the confidence values ​​are adjusted.

[0068] When a policy is repealed, the repealed policy is set to a historical state, the relationship records of the policy before repeal are retained, confidence decay is performed on the downstream relationship edges associated with the policy, and the replacement basis document number is identified through semantic analysis.

[0069] Specifically, the updating of the energy policy knowledge graph is divided into three types: adding (policy documents into the database), updating (policy documents being revised), and deleting (policy documents being repealed).

[0070] New policy document processing procedures:

[0071] When a new policy is added to the database, the system automatically triggers the following process:

[0072] Entity initialization: Extract metadata (document number, issuing department, time, etc.) from policy documents to create graph nodes, and simultaneously build a spatiotemporal index (generate a joint index key according to the policy release time and geographical level).

[0073] Relationship mining: By calling the energy policy correlation analysis module, the system automatically associates the current policy with existing policies in the database based on or in terms of "implementation details" through explicit reference resolution and implicit semantic analysis.

[0074] Confidence level classification: Automatically identified associations are stored according to their confidence level (high-confidence associations take effect directly, while low-confidence associations are queued for manual review).

[0075] Dynamic index building: The domain subgraph index is automatically updated based on policy type (such as photovoltaic subsidies) to support subsequent fast retrieval.

[0076] Policy revision and update process:

[0077] When a policy document revision is detected, perform the following actions:

[0078] Version evolution management: Create new version nodes for revised policies, connect them to the original versions through "revision inheritance", and record revision summaries (e.g., the photovoltaic subsidy standard is adjusted from 0.42 yuan / kWh to 0.48 yuan / kWh).

[0079] Relationship weight adjustment: Based on the impact analysis of the revised content (such as the proportion of clauses involved and the magnitude of indicator changes), the weight value of the policy and its upstream and downstream relationships is automatically adjusted. If the weight decays to the threshold (such as <0.5), a warning of relationship failure is triggered.

[0080] Conflict clause detection: By comparing the semantic vector similarity between the old and new versions of the clauses, we can identify content that has substantial conflicts before and after the revision, automatically label the conflict, and notify the relevant policies.

[0081] Policy repeal process:

[0082] When a policy is officially repealed, the system executes:

[0083] Logical deletion mechanism: Mark the repealed policy node as a "historical version" and retain its association records before the repeal time, but stop it from participating in new association calculations.

[0084] Relationship cleanup: Iterate through all downstream policies that are linked to the repealed policy, calculate the decay of the confidence of the linked edges (e.g., decrease by 20% per month), and automatically disconnect the link when the confidence is below 0.3.

[0085] Alternative basis recommendation: For the lack of relevance due to policy repeal, the semantics of downstream policy clauses are analyzed using a large language model, and the most matching alternative basis is recommended from the existing effective policy database (the recommendation results need to be manually confirmed before they take effect).

[0086] Update verification mechanism:

[0087] The system incorporates the following safeguards to ensure update reliability:

[0088] Consistency verification rule base: Contains 200+ domain rules (such as municipal policies cannot directly cite national policies that have been repealed), and automatically performs rule verification on each update;

[0089] Automated test case set: simulates typical test cases in three scenarios: policy addition, revision, and repeal (such as provincial photovoltaic policies referencing national planning documents) to verify the completeness of the map update and the correctness of the associated logic.

[0090] In this embodiment of the application, step S4 utilizes a knowledge graph structure to execute a multi-path recall strategy that includes vertical source tracing recall and horizontal comparison recall, generating relevant recall results related to the target policy.

[0091] Based on the relationships existing in the constructed knowledge graph, vertical tracing and recall are performed, querying higher-level policies level by level along local, provincial, and national paths.

[0092] Based on the detailed implementation relationship edge, a horizontal comparison recall is performed to identify the different implementation plans among multiple local policies under the same superior policy.

[0093] By combining the vectorized representations of entities in the knowledge graph, the entity most similar to the target policy text is selected through Euclidean distance calculation, which helps to complete and compare related policy documents.

[0094] Specifically, the intelligent recall strategy of this invention adopts a multi-path hybrid recall mechanism, which achieves accurate tracing and comparative analysis of energy policies through the collaborative work of knowledge graphs and text encoding models. The specific technical approach is as follows:

[0095] Vertical source tracing recall strategy:

[0096] The core of the vertical tracing and recall strategy lies in utilizing the "based on" relationship edges and policy hierarchy features constructed in the knowledge graph to accurately trace the basis for the formulation of local policies by higher authorities. The deep synergy mechanism between this strategy and the knowledge graph illustrates:

[0097] Explicit basis chain retrieval: Directly traverse the pre-constructed basis and relation edges in the knowledge graph, tracing back level by level along the document issuance hierarchy path (local → provincial → national). For example... Figure 2 As shown.

[0098] Semantic similarity recall: This method calculates Euclidean distances between the vectorized representations of all entities in the graph and recalls the 10 policy documents with the closest distances. This approach requires a vector database as its underlying technology.

[0099] Horizontal comparison of recall strategies:

[0100] The core of the horizontal comparison recall strategy lies in utilizing the "refined implementation" relationship edges and multi-dimensional policy features constructed in the knowledge graph to accurately identify differentiated local implementation plans under the same superior policy. The deep synergy mechanism between this strategy and the knowledge graph is explained below:

[0101] Explicit basis chain recall: This involves locating the corresponding superior policy through the "based on" relationship edge constructed in the knowledge graph, and then precisely recalling the same-level policy through the "detailed implementation" relationship edge. For example... Figure 3 As shown.

[0102] Semantic similarity recall: This method calculates Euclidean distances between the vectorized representations of all entities in the graph and recalls the 10 policy documents with the closest distances. This approach requires a vector database as its underlying technology.

[0103] Example 3 is an embodiment of the present invention, which provides a method for multi-level construction and intelligent recall strategy of energy policy knowledge graph. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0104] Step 1: Preprocessing of multi-source heterogeneous data.

[0105] Step 1.1: Data Acquisition and Format Normalization.

[0106] The system automatically retrieves energy policy documents (PDF / Word format) from government portals via API, uses Apache PDFBox to parse PDF text, and employs the POI library to parse Word documents.

[0107] OCR recognition was performed on the scanned document using the PaddleOCRv2.6 model, with the text recognition confidence threshold set to 0.85.

[0108] Step 1.2: Structured cleaning and labeling.

[0109] The policy document number is accurately extracted using the regular expression \b〔\d{4}〕\d+\b, and a document number-document mapping index is established.

[0110] Paragraph-level word segmentation was performed using the spaCy_core_web_trf model.

[0111] Use the UIE model to extract metadata such as publishing organization and effective time.

[0112] Step 2: Implementation of multimodal correlation analysis.

[0113] Step 2.1: Explicit reference resolution.

[0114] Design multi-level nested regular expressions to match common policy citation formats:

[0115] Establish a confidence formula for citation relationships according to the regular expression ` / According to (([^》]{5,30}?)》(〔\d{4}〕\d+号)|([^,]{10,50}? management method)) / g`, and calculate the relationship confidence according to the number of hops in the citation path:

[0116] Ce = 0.7n

[0117] Where n is the number of hops in the citation path. For direct citation, n = 1; for indirect citation, n = 2.

[0118] Step 2.2, UIE entity relationship extraction.

[0119] Adopt the UIE-base model, inject entity hints in the energy field at the input layer, and the Schema for entity extraction by the UIE model:

[0120] schema = ["Issuing Department", "Energy Type", "Policy Category", "Applicable Region"]

[0121] Set the entity recognition threshold: When F1-score ≥ 0.82, trigger relationship extraction; otherwise, transfer it to the large language model extraction queue.

[0122] Step 2.3, large language model relationship extraction.

[0123] When the confidence of the key content extracted by the UIE model is insufficient, use the large language model to extract the key information from the policy document.

[0124] Step 2.4, text vectorization.

[0125] Use the BGE-M3Embedder model combined with the ChromaDB vector database to implement the encoding and storage of policy texts. The specific implementation is as follows:

[0126] Model selection and configuration:

[0127] Adopt the BAAI / bge-m3 multilingual embedding model (1024-dimensional output).

[0128] Set the text truncation length to 512 tokens. For the exceeded part, use the sliding window segmentation method (step size 256) to crop the text in the way of a sliding window (window size 512, step size 256):

[0129] text c hunks = [text[i:i + 512] for i in range(0, len(text), 256)]

[0130] Enable dual-mode output: dense retrieval and sparse retrieval.

[0131] Domain adaptation optimization:

[0132] A specialized corpus of energy policy data (containing 5,000 policy texts) was constructed for domain adaptation training.

[0133] Adjust the temperature parameter τ = 0.02 to enhance the distinguishability of technical terms.

[0134] Vector storage:

[0135] Initialize parameter settings, initialize the chromadb vector database and save it to a local file, and set BGE-M3-Embedder as the encoding model:

[0136]

[0137] Index optimization configuration: Set the search range of chromadb's HNSW algorithm to 200, the number of inter-layer connections to 16, and scalar quantization to 8-bit integer quantization.

[0138]

[0139] (5) Relationship mining

[0140] Based on the relationship extraction results and semantic similarity retrieval results, the policy text is fed into a large model for relation mining:

[0141] The Few-shot prompt template is shown in Table 3.

[0142] Table 3 Energy Policy Relationship Mining Prompt Template

[0143]

[0144] Select model:

[0145] DeepSeekV3, a domestically developed open-source large language model, was selected as the large model for relation mining. A distillation model with 70 bytes of parameters was called to save computing resources.

[0146] Step 3: Dynamic construction of the knowledge graph.

[0147] Step 3.1: Create nodes and relationships.

[0148] Entity node initialization rules, entity node definition, including node ID and attributes (power level, jurisdiction):

[0149]

[0150] Calculation of relation edge weights:

[0151] W r =α·C e +β·S s +γ·C llm

[0152] Among them, W r C represents the comprehensive weight value (0-1) of the relation edges, indicating the credibility of the final policy relationship. e To determine the association confidence based on explicit citations, S is obtained by parsing the cited statements in the original policy text using regular expressions. s To calculate the cosine similarity score (0-1) based on text vectors, the C value is calculated using the BGE-M3 model. llm The association confidence scores determined by the large language model are obtained by analyzing the implicit logic of the large language model, with coefficients α = 0.4, β = 0.3, and γ = 0.3.

[0153] Step 3.2, Spatiotemporal index construction.

[0154] Timeline Index: A B+ tree index is built according to the policy effective date, supporting range queries (such as photovoltaic policies from 2020 to 2023).

[0155] Geographic hierarchical index: The administrative regions are encoded using a quadtree structure (national level → provincial level → municipal level), and the leaf nodes store the policy density heat values.

[0156] Step 3.3, Version Management.

[0157] Design revision impact assessment model:

[0158]

[0159] Among them, I r The policy change impact index quantifies the actual impact of policy revisions on relevant industries (range 0-1), N. changed N represents the number of clauses that have actually changed after the policy revision. total The total number of policy clauses, the total number of clauses in the current policy text, Δt is the change time interval, the number of days since the policy revision took effect, and log is a function used to calculate the decay of policy impact over time. r A recalculation of the association relationship is triggered when the value is ≥0.5.

[0160] Conflict clause detection: Use BGE-M3 to compare the semantic vectors of the old and new versions, and generate a conflict warning when the cosine similarity is <0.6.

[0161] Step 4: Execute the intelligent recall strategy.

[0162] Step 4.1: Vertical tracing and recall.

[0163] The path backtracking algorithm identifies the superior policy (the parent node in the "based on" relationship) of a given policy:

[0164]

[0165] The maximum backtracking depth is set to 5 levels. When the depth exceeds the specified level or no results are found, semantic compensation recall will be initiated.

[0166] Semantic Recall Algorithm:

[0167] The BGE-M3 model was used to compute the vector representation of the policy text.

[0168] Perform a cross-level semantic search in ChromaDB, with a conditional vector similarity query. The conditions are: Policy effectiveness level: National level, Energy type: Photovoltaic.

[0169]

[0170] Step 4.2: Horizontal comparison recall.

[0171] Algorithm for positioning superior policies:

[0172] A precise search was performed using the graph query engine to find the policy parent node (formulation basis) with doc_id of National Energy Administration

[2022] No. 1 in the knowledge graph:

[0173] MATCH(parent:Policy{doc_id:'National Energy Administration

[2022] No. 1'})

[0174] RETURNparent

[0175] Policy findings at the same level:

[0176] Traversing sibling nodes along the "Detailed Implementation" relationship, the Cypher query is used to find differentiated policy implementation plans at the same level. Tracing upwards: `current` is based on `parent`, expanding downwards: `parent` detailed implementation siblings, retrieving 50 sibling policies (siblings) of the same type (policy_type) as the current policy (`current`) but in different regions (`region`).

[0177] MATCH(parent) <- [:based on] - (current)

[0178] MATCH(parent) - [:Detailed Implementation] -> (sibling)

[0179] WHEREsibling.region<>current.region

[0180] ANDsibling.policy_type=current.policy_type

[0181] RETURNsibling

[0182] LIMIT50

[0183] Semantic Recall Algorithm:

[0184] The BGE-M3 model was used to compute the vector representation of the policy text.

[0185] Perform cross-level semantic search and multi-condition vector retrieval in ChromaDB to find the 5 most matching national-level photovoltaic policies in the policy knowledge base using a given vector `param_vector`. First, perform precise filtering based on the metadata fields of effectiveness level and energy type (requiring the effectiveness level to be "national-level" and the energy type to be "photovoltaic"). Then, calculate the similarity between each policy document and the input vector in the filtered results, and finally return the 5 policy records with the highest similarity. This is represented as:

[0186]

[0187] Reordering Algorithm:

[0188] The recall policy text was reordered using the BAAI / bge-reranker-large model.

[0189] Input processing: Concatenate the policy text pairs (current policy vs. candidate policy) from the recall results into:

[0190] """

[0191] [Current Policy]: {Policy Title}\n{Summary of Core Clauses}

[0192] [Candidate Policy]: {Policy Title}\n{Relevance Explanation}\n{Technical Indicator Comparison}

[0193] """

[0194] The truncation length is set to 1024 tokens, and the first and last key paragraphs are retained for any excess.

[0195] Traditional policy retrieval relies on manual step-by-step backtracking and cross-departmental document verification, with a single retrieval taking an average of 3-5 hours. This invention achieves automatic penetration of the policy hierarchy transmission link through a multi-level association parsing mechanism and spatiotemporal dimension indexing technology. The explicit relational edge traversal algorithm based on knowledge graphs can compress the vertical retrieval response time from national policies to local regulations to within 10 seconds, improving efficiency by over 1000 times. Simultaneously, the semantic similarity pre-screening module reduces the frequency of large language model calls, and the candidate set generation efficiency for horizontal comparison tasks is improved by 85% compared to traditional methods. Experiments show that in a test scenario involving 5 policy levels and 2000+ related documents, the system's end-to-end retrieval time is consistently below 300 seconds (5 minutes), meeting the needs of high-concurrency business operations.

[0196] Traditional keyword matching methods achieve an accuracy of less than 62% in cross-regional policy comparisons. This invention, through the synergistic effect of a multimodal association engine (explicit reference resolution + semantic vector matching + large language model inference), improves the accuracy of vertical source tracing to 96.7%. In horizontal comparison scenarios, a hybrid recall strategy based on knowledge graph attribute constraints (such as level of effectiveness and jurisdiction) achieves an identification accuracy of 91.2% for differentiated clauses of similar policies, which is 42 percentage points higher than single text similarity methods.

[0197] Example 4, refer to Figure 4 This embodiment of the present invention provides a multi-level construction and intelligent recall strategy implementation system for an energy policy knowledge graph, including an energy policy correlation analysis module, an energy policy knowledge graph maintenance module, and an intelligent recall module.

[0198] The Energy Policy Correlation Analysis module is used to acquire energy policy text data, conduct energy policy correlation analysis, and extract entity information, relationship clues, and text vector representations from the energy policy text.

[0199] The energy policy correlation analysis module is used to construct a knowledge graph using extracted entity information, relationship clues, and vector representations, generating a knowledge graph structure that includes entity type, entity attributes, relationship type, and confidence level; and updates the knowledge graph for the addition, revision, and repeal of policy documents.

[0200] The intelligent recall module utilizes a knowledge graph structure to execute a multi-path recall strategy, including vertical source tracing recall and horizontal comparison recall, to generate relevant recall results related to the target policy.

[0201] This embodiment also provides an electronic device applicable to the implementation method of multi-level construction and intelligent recall strategy for energy policy knowledge graph, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the multi-level construction and intelligent recall strategy for energy policy knowledge graph as proposed in the above embodiment.

[0202] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the multi-level construction and intelligent recall strategy implementation method for the energy policy knowledge graph proposed in the above embodiments.

[0203] The storage medium proposed in this embodiment and the method for implementing the multi-level construction and intelligent recall strategy of energy policy knowledge graph proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0204] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0205] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for multi-level construction and intelligent recall strategy implementation of energy policy knowledge graph, characterized by: include, Acquire energy policy text data, conduct energy policy correlation analysis, and extract entity information, relationship clues, and text vector representations from the energy policy texts; Knowledge graphs are constructed using extracted entity information, relational clues, and vector representations, generating knowledge graph structures that include entity types, entity attributes, relational types, and confidence levels. The knowledge graph is updated in response to the addition, revision, and repeal of policy documents; By leveraging the knowledge graph structure, a multi-path recall strategy, including vertical source tracing and horizontal comparison recall, is implemented to generate relevant recall results related to the target policy.

2. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 1, characterized in that: The energy policy correlation analysis includes using a combination of various analytical methods to process energy policy texts, thereby expanding the coverage of correlation judgments.

3. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 2, characterized in that: The knowledge graph construction involves using text extraction results as a basis and employing a unified entity relationship modeling approach to complete the structured expression of energy policies.

4. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 3, characterized in that: The knowledge graph update includes the processing flow of policy documents, covering three scenarios: addition, revision, and repeal, and adjusts the graph relationships based on interaction rules and confidence information.

5. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 4, characterized in that: The knowledge graph construction also includes constructing knowledge graph nodes with energy policy documents as entities based on the extracted entity information and relationship clues. The nodes include three types of attribute information: basic attributes, classification attributes, and derived attributes. The basic attributes include policy document number, issuing department, issuance time and level of effectiveness; the classification attributes include energy type, jurisdiction and policy type; and the derived attributes include a list of related documents, revision history and vectorized representation. Establish relationship edges between nodes, with relationship types including based on, detailed implementation, reference association, and clause conflict; And a confidence value is assigned to each relation edge to manage the relationships during the update process.

6. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 4, characterized in that: The knowledge graph update also includes, when adding a new policy document, extracting the document number, issuing department, and release time of the new policy document as basic entity attributes for initialization, performing energy policy correlation analysis, establishing or refining the implementation relationship with existing policies, and storing the correlation relationship in a hierarchical manner according to confidence level. When revising policies, new version nodes are generated and revision history information is recorded. By comparing the semantic vectors of the clauses in the new and old versions, conflicting relationship edges are identified and the confidence values ​​are adjusted. When a policy is repealed, the repealed policy is set to a historical state, the relationship records of the policy before repeal are retained, confidence decay is performed on the downstream relationship edges associated with the policy, and the replacement basis document number is identified through semantic analysis.

7. The method for implementing multi-level construction and intelligent recall strategy of energy policy knowledge graph as described in claim 4, characterized in that: The multi-path recall strategy includes performing vertical source tracing recall based on the relationship edges existing in the constructed knowledge graph, querying higher-level policies level by level along local, provincial, and national paths; Based on the detailed implementation relationship edge, a horizontal comparison recall is performed to identify the different implementation plans among multiple local policies under the same superior policy; By combining the vectorized representations of entities in the knowledge graph, the entity most similar to the target policy text is selected through Euclidean distance calculation, which helps to complete and compare related policy documents.

8. A system for implementing a multi-level construction and intelligent recall strategy for an energy policy knowledge graph, comprising applying the method for implementing a multi-level construction and intelligent recall strategy for an energy policy knowledge graph as described in any one of claims 1 to 7, characterized in that, include: The module includes an energy policy correlation analysis module, an energy policy knowledge graph maintenance module, and an intelligent recall module. The energy policy correlation analysis module is used to acquire energy policy text data, perform energy policy correlation analysis, and extract entity information, relationship clues, and text vector representations from the energy policy text. The energy policy correlation analysis module is used to construct a knowledge graph using extracted entity information, relationship clues, and vector representations, generating a knowledge graph structure that includes entity type, entity attributes, relationship type, and confidence level; and updates the knowledge graph for the addition, revision, and repeal of policy documents. The intelligent recall module is used to utilize a knowledge graph structure to execute a multi-path recall strategy, including vertical source tracing recall and horizontal comparison recall, to generate relevant recall results related to the target policy.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for implementing the multi-level construction and intelligent recall strategy of the energy policy knowledge graph as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for implementing the multi-level construction and intelligent recall strategy of the energy policy knowledge graph as described in any one of claims 1 to 7.

Images