Dynamic knowledge distillation method and system based on knowledge graph

Abstract

The invention belongs to the field of artificial intelligence and knowledge engineering, and discloses a dynamic knowledge distillation method and system based on a knowledge graph, which are used for extracting and reconstructing high-value knowledge from structured or unstructured documents. Aiming at the existing problems of static processing, semantic segmentation, recombination stiffness, cross-document association missing and the like, the method realizes dynamic distillation through the following steps: generating separators in real time based on a domain knowledge graph core entity set, breaking through a fixed partitioning rule, and adapting to multi-type document structures; a concept is utilized to trigger a correlation decision model to filter non-correlation texts, and semantic fragmentization is reduced; loading an adaptive prompt template according to the domain attribute of the knowledge graph, and fitting the characteristics of the fields of medical treatment, education and the like; knowledge is dynamically reconstructed through a logic framework defined by a knowledge graph, and cross-document knowledge point semantic association (such as policy document complementary clause linkage) is achieved. According to the system, through dynamic blocking, semantic filtering, field adaptation and logic recombination, the limitation of a traditional technology is solved, the flexibility and field adaptability of knowledge purification are remarkably improved, and the system is suitable for high-value knowledge mining and structured output of massive documents.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and knowledge engineering, and specifically relates to a dynamic knowledge distillation method and system based on knowledge graph, which is suitable for extracting and reconstructing high-value knowledge from structured or unstructured documents. Background Art

[0002] Knowledge distillation refers to the processing and purification of information. In the information age, especially with the widespread application of artificial intelligence technology and the face of massive amounts of knowledge, information processing and knowledge purification are becoming increasingly important. Existing knowledge purification and processing technologies have the following shortcomings: 1. Often adopt static processing methods: Traditional methods use fixed block rules (such as segmentation by chapter), which cannot adapt to various document structures; 2. Semantic fragmentation: Lack of domain knowledge guidance leads to fragmentation of key concepts (e.g., "credit recognition" in academic documents is handled separately); 3. Restructuring Rigidity: Output follows a general template, ignoring domain specificity (e.g., medical care requires a "cause → diagnosis → treatment" logic); 4. Lack of cross-document association: It is impossible to establish semantic associations between scattered knowledge points (such as complementary clauses in different policy documents). Summary of the Invention

[0003] To address the above problems, it is necessary to provide a dynamic knowledge distillation method based on knowledge graph to adapt to various document types, solve the problem of knowledge fragmentation, and adapt to the characteristics of different fields.

[0004] In addition, it is necessary to provide a dynamic knowledge distillation system based on knowledge graph to adapt to various document types, solve the problem of knowledge fragmentation, and be able to adapt to the characteristics of different fields.

[0005] A dynamic knowledge distillation method based on a knowledge graph, a dynamic knowledge distillation method based on a knowledge graph, (S1) generates document delimiters in real time based on the core entity set of the domain knowledge graph, and divides the text into blocks based on the delimiters, wherein the dynamic nature is reflected in the generation of block rules driven by the knowledge graph; (S2) uses a concept-triggered relevance decision model to filter non-relevant text blocks; (S3) loads a domain-adapted large language model distillation prompt template based on the domain attributes of the knowledge graph, and uses the large language model to compress the text; (S4) dynamically reconstructs and saves the output knowledge according to the logical framework defined by the knowledge graph, wherein the dynamic nature is reflected in the switching of the reorganization logic driven by the knowledge graph; wherein the dynamic nature specifically refers to the ability of the knowledge graph to generate block rules and reorganization logic in real time, rather than data weight adjustment. The real-time generation of document delimiters includes: (S11) extracting core entities and their aliases from the knowledge graph to form an entity set; (S12) generating a delimiter in the form of a regular expression for each entity; (S13) adding structural delimiters predefined in the knowledge graph; and (S14) merging the above delimiters to form a final delimiter set.

[0006] The calculation method of the concept-triggered relevance decision model is: relevance value = parameter α × BERT semantic similarity + parameter β × knowledge graph centrality + parameter γ × timeliness factor, where: (S21) BERT semantic similarity: the semantic matching degree between the concept and the text; (S22) knowledge graph centrality: the intermediary centrality weight of the concept in the knowledge graph; (S23) timeliness factor: the timeliness weight assigned to policy concepts, with a value range of 0 to 1.

[0007] Domain-adapted distillation prompt templates include: (S31) Medical Domain Template: Guides the extraction of etiology mechanisms, diagnostic criteria, and treatment plans; (S32) Academic Affairs Domain Template: Guides the extraction of institutional basis, process nodes, and risk control thresholds; (S33) Legal Domain Template: Guides the extraction of legal provisions, judicial interpretations, and sentencing standards. The logical framework is as follows: (S41) The Academic Affairs Domain adopts a four-layer structure of "institutional basis → process topology → implementation details → risk control red lines"; (S42) The Medical Domain adopts a causal chain of "etiology mechanisms → diagnostic criteria → treatment plans → medication contraindications"; and (S43) The Engineering Domain adopts a workflow of "design specifications → construction standards → acceptance indicators → accident response plans."

[0008] Conflict resolution is performed in the reorganization phase: (S51) semantically conflicting clauses in the knowledge output set are identified; (S52) when the knowledge graph determines that two clauses are contradictory, their authority values ​​are compared; (S53) the clause with higher authority is retained and the clause with lower authority is discarded; (S54) the authority is a graded value based on the reliability of the knowledge graph source, and the value range is 1 to 5.

[0009] The large language model is also referred to as the large model or LLM, and the open source large model Qwen-72B can be used.

[0010] The present invention also provides a dynamic knowledge distillation system, comprising: (1) a knowledge graph management module (201) for storing entities, relationships and domain templates; (2) a dynamic block engine (202) for realizing the generation of separators as described in claim 2; (3) a concept filtering engine (203) for realizing the correlation calculation as described in claim 3; (4) a domain distillation module (204) for loading the prompt template in the knowledge graph to control the large language model; (5) a logic reorganization module (205) for reconstructing the knowledge flow according to the framework defined by the knowledge graph; the "dynamic" specifically refers to the knowledge graph guiding the generation of block rules and the switching of reorganization logic in real time.

[0011] The knowledge graph management module (201) includes: (1) a time-weighted labeling unit (201a) for marking the effective / ineffective time of policy entities; (2) an authority calculation unit (201b) for setting an authority value from 1 to 5 according to the source of the entity; and (3) a domain template warehouse (201c) for storing distilled prompt templates in the fields of medical care, education, and law.

[0012] The logic reorganization module (205) includes: (1) a frame selector (205a) for selecting reorganization rules according to the knowledge graph domain label; (2) a conflict detector (205b) for implementing the conflict resolution described in claim 6; and (3) a knowledge stitcher (205c) for integrating related terms across documents into a continuous knowledge flow.

[0013] The present invention also provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the program, a dynamic knowledge distillation method based on a knowledge graph is implemented.

[0014] The beneficial effects of this invention lie in overcoming the limitations of traditional technologies through dynamic segmentation, semantic filtering, domain adaptation, and logical reorganization. This significantly improves the flexibility and domain adaptability of knowledge extraction, making it suitable for high-value knowledge mining and structured output from massive amounts of documents. Through the dual engines of "dynamic knowledge graph guidance" and "domain logical reorganization," this approach addresses the three major pain points of traditional distillation: rigid rules, unsolvable conflicts, and difficulty in domain migration. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The accompanying drawings are primarily intended to illustrate preferred embodiments and are not intended to limit the present invention. Throughout the accompanying drawings, the same reference symbols denote the same components. In the accompanying drawings: Figure 1 This is a flow chart of a dynamic knowledge distillation method based on a knowledge graph according to an embodiment of the present invention; Figure 2 It is an algorithm flow chart of the concept triggering correlation decision model of the embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and beneficial effects of the present invention clearer and more understandable, the present invention is further described below with reference to the accompanying drawings and embodiments. Example 1

[0017] like Figure 1 The figure shows a dynamic knowledge distillation method based on a knowledge graph, comprising: (S1) KG-guided chunking: document delimiters are generated in real time based on the core entity set of the domain knowledge graph, and the text is chunked according to the delimiters. The dynamic nature is reflected in the generation of chunking rules driven by the knowledge graph. KG in the figure is the abbreviation of knowledge graph. (S2) Concept-triggered filtering: a concept-triggered relevance decision model is used to filter non-relevant text chunks; (S3) LLM knowledge distillation: domain-adapted LLM distillation prompt templates are loaded based on the domain attributes of the knowledge graph, and the text is compressed using LLM; (S4) logical reorganization: output knowledge is dynamically reconstructed and saved according to the logical framework defined by the knowledge graph. The dynamic nature is reflected in the switching of reorganization logic driven by the knowledge graph. The dynamic nature specifically refers to the ability of the knowledge graph to generate chunking rules and reorganization logic in real time, rather than data weight adjustment. The real-time generation of document delimiters includes: (S11) extracting core entities and their aliases from the knowledge graph to form an entity set; (S12) generating a delimiter in the form of a regular expression for each entity; (S13) adding predefined structural delimiters in the knowledge graph; (S14) merging the above delimiters to form a final delimiter set.

[0018] The calculation method of the concept-triggered relevance decision model is: relevance value = parameter α × BERT semantic similarity + parameter β × knowledge graph centrality + parameter γ × timeliness factor, where: (S21) BERT semantic similarity: the semantic matching degree between the concept and the text; (S22) knowledge graph centrality: the intermediary centrality weight of the concept in the knowledge graph; (S23) timeliness factor: the timeliness weight assigned to policy concepts, with a value range of 0 to 1.

[0019] Domain-adapted distillation prompt templates include: (S31) Medical Domain Template: Guides the extraction of etiology mechanisms, diagnostic criteria, and treatment plans; (S32) Academic Affairs Domain Template: Guides the extraction of institutional basis, process nodes, and risk control thresholds; (S33) Legal Domain Template: Guides the extraction of legal provisions, judicial interpretations, and sentencing standards. The logical framework is as follows: (S41) The Academic Affairs Domain adopts a four-layer structure of "institutional basis → process topology → implementation details → risk control red lines"; (S42) The Medical Domain adopts a causal chain of "etiology mechanisms → diagnostic criteria → treatment plans → medication contraindications"; and (S43) The Engineering Domain adopts a workflow of "design specifications → construction standards → acceptance indicators → accident response plans."

[0020] Conflict resolution is performed in the reorganization phase: (S51) semantically conflicting clauses in the knowledge output set are identified; (S52) when the knowledge graph determines that two clauses are contradictory, their authority values ​​are compared; (S53) the clause with higher authority is retained and the clause with lower authority is discarded; (S54) the authority is a graded value based on the reliability of the knowledge graph source, and the value range is 1 to 5.

[0021] In step (S3), the LLM is a large language model or simply a large model, and the open source large model Qwen-72B is used.

[0022] Specifically, the domain knowledge graph KG = (E, R, A) is initialized, where E represents the core entity set (e.g., "credit recognition" and "student status change" in the academic affairs domain), R represents the entity relationship network, and A represents the dynamic rule base (including structural delimiters, domain prompt templates, and logical framework mapping). When a new entity enew is added to the graph (e.g., "cross-border online courses"), the system automatically generates a regular expression delimiter (e.g., _) based on the entity's characteristics and combines it with pre-existing rules to form a dynamic segmentation rule set Rdynamic = Rold ⊕ rdelim.

[0023] The document segmentation phase uses a recursive segmentation algorithm to process the input text according to rule priority. Taking a 300-page academic manual as an example, the system uses entity linking to identify the delimiters corresponding to entities ei (e.g., "credit recognition"), accurately segmenting the relevant sections and generating N=85 semantically coherent text blocks {C1, C2, ..., CN}. Concept-triggered filtering is then performed: for each text block Cj, its relevance value to the graph entity ei is calculated.

[0024] Among them, sim sem Cosine similarity based on BERT embedding

[0025] Where α, β, γ are weight coefficients (α+β+γ=1), which control the influence of semantic similarity, entity centrality, and timeliness factor respectively.

[0026] sim sem (Cj,ei): The semantic similarity between the text block Cj and the entity ei, calculated by the cosine similarity of the BERT embedding vector.

[0027] PR(ei): The PageRank centrality of entity ei in the knowledge graph, which measures the structural importance of the entity.

[0028] τ: Timeliness factor (0≤τ≤1), reflecting the timeliness of the text content (such as policy release time, data update cycle).

[0029] PR(ei) is the PageRank centrality of the entity in the KG graph, and τ is the time decay factor. We set a threshold of θ = 0.7 and retain text blocks with Rj > θ. For example, a text block containing "2024 Edition of Degree Standards" is retained with τ = 1.0, while historical policies are filtered out with τ = 0.3.

[0030] In the distillation phase, dynamically load the domain template Td (d is the graph domain label): when d="education management", enable template extraction policy basis compression to. Input the filtered text block into the large language model LLM (such as Qwen-72B), and minimize the distillation loss.

[0031] To achieve knowledge compression, the original 500-word "deferred examination approval" clause was refined into a 150-word core process.

[0032] Restructuring phase response map domain label d: The educational affairs domain uses the four-layer framework Fedu = {system basis, process topology, implementation details, risk control red line}. In this process, semantic conflict clauses (such as clauses s and t conflict) are detected and the authority decision function is used. Keep(s,t)={stif Auth(s)>Auth(t)otherwise Where Auth(x) is the authority of the source of the clause (Auth from the Ministry of Education = 5 > Auth from the college = 3). Finally, a structured knowledge flow is constructed.

[0033] When outputting knowledge, it is saved in the node corresponding to the framework. The framework can be a JSON file isomorphic to the knowledge graph, or it can be saved directly in the JSON file of the knowledge graph. Example 2

[0034] Corresponding documents: "2024 Undergraduate Student Registration Management Measures", "Graduation Audit Implementation Rules" and other 6 documents (a total of 183 pages), for knowledge distillation, the following steps can be adopted.

[0035] Step 1: Dynamically generate document separators (S101) Knowledge graph entity extraction Core entity sets: Vcore={credit recognition, deferred exam approval, student status change, graduation review, course replacement} Entity Alias: "Credit Recognition" → ["Credit Transfer", "Credit Recognition Standards"] "Graduation Review" → ["Graduation Qualification Review", "Degree Review"] Regular delimiter generation

[0036] Structural delimiters added Chapter Chapter Final delimiter set Ssep={credit recognition rule, deferred examination approval rule, …, Sstruct} Chunking effect: splits a 183-page document into 62 semantic chunks (e.g., "Credit Recognition - International Courses" chunk, "Graduation Review - Core Course Requirements" chunk) Step 2: Concept triggers relevance decision model filtering (S102) Target text block: "Applications for international course credit recognition must be submitted 15 working days before the end of the semester. Late applications will not be accepted." Correlation calculation (parameters: α=0.6, β=0.3, γ=0.1) BERT sim ("credit recognition", text block) = 0.92 (including the keyword "credit recognition") Centrality("credit recognition")=0.85 (core node) T fresh ("Credit Recognition") = 1.0 (New policy in 2024) Relevance=0.6×0.92+0.3×0.85+0.1×1.0=0.907 1) Filtering decisions 0.907>θ(0.7) → keep this text block Filtering results: 38 of 62 text blocks were retained (24 outdated policy / explanatory content was discarded) Step 3: Domain Adaptation Distillation (S103) 2) Load the academic affairs template Extract: 1) System basis (marked with document number) 2) Process nodes 3) Risk control thresholds, delete examples, compress to 30% volume 3) LLM distillation treatment Enter a text block: According to Article 5.2 of the "Student Registration Management Regulations," applications for international course credit recognition must be submitted 15 working days prior to the end of the semester. Applications submitted after the deadline will be deemed as automatic withdrawal. Original copies of the course syllabus and transcripts are required; electronic copies will not be accepted. Distillation output: json { "System Basis": "Article 5.2 of the "Student Registration Management Regulations"", "Process Node": "Submit application 15 working days before the end of the semester", "Risk Control Threshold": "Overdue documents are invalid, only paper documents will be accepted" } Compression rate: Original text 128 words → Distilled text 36 words (28%) Step 4: Logical Framework Reorganization (S104) 1) Reorganization according to the academic affairs framework Source file of knowledge fragments embedded in framework nodes The system is based on Article 5.2 of the "Student Registration Management Regulations" (Student Registration Management Regulations.pdf). Submit the International Course Implementation Rules.docx 15 working days before the end of the semester. Risk control red line expires and is invalid + only paper materials are accepted Student Registration Management Measures.pdf 2) Conflict Detection and Arbitration Detecting conflicts: Document A: "Paper documents require the college's seal" (Auth=4) File B: "Electronic signature valid" (Auth=2) Authority judgment: max(Auth)=4→Retention requires the college's seal 3) Reorganization output: json { "System Basis": ["Student Registration Management Regulations" Article 5.2], "Process topology": ["Submit application 15 working days before the end of the semester"], "Implementation Details": ["Submit application form to the International College"], "Risk Control Red Line": ["Invalid after the deadline", "Paper materials + college seal required"] } Dynamic performance example 1. Add real-time response to new entities Event: Added "Online Exam Violation Handling" entity to the knowledge graph Dynamic response: Automatically generate separators:

[0037] When processing the new document "Online Examination Management Measures" later, the "Chapter 4 Violation Handling" will be automatically divided into blocks 2. Domain switching takes effect immediately Event: Knowledge graph label switched from "edu" to "law" Dynamic response: Enable legal templates: "Extract legal provisions, judicial interpretations, and sentencing standards" Switch the reorganization framework to: "Law → Judicial Interpretation → Case Reference" The dynamic response mechanism is the core innovation of this method: When a new entity enew is added, the block rule set is updated to Rnew=R∪{rdelim(enew)} When switching fields, the template switches from Td to Td' and the frame switches from Fd to Fd' No manual parameter adjustment is required during the entire process, and the performance improvement is as follows: Block accuracy ΔF1=0.92-0.68=+0.24 Time optimization toldtold-tnew=142142-67≈53% Cross-document correlation Number of documents Number of terms = 83327 ≈ 3.94 terms / document. Example 3

[0038] This example mainly introduces the document distillation process for textbooks. Step 1: Graph initialization Load the physical experiment knowledge graph KG=(E,R,A), where: Core entity set E = {"Rigid body moment of inertia", "Young's modulus determination", ...} Preset Structure Delimiters: Experimental Step 2: Dynamic Chunking When adding the entity enew="Hall Effect Experiment", a separator is generated: Hall effect experiment The positioning function Locate(enew,text) based on the entity regular expression is used to accurately divide the blocks in the "Experiment 3.8 Hall Effect" section.

[0039] Step 3: Concept Filtering The cosine similarity calculation formula is used to calculate the semantic similarity between the text block Cj and the entity ei:

[0040] vCj: BERT embedding vector (high-dimensional semantic space vector) of text block Cj.

[0041] vei: BERT embedding vector of entity ei (usually generated from the entity name or description).

[0042] T: Vector transpose symbol, indicating inner product operation.

[0043] ∥•∥: L2 norm (vector modulus), used to normalize vector length.

[0044] Value range: [0,1], the larger the value, the higher the semantic similarity.

[0045] Example: The "Application of Gauss's Theorem" text block is retained because sim(Cj,"Electrostatic Experiment")=0.85>θsim.

[0046] Step 4: Domain Distillation Enable the physical realm template: Tphy="Extract the experimental purpose, principle formula, instrument list, and core steps" Distillation effect (taking "Young's modulus determination" as an example): Original text: 600 words → Distilled to 180 words (70% compression) Keep the key formula: ΔL=πr2EFL (Hooke's law variant) Purpose: To measure the elastic modulus of metal wire Instrument: Optical lever + weight set Steps: Install the specimen → Load and measure ΔL → Calculate E Step 5: Logical Reorganization Reorganize by physical frame: Fphy={experimental purpose, theoretical formula, instrument, steps, data processing} Conflict resolution: When two error algorithms conflict, the authority of the graph Auth is used for arbitration (e.g., national standard Auth=5 takes precedence over industry standard Auth=3).

[0047] Key symbol description Gphy: Physics Experiment Knowledge Graph vCj,vei: BERT embedding vectors of text blocks and entities θsim: semantic similarity threshold (usually 0.7) ΔL: wire elongation, F: tension, L: original length of wire, r: radius, E: elastic modulus. Example 4

[0048] This embodiment introduces a specific implementation of the correlation decision model, which adopts "correlation matrix construction + threshold decision", such as Figure 2 The flowchart of the relevance decision model is shown in FIG. The specific steps of the concept triggering relevance decision model to filter non-relevant text blocks include: (a) Concept-text correlation matrix construction: Mc×t=[R(ci,tj)] in: ci∈VKG: entity set in the knowledge graph tj: input text block collection R(ci,tj): The correlation between entity ci and text block tj (calculated using the formula R=α•BERTsim+β•Centrality+γ•Tfresh) (b) Threshold decision rule tj is retained if and only if maxciR(ci,tj)>θ Logic: For each text block tj, calculate its maximum correlation value with all entities. If it exceeds the threshold θ (such as 0.7), it is determined to be a relevant text block.

[0049] Example: If R=0.907>0.7 between tj and the entity “credit recognition”, then retain the text block.

[0050] (c) Dynamic update mechanism When VKG adds or deletes entities, Mc×t←Rebuild(VKGnew,T) Trigger condition: Changes to the knowledge graph entity set VKG (such as the addition of the "Online Exam Violation Handling" entity).

[0051] Operation: Recalculate the relevance values ​​of all entities and text blocks in real time, rebuild the matrix Mc×t, and ensure that the filtering logic is always based on the latest knowledge system.

[0052] The correlation matrix calculation includes three-channel weighting: R(ci,tj)=semantic channel α•BERTsim(ci,tj)+graph structure channel β•Centrality(ci)+temporal channel γ•Tfresh(ci).

[0053] The weight coefficient satisfies: α+β+γ=1 Parameter Description: α: semantic similarity weight (controls the importance of BERT semantic matching) β: Knowledge graph centrality weight (controls the importance of entity structure) γ: Timeliness factor weight (controls the impact of content timeliness) The dynamic threshold setting rule of the threshold θ is: θ=0.6+0.2•m∑i=1mCentrality(ci) explain: The basic threshold is 0.6 The dynamic adjustment term is proportional to the average centrality of the knowledge graph entity The higher the centrality of the entity set, the closer the threshold is to 0.8 The dynamic update mechanism of the correlation matrix, whose response events include three-channel weighting: 1. Adding new entity vnew to the knowledge graph M←[M[R(vnew,t1),…,R(vnew,tn)]] 2. Knowledge graph deletion entity vdel M←M\Row(vdel) 3. Add a new text block tadd to the document M←[M[R(c1,tadd),…,R(cm,tadd)]T].

Claims

1. A dynamic knowledge distillation method based on knowledge graph, characterized by include: (S1) Generating document delimiters in real time based on the core entity set of the domain knowledge graph, and dividing the text into blocks based on the delimiters. The dynamic nature is reflected in the generation of block rules driven by the knowledge graph; (S2) Using the concept-triggered relevance decision model to filter out irrelevant text blocks; (S3) Loading domain-adapted large language model distillation prompt templates based on domain attributes of the knowledge graph, and compressing text using the large language model; (S4) dynamically reconstructing and saving the output knowledge according to the logical framework defined by the knowledge graph, wherein the dynamic nature is reflected in the switching of the reorganization logic driven by the knowledge graph; The dynamic nature specifically refers to the ability of the knowledge graph to generate segmentation rules and reorganization logic in real time, rather than data weight adjustment.

2. The method according to claim 1, characterized in that The real-time generation of document delimiters includes: (S11) extracting core entities and their aliases from the knowledge graph to form an entity set; (S12) generating a delimiter in the form of a regular expression for each entity; (S13) Adding predefined structural separators in the knowledge graph; (S14) Merge the above delimiters to form a final delimiter set.

3. The method according to claim 1, characterized in that The calculation method of the concept trigger relevance decision model is: Relevance value = parameter α × BERT semantic similarity + parameter β × knowledge graph centrality + parameter γ × timeliness factor, where: (S21) BERT semantic similarity: semantic matching between concepts and text; (S22) Knowledge graph centrality: the betweenness centrality weight of a concept in the knowledge graph; (S23) Timeliness factor: The timeliness weight assigned to policy concepts, with a value range of 0 to 1.

4. The method according to claim 1, characterized in that The distillation prompt templates adapted for the field include: (S31) Medical field template: guides the extraction of etiology mechanisms, diagnostic criteria, and treatment plans; (S32) Academic Affairs Template: Provides guidance on extracting system basis, process nodes, and risk control thresholds; (S33) Legal field template: provides guidance for extracting legal requirements, judicial interpretations, and sentencing standards.

5. The method according to claim 1, characterized in that The logical framework is: (S41) The academic affairs field adopts a four-layer structure of "system basis → process topology → implementation details → risk control red lines"; (S42) The medical field adopts the causal chain of "etiology mechanism → diagnostic criteria → treatment plan → medication contraindications"; (S43) The engineering field adopts the "design specifications → construction standards → acceptance indicators → accident plan" workflow.

6. The method according to claim 5, characterized in that Perform conflict resolution during the reorganization phase: (S51) identifying semantically conflicting clauses in the knowledge output set; (S52) When the knowledge graph determines that two clauses are contradictory, their authority values ​​are compared; (S53) retain the clauses with higher authority and discard the clauses with lower authority; (S54) The authority is a graded value based on the reliability of the knowledge graph source, and the value range is 1 to 5.

7. A dynamic knowledge distillation system based on knowledge graph, characterized by include: A knowledge graph management module (201) is used to store entities, relationships and domain templates; A dynamic block engine (202) for implementing the delimiter generation as claimed in claim 2; A concept filtering engine (203), configured to implement the correlation calculation described in claim 3; The domain distillation module (204) is used to load the prompt template in the knowledge graph to control the large language model; A logic reorganization module (205) is used to reconstruct the knowledge flow according to the framework defined by the knowledge graph; The "dynamic" mentioned above specifically refers to the real-time guidance of block rule generation and reorganization logic switching by the knowledge graph.

8. The system according to claim 7, characterized in that The knowledge graph management module (201) includes: The timeliness weight marking unit (201a) marks the effective / ineffective time of policy entities; An authority calculation unit (201b) sets an authority value from 1 to 5 according to the entity source; Domain Template Repository (201c), stores distilled prompt templates in the medical, educational, and legal fields.

9. The system according to claim 7, characterized in that The logic reorganization module (205) includes: Frame selector (205a), which selects reorganization rules based on the knowledge graph domain label; A conflict detector (205b) for implementing the conflict resolution as claimed in claim 6; The knowledge stitcher (205c) integrates related terms across documents into a continuous knowledge stream.

10. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images