Knowledge graph dynamic construction and evolution method based on LLM and KG collaboration

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a knowledge graph dynamic construction and evolution method based on LLM and KG collaboration. The method comprises the following steps: an intelligent perception step: generating a structured knowledge proposal from multi-source data by using a large language model; a collaborative verification step of performing multi-dimensional consistency verification on the proposal based on the knowledge graph and generating a verification score; an intelligent evolution decision-making step: fusing multi-source information to calculate composite confidence, and outputting a refined evolution operation instruction; a dynamic ontology evolution step: monitoring a new mode and realizing self-adaptive updating of the ontology; a trust management step of endowing knowledge with trust values and dynamically adjusting the trust values; an execution and optimization step; and a knowledge tracing step of recording full life cycle information. According to the invention, automatic, dynamic and credible construction and continuous evolution of the knowledge graph are realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method for dynamic construction and evolution of knowledge graphs based on the collaboration of LLM and KG. Background Technology

[0002] Knowledge graphs, as structured semantic knowledge bases, are core infrastructure supporting advanced artificial intelligence applications such as intelligent search, question answering, recommendation, and decision-making. Traditional knowledge graph construction heavily relies on domain experts for ontology design and data annotation, or on extraction from limited structured data using supervised learning models. This results in inherent drawbacks such as long construction cycles, high costs, limited scalability, and lagging updates. Essentially, it is a "static knowledge base" that cannot adapt to rapidly changing real-world information. Therefore, this paper proposes a dynamic construction and evolution method for knowledge graphs based on the collaboration of LLM and KG.

[0003] The existing technology has the following technical defects, specifically:

[0004] 1. Traditional knowledge graphs rely on manual construction by experts or supervised learning, resulting in high update costs and long cycles. This static architecture cannot capture and integrate rapidly evolving information in fields such as news and technology in real time, forming "data silos." The lag in knowledge bases severely restricts their application value in scenarios requiring real-time perception and decision-making.

[0005] 2. Existing methods typically treat large language models as shallow "extraction tools," and their inherent "illusion" problem directly contaminates the knowledge base. Furthermore, there is a lack of deep integration of reasoning and verification mechanisms between LLMs and knowledge graphs. More importantly, most of these methods can only perform simple knowledge "addition," completely failing to automate complex evolutionary logic such as knowledge conflicts, fact obsolescence, and entity merging.

[0006] 3. Existing technologies lack quantitative assessment and dynamic management mechanisms for the credibility of knowledge itself, leading to doubts about the reliability of knowledge bases. Furthermore, predefined ontology structures cannot be updated automatically as instance knowledge grows, and the system as a whole cannot learn and optimize from historical decisions, remaining at an initial, low-intelligence level of automation. Summary of the Invention

[0007] The purpose of this invention is to provide a method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention aims to provide a method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration, comprising: an intelligent perception step: perceiving potential new knowledge from multi-source unstructured data, extracting candidate knowledge units using a large language model and generating relevant evidence and initial confidence scores to form a structured knowledge proposal, wherein the knowledge proposal includes candidate triples, original text evidence fragments supporting the candidate triples, initial confidence scores generated by the large language model, and source metadata.

[0009] Collaborative verification steps: Based on the knowledge proposal, relevant subgraphs are retrieved from the knowledge graph, the knowledge proposal is subjected to multi-dimensional consistency verification, and verification scores corresponding to each verification dimension are generated.

[0010] Intelligent evolutionary decision-making steps: Based on the verification score and the initial confidence score in the knowledge proposal, a decision model is used to calculate the composite confidence score and output refined knowledge evolution operation instructions.

[0011] Dynamic ontology evolution steps: monitor new knowledge patterns, generate ontology evolution proposals, and integrate them into existing ontology based on consistency verification.

[0012] Trust management steps: Assign and store a dynamic trust value for each fact in the knowledge graph, and dynamically adjust the dynamic trust value based on the execution result of the evolution operation instructions.

[0013] Execution and optimization steps: Execute the evolution operation instructions to update the knowledge graph, analyze the impact of changes, and optimize system performance.

[0014] Knowledge tracing steps: Record the full lifecycle information of each knowledge unit to form a traceable change chain. The full lifecycle information includes at least: the source of knowledge generation, the results of each verification, evolutionary decision and operation records, the history of trust value changes, and the interaction records with manual review.

[0015] As a further improvement to this technical solution, the intelligent perception step specifically includes: a context-enhanced prompt construction sub-step: based on the text content of the multi-source unstructured data and preset domain knowledge, constructing a context-enhanced prompt that includes task instructions, output format examples, and text to be processed.

[0016] Multi-turn reasoning generation sub-step: The context-enhanced prompts are input into a pre-trained large language model, and a natural language response containing candidate triples and their semantic interpretations is generated through multi-turn reasoning and thought chain technology.

[0017] Structured parsing and confidence calculation sub-steps: Based on preset parsing rules, candidate triples and corresponding evidence text fragments are extracted from the natural language response, and the initial confidence score is calculated based on the probability distribution output by the large language model.

[0018] Metadata annotation sub-step: Attach source metadata to each structured knowledge proposal, which includes data source identifier, acquisition timestamp, text location information, and processing batch identifier.

[0019] As a further improvement to this technical solution, the confidence calculation includes: semantic consistency score: based on the logical probability of each candidate triple in the complete response output by the large language model, the semantic consistency score is calculated.

[0020] Pattern matching score: The extracted candidate triples are matched with a preset relation pattern library to calculate the pattern matching score.

[0021] Evidence Coverage Score: The evidence text fragment is evaluated to assess the degree of support coverage for the candidate triples, and the evidence coverage score is calculated.

[0022] The initial confidence score is a weighted sum of the semantic consistency score, pattern matching score, and evidence coverage score, with the weights dynamically adjusted based on historical extraction accuracy.

[0023] As a further improvement to this technical solution, the multi-dimensional consistency verification includes factual consistency, logical consistency, temporal consistency, and source reliability.

[0024] The fact consistency verification involves checking whether the knowledge proposal conflicts with existing knowledge in the relevant subgraph, calculating the conflict intensity and type, and outputting a fact consistency score.

[0025] Logical consistency verification: Based on ontology constraints and predefined reasoning rules of the knowledge graph, verify the logical rationality of the knowledge proposal and output a logical consistency score.

[0026] Time Consistency Verification: For knowledge with time attributes, examine the logical relationship between the knowledge proposal and the event timeline in the relevant subgraph, and output a time consistency score.

[0027] Source reliability verification: The reliability of the knowledge proposal is evaluated based on the source metadata and historical accuracy, and cross-validation is performed on similar proposals from multiple independent sources to output a source reliability score.

[0028] As a further improvement to this technical solution, the intelligent evolutionary decision-making step specifically includes: the decision model receiving the verification score and the initial confidence score, and calculating the composite confidence score by integrating the source metadata of the knowledge proposal and the structured context information provided by the relevant subgraph.

[0029] Based on the preset set of decision strategies and the composite confidence level, a refined evolutionary operation instruction is output. The evolutionary operation instruction includes at least the operation type and operation parameters. The operation type includes: adding, merging, updating, abolishing, and triggering causal review.

[0030] As a further improvement to this technical solution, when the operation type is "triggering causal review", the intelligent evolutionary decision-making step further includes a causal conflict deep resolution sub-process: using the large language model to perform counterfactual reasoning analysis on conflicting knowledge and generating a causal analysis report.

[0031] Based on the causal analysis report and the related subgraph, a composite knowledge structure correction scheme is constructed, and this scheme is output as a new evolutionary operation instruction.

[0032] As a further improvement to this technical solution, the dynamic ontology evolution step specifically includes: a new pattern detection sub-step: continuously analyzing newly added instance knowledge to identify new relational patterns or entity categories that exceed the current ontology definition.

[0033] Evolutionary proposal generation sub-step: Using the large language model, generate ontology evolution proposals for the new relation schema or entity category, including definitions, attributes, and relationships with existing ontology.

[0034] Ontology verification sub-step: Evaluate the logical consistency, necessity, and scope of impact on existing knowledge of the proposed ontology evolution.

[0035] Ontology integration sub-step: Integrate the verified new relational patterns or entity categories as new ontology elements into the ontology of the knowledge graph.

[0036] As a further improvement to this technical solution, the trust management steps specifically include: dynamic trust assessment: when a new fact is adopted according to the evolutionary operation instruction, its dynamic trust value is initialized based on its composite confidence level.

[0037] Trust propagation mechanism: The positive change of the dynamic trust value is propagated to its associated entities and information sources in a limited way along the relational path in the knowledge graph; when an existing fact is invalidated, its dynamic trust value is cleared to zero and negative trust propagation is triggered, reducing the credibility weight of its associated entities and information sources.

[0038] Trust decay mechanism: A time decay factor is set for the dynamic trust value, so that the dynamic trust value of knowledge facts that have not been continuously verified decreases over time.

[0039] As a further improvement to this technical solution, in the knowledge tracing step, the full lifecycle information is stored in a chain log structure organized in chronological order, supporting rapid historical backtracking and audit tracing based on knowledge identifiers, timestamps, or operation user identifiers.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] 1. This invention deeply integrates the semantic understanding and generation capabilities of LLM with the structured logic and reasoning capabilities of KG through a collaborative verification and intelligent decision-making module. The two are no longer in a one-way "tool usage" relationship, but constitute an intelligent system capable of mutual verification and joint reasoning, fundamentally surpassing the shallow modes of traditional pipelines or retrieval enhancement.

[0042] 2. This invention not only automatically adds new knowledge, but also automatically resolves knowledge conflicts, discards outdated information, and merges entities through refined operation instructions and causal review mechanisms, achieving a "metabolism" of knowledge. Simultaneously, it can automatically identify new patterns and drive the evolution of the ontology structure, enabling the knowledge graph to possess the vitality to continuously grow and adapt to new domains.

[0043] 3. This invention assigns quantifiable credibility to each piece of knowledge through multi-dimensional verification, composite confidence calculation, and dynamic trust management. End-to-end "knowledge traceability" records ensure the explainability and auditability of the decision-making process. Furthermore, a feedback-based self-optimization mechanism enables the entire system to continuously learn, constantly reducing the cost of manual intervention and forming a virtuous cycle of becoming increasingly intelligent with use. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0045] Figure 1 This is a schematic diagram of the implementation steps of the method of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Example: Please refer to Figure 1 As shown, a method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration is provided, including: intelligent perception step: perceiving potential new knowledge from multi-source unstructured data, extracting candidate knowledge units using a large language model and generating relevant evidence and initial confidence scores to form a structured knowledge proposal. The knowledge proposal includes candidate triples, original text evidence fragments supporting the candidate triples, initial confidence scores generated by the large language model, and source metadata.

[0048] The candidate triples refer to knowledge units that conform to the (head entity, relation, tail entity) structure initially extracted from unstructured text by the large language model. They are candidate inputs for the dynamic evolution of the knowledge graph, and their accuracy needs to be verified in subsequent steps. For example, from the sentence 'Apple Inc. released a new iPhone,' the possible candidate triples extracted would be (Apple Inc., released, new iPhone).

[0049] In one specific embodiment, the intelligent perception step specifically includes: a context-enhanced prompt construction sub-step: based on the text content of the multi-source unstructured data and preset domain knowledge, constructing a context-enhanced prompt that includes task instructions, output format examples, and text to be processed.

[0050] The specific sub-step for constructing context-enhanced suggestions includes:

[0051] Dynamic context retrieval: Based on the key entities of the text to be processed, relevant entity, relationship and attribute information are retrieved from the knowledge graph to construct the domain knowledge context.

[0052] Few-sample example selection: Select the N examples with the highest semantic similarity to the current text from the knowledge proposals that have been successfully extracted in the past as few-sample examples.

[0053] Instruction template adaptation: Based on the domain type and language style of the text to be processed, specific task instructions are generated by selecting or adapting from the preset instruction template library.

[0054] The structure of the context-enhanced prompt is: [System Role Instruction] + [Domain Knowledge Context] + [Few Sample Examples] + [Text to be Processed] + [Output Format Constraints].

[0055] Multi-turn reasoning generation sub-step: The context-enhanced prompts are input into a pre-trained large language model, and a natural language response containing candidate triples and their semantic interpretations is generated through multi-turn reasoning and thought chain technology.

[0056] Structured parsing and confidence calculation sub-steps: Based on preset parsing rules, candidate triples and corresponding evidence text fragments are extracted from the natural language response, and the initial confidence score is calculated based on the probability distribution output by the large language model.

[0057] Metadata annotation sub-step: Attach source metadata to each structured knowledge proposal, which includes data source identifier, acquisition timestamp, text location information, and processing batch identifier.

[0058] In one specific embodiment, the confidence calculation includes: semantic consistency score: calculating the semantic consistency score based on the logical probability of each candidate triplet output by the large language model in the complete response.

[0059] Evaluate whether the LLM itself is "self-consistent" with respect to the generated triples, and capture the inherent determinism of its output logic.

[0060] Computational basis: Utilizing the generation probabilities of LLM (such as token probability or sequence probability).

[0061] Calculation formula:

[0062] Let candidate triplet be set .

[0063] Let the complete natural language response generated by the LLM for this triple be denoted as . .calculate The normalized log probability (or average token probability) is used as the benchmark for its overall fluency.

[0064] Furthermore, conditional probabilities can be calculated by masking important elements. For example, masking relationships... Make the model based on the head entity Tail-end entity Predictive relationship, calculate prediction as probability .

[0065] Semantic consistency score It can be defined as a weighted sum or combination of the above probabilities:

[0066]

[0067] in, and Represents harmonic weights, usually It has a higher weight because it directly tests whether the relationship is reasonable.

[0068] Example: For the text "Musk founded SpaceX", LLM generates the response "Entity: Musk, Relationship: Founder, Entity: SpaceX". A high value means the model is very confident that the "founder" is a reasonable link between "Musk" and "SpaceX".

[0069] Pattern matching score: The extracted candidate triples are matched with a preset relation pattern library to calculate the pattern matching score.

[0070] By using the existing structured knowledge (patterns) in the knowledge graph as external constraints, we can check whether candidate triples conform to domain common sense and established patterns.

[0071] Computational foundation: A "relation schema library" pre-built or mined from an existing KG. Schemas may include: relations Domain (head entity type) and range (tail entity type), symmetric / transitive attributes of relations, common entity pairs, etc.

[0072] Calculation formula:

[0073] Type constraint check: If relation The schema library requires the header entity type to be... If the tail entity type is Company, then check... and Whether it is classified as the corresponding type in KG (or determined by LLM / entity link). 1 point is awarded if it is, otherwise 0 points are awarded.

[0074] Statistical likelihood check: Calculating relationships from KG Connecting in existing instances Similar entities and The frequency of similar entities is used as statistical support.

[0075] Pattern matching score It can be a weighted sum of the scores from the above checks:

[0076]

[0077] ( For indicator functions, (Normalized statistical frequency).

[0078] Example: For (Beijing, is the capital, China), the domain of the "is the capital" relation in the schema library should be City, and the value domain should be Country. If the type of "Beijing" in the schema library is City, and the type of "China" is Country, then this item will score higher.

[0079] Evidence Coverage Score: The evidence text fragment is evaluated to assess the degree of support coverage for the candidate triples, and the evidence coverage score is calculated.

[0080] Assess whether the extracted original text evidence fragments adequately and accurately support the triples, preventing LLM from "over-interpreting" or "misgeneralizing".

[0081] Computational basis: alignment of text with triplet elements.

[0082] Calculation formula:

[0083] Named entity overriding: Checking whether a head entity explicitly appears in the evidence fragment. Tail-end entity The text mentions it. Matching can be done using string matching or named entity recognition.

[0084] Relationship expression coverage: Checking whether the evidence fragment contains expressions related to the relationship. Keywords or phrases that are highly semantically related. Predefined relationships can be used to trigger the thesaurus or calculations. Similarity between the text embeddings of the evidence fragments and the text embeddings.

[0085] Evidence Coverage Score It can be defined as:

[0086]

[0087] Entity coverage score (0 or 1, or a continuous value based on the degree of matching).

[0088] : Semantic similarity between relation and evidence (normalized to [0,1] by the Sigmoid function).

[0089] Example: For the triple (Apple Inc., release, iPhone 15), if the evidence fragment is "Apple Inc. recently released the iPhone 15", then the entity is fully covered, and "released" and "launched" are highly similar. The score is close to 1. If the evidence fragment is "Apple's performance hits a new high," then the coverage is insufficient, and the score is extremely low.

[0090] The initial confidence score is a weighted sum of the semantic consistency score, pattern matching score, and evidence coverage score, with the weights dynamically adjusted based on historical extraction accuracy.

[0091]

[0092] in, .

[0093] Dynamic adjustment implementation steps:

[0094] Historical data records: Preserving the final fate of every knowledge proposal (adopted, rejected, or subject to manual correction).

[0095] Dimensional reliability assessment: Periodic statistics, for example, calculating the average value of each score dimension from the past M ultimately adopted knowledge proposals. ) and variance ( ). High and Lower dimensions are more reliable.

[0096] Weight update formula (a feasible solution):

[0097] For dimensions :

[0098] Reliability indicators / / To prevent small constants from being divided by zero

[0099] New weights / / For all dimensions .

[0100] Collaborative verification steps: Based on the knowledge proposal, relevant subgraphs are retrieved from the knowledge graph, the knowledge proposal is subjected to multi-dimensional consistency verification, and verification scores corresponding to each verification dimension are generated.

[0101] In one specific embodiment, the multi-dimensional consistency verification includes factual consistency, logical consistency, temporal consistency, and source reliability.

[0102] The fact consistency verification involves checking whether the knowledge proposal conflicts with existing knowledge in the relevant subgraph, calculating the conflict intensity and type, and outputting a fact consistency score.

[0103] Detect whether new knowledge directly contradicts assertions in the existing knowledge base.

[0104] Verification process:

[0105] Retrieve conflicting facts: In the retrieved relevant subgraph, find candidate triples. Existing facts leading to direct conflict .

[0106] Conflict type determination:

[0107] Type A (Direct Negation): .For example, , .

[0108] Type B (mutual exclusion): ,but and They are mutually exclusive by definition. For example, , .

[0109] Type C (Attribute Numerical Conflict): For numerical attributes, the new value and the old value exceed the reasonable error range. For example, , .

[0110] Calculate conflict intensity and score:

[0111] If there is no conflict, then .

[0112] If a conflict exists, then:

[0113]

[0114] : No. The weight of conflicting facts is determined by the following criteria: direct denial (Category A) has the highest weight (e.g., 1.0), followed by mutual exclusion (Category B) (e.g., 0.7), and numerical contradiction (Category C) can be calculated based on the proportion of difference.

[0115] Conflict Facts The current dynamic trust value. The higher the trust value, the stronger the opposition to new proposals.

[0116] Example: Knowledge proposal (Musk, position, Tesla CEO). The relevant subgraph retrieves high-trust facts (Tesla, CEO, Musk). While not a direct denial, this could be considered a weak correlation conflict (subject inversion). It may be slightly lower than 1.0, such as 0.9.

[0117] Logical consistency verification: Based on ontology constraints and predefined reasoning rules of the knowledge graph, verify the logical rationality of the knowledge proposal and output a logical consistency score.

[0118] Based on the domain ontology and axioms, examine whether the knowledge proposal is reasonable at the conceptual level.

[0119] Verification process and formula:

[0120] The following checks will be performed based on the ontology; each violation will result in a deduction of points:

[0121] Domain / range check: Checking relationships Domain Does it contain entities? Types ( ),range( Does it contain entities? The type.

[0122]

[0123]

[0124]

[0125] Attribute cardinality check: Check entity Do you already possess more than The relationship specifies the maximum cardinality ( The same type of relationship.

[0126]

[0127] Class disjointness check: Check or Is the type similar to The types implied by the relationship are logically disjoint. )conflict.

[0128]

[0129] Logical consistency score :

[0130]

[0131] in, Indicates domain penalty, Indicates range penalty, Indicates cardinal penalty. Indicates the penalty for classes not intersecting. , The pre-defined violation weights are usually less than 1, ensuring that a single violation does not cause the score to drop to zero.

[0132] Example: The ontology defines the domain of the relation as Person and the value range as Organization. If the proposal is ( Because the "Artificial Intelligence" type is Concept, not Person, , It will be significantly reduced.

[0133] Time Consistency Verification: For knowledge with time attributes, examine the logical relationship between the knowledge proposal and the event timeline in the relevant subgraph, and output a time consistency score.

[0134] Ensure that events or facts involving time attributes are consistent across the timeline.

[0135] Verification process and formula:

[0136] For those with timestamps ( ) or time period ( Knowledge suggestions:

[0137] Constructing a local timeline: Extracting all entities from the relevant subgraph and / or Relevant facts with time information are sorted chronologically.

[0138] Execution time logic check:

[0139] Existential conflict: At any given time, an entity cannot simultaneously possess two mutually exclusive states or attributes. For example, a person cannot be "employed by Company A" and "employed by Company B" simultaneously during [2020-2023] (unless it is a part-time job, but this must be permitted by the ontology definition).

[0140] Sequence conflict: Some events have a fixed order. For example, "graduation" must be after "enrollment"; "divorce" must be after "marriage".

[0141] Interval rationality: The end time of a time period cannot be earlier than the start time.

[0142] Calculate time consistency score :

[0143] If no relevant facts about time are found or no conflict is discovered .

[0144] If a conflict is found, points will be deducted based on the severity of the conflict (e.g., complete overlap vs. partial overlap, violation of rigid order vs. soft order).

[0145]

[0146] : No. The severity coefficient of each time conflict (0~1).

[0147] Trust value of existing time facts that triggered the conflict.

[0148] Example: New proposal Date: 2024-06-01. Related sub-graphs show... The time frame is from January 1, 2023 to the present. If the ontology stipulates that "a person cannot simultaneously serve as a core executive in two competing companies within the same time period," then a conflict arising from overlapping time intervals will occur. reduce.

[0149] Source reliability verification: The reliability of the knowledge proposal is evaluated based on the source metadata and historical accuracy, and cross-validation is performed on similar proposals from multiple independent sources to output a source reliability score.

[0150] Assess the "origin" of knowledge proposals and quantify the credibility of information sources.

[0151] Verification process and formula:

[0152] Single-source reliability score:

[0153] Extract source identifiers (such as "News Website X" or "Academic Journal Y") from source metadata.

[0154] Query the historical accuracy archive of this source, which is maintained by statistically analyzing the percentage of knowledge proposals provided by this source in the past that were ultimately adopted by the system.

[0155] / / A dynamically updated value, the initial value can be a prior value from the domain.

[0156] Multi-source cross-validation gain:

[0157] The system searches whether it has recently received semantically similar candidate knowledge proposals from other independent sources.

[0158] If there are L independent sources supporting similar facts, then a confidence gain is given.

[0159] Source reliability score Comprehensive calculation:

[0160]

[0161] The main sources are historically reliable.

[0162] Multi-source corroboration coefficients. Can be designed as: , where L is the number of independent corroborating sources. When L=0, C=0, and the larger L is, the closer C is to 1.

[0163] The tradeoff parameter (e.g., 0.7) indicates the emphasis placed on the credibility of a single source itself.

[0164] Example: A knowledge proposal comes from a "well-known scientific journal" ( Furthermore, the system detected another "authoritative industry report" that mentioned a similar fact (L=1, C≈0.1). Therefore... .

[0165] Intelligent evolutionary decision-making steps: Based on the verification score and the initial confidence score in the knowledge proposal, a decision model is used to calculate the composite confidence score and output refined knowledge evolution operation instructions.

[0166] The decision model is a machine learning-based classifier (such as a gradient boosting decision tree or a neural network). Its input is a fused feature vector containing the validation score, initial confidence score, etc., and its output is a probability distribution of the evolutionary operation type. The model is trained under supervised supervision using historical decision data (knowledge proposal features and final adoption / rejection labels).

[0167] In one specific embodiment, the intelligent evolutionary decision-making step specifically includes: the decision model receiving the verification score and the initial confidence score, and calculating the composite confidence score by integrating the source metadata of the knowledge proposal and the structured context information provided by the relevant subgraph.

[0168] Composite confidence score is a quantitative assessment of the final credibility of a knowledge proposal by a decision-making model. It is calculated through the fusion of multi-source evidence. The specific process is as follows:

[0169] 1. Feature Vector Construction

[0170] The decision model first integrates various types of input evidence into a multi-dimensional feature vector:

[0171] Initial confidence score

[0172] Multidimensional Validation Score: Four-Dimensional Vector

[0173] Source metadata features: Accuracy of extracting source history Authority level Release time and freshness , etc., normalized to the interval [0, 1].

[0174] Structured context features: calculated from the relevant subgraph:

[0175] Context support ,in, Represents a related subgraph. Facts in the subgraph Trust value.

[0176] Neighborhood Consistency This measures the compatibility of candidate triples with their neighboring nodes.

[0177] Calculate the composite confidence level using a weighted linear combination or neural network model.

[0178]

[0179] in, To verify the set of dimensions, , 2. 3. 4. 5 represents the weighting coefficient. This weighting coefficient is not fixed but dynamically adjusted based on the historical performance of each feature dimension.

[0180] Record the contribution of each feature dimension in historical decisions, use a sliding window to calculate the correlation between the scores of each dimension and the adopted results, and update the weights regularly so that more reliable dimensions receive higher weights.

[0181] Based on the preset set of decision strategies and the composite confidence level, a refined evolutionary operation instruction is output. The evolutionary operation instruction includes at least the operation type and operation parameters. The operation type includes: adding, merging, updating, abolishing, and triggering causal review.

[0182] The decision model, based on composite confidence and a pre-defined set of decision strategies, generates evolutionary operation instructions through multi-level judgment logic: First, it compares the composite confidence with a dynamic threshold to preliminarily determine the credibility level of the knowledge proposal; second, it selects the most matching decision rule from the strategy set by combining specific verification score patterns (such as which dimensions have lower scores and the specific type of conflict) and structured contextual information; finally, based on the selected operation type, it extracts relevant knowledge elements (such as conflicting entities, time information, source comparison, etc.) to generate detailed operation parameters, forming a complete and executable instruction.

[0183] In one specific embodiment, when the operation type is "triggering causal review", the intelligent evolutionary decision-making step further includes a causal conflict deep resolution sub-process: using the large language model to perform counterfactual reasoning analysis on conflicting knowledge and generating a causal analysis report.

[0184] The generation of the causal analysis report is a structured, guided, automated reasoning process. The system first transforms the detected conflicting knowledge (candidate triples and existing high-trust facts) and their related subgraphs (containing entities, historical relationships, and attributes) into a specific analysis task prompt, which is then input into the large language model. This prompt explicitly requires the model to perform "counterfactual reasoning," that is, to consider whether the outcome would still occur if a certain cause were absent, and to analyze the interactions between multiple possible causes (such as substitution, promotion, and inhibition). Based on its world knowledge and chain reasoning capabilities, the large language model outputs a structured analysis report. This report typically includes: a determination of the conflict type (e.g., direct causal conflict, common cause confusion, missing causal chain, etc.), an assessment of the causal strength of each potential cause, an analysis of the existence of mediating or confounding variables, and a preliminary causal explanation. This report serves as the direct basis for subsequently constructing corrective solutions.

[0185] Based on the causal analysis report and the related subgraph, a composite knowledge structure correction scheme is constructed, and this scheme is output as a new evolutionary operation instruction.

[0186] The construction of a composite knowledge structure correction scheme is an analysis-based, execution-oriented knowledge engineering process. The system receives a causal analysis report and, in conjunction with relevant subgraphs of the knowledge graph (providing structured context such as entity types and relational constraints), formulates a correction scheme through a scheme generation module. This module matches or combines corresponding operation sequences from a predefined correction pattern library based on the conclusions in the analysis report. For example, if the analysis identifies "multiple causes," the scheme might add a composite relation node representing "joint effect" and adjust the confidence of the original relation; if it identifies "reversed causal time sequence," the scheme might abolish the old relation and add a relation with the correct time sequence, while supplementing time attribute constraints. The final generated correction scheme is a directly executable, finely detailed operation instruction package, which clearly specifies the specific operations to be performed, such as adding, merging, updating, or abolishing, and their parameters, ensuring that the corrected knowledge fragments maintain consistency and rationality in logical, temporal, and factual dimensions.

[0187] Dynamic ontology evolution steps: monitor new knowledge patterns, generate ontology evolution proposals, and integrate them into existing ontology based on consistency verification.

[0188] In one specific embodiment, the dynamic ontology evolution step specifically includes: a new pattern detection sub-step: continuously analyzing newly added instance knowledge to identify new relational patterns or entity categories that exceed the current ontology definition.

[0189] Evolutionary proposal generation sub-step: Using the large language model, generate ontology evolution proposals for the new relation schema or entity category, including definitions, attributes, and relationships with existing ontology.

[0190] Ontology verification sub-step: Evaluate the logical consistency, necessity, and scope of impact on existing knowledge of the proposed ontology evolution.

[0191] In the ontology verification sub-step, the evaluation of ontology evolution proposals is a multi-dimensional, systematic review process: First, a logical consistency check is performed to ensure that the newly proposed concept definitions and relation attributes do not conflict with the existing ontology architecture. For example, new classes do not form circular inheritance or logical contradictions with existing classes, and the domain and value range of new relations are compatible with the existing type system. Second, a necessity assessment is conducted, based on the statistical significance of the new pattern (such as frequency of occurrence and number of associated instances) and its semantic uniqueness, to determine whether it is a redundant representation of existing ontology elements or whether it truly represents a new category worthy of abstraction in knowledge evolution. Finally, an impact scope analysis is performed, through simulated reasoning on the graph structure, to predict which existing queries, reasoning rules, and instance classifications will be affected by the new ontology elements, and to assess its complexity and controllability. Only proposals that meet the preset standards in all three assessments (such as logical consistency, statistical significance, and controllable impact) will be approved to enter the integration phase.

[0192] Ontology integration sub-step: Integrate the verified new relational patterns or entity categories as new ontology elements into the ontology of the knowledge graph.

[0193] Trust management steps: Assign and store a dynamic trust value for each fact in the knowledge graph, and dynamically adjust the dynamic trust value based on the execution result of the evolution operation instructions.

[0194] In one specific embodiment, the trust management step specifically includes: dynamic trust assessment: when a new fact is adopted according to the evolutionary operation instruction, its dynamic trust value is initialized based on its composite confidence level.

[0195] Trust propagation mechanism: The positive change of the dynamic trust value is propagated to its associated entities and information sources in a limited way along the relational path in the knowledge graph; when an existing fact is invalidated, its dynamic trust value is cleared to zero and negative trust propagation is triggered, reducing the credibility weight of its associated entities and information sources.

[0196] Trust decay mechanism: A time decay factor is set for the dynamic trust value, so that the dynamic trust value of knowledge facts that have not been continuously verified decreases over time.

[0197] Execution and optimization steps: Execute the evolution operation instructions to update the knowledge graph, analyze the impact of changes, and optimize system performance.

[0198] Knowledge tracing steps: Record the full lifecycle information of each knowledge unit to form a traceable change chain. The full lifecycle information includes at least: the source of knowledge generation, the results of each verification, evolutionary decision and operation records, the history of trust value changes, and the interaction records with manual review.

[0199] In the example, during the knowledge tracing step, the full lifecycle information is stored in a chain log structure organized in chronological order, supporting rapid historical backtracking and audit tracing based on knowledge identifiers, timestamps, or user identifiers.

[0200] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration, characterized in that, include: Intelligent perception steps: Perceive potential new knowledge from multi-source unstructured data, extract candidate knowledge units using a large language model and generate relevant evidence and initial confidence scores to form a structured knowledge proposal. The knowledge proposal includes candidate triples, original text evidence fragments supporting the candidate triples, initial confidence scores generated by the large language model, and source metadata. Collaborative verification steps: Based on the knowledge proposal, relevant subgraphs are retrieved from the knowledge graph, the knowledge proposal is subjected to multi-dimensional consistency verification, and verification scores corresponding to each verification dimension are generated; Intelligent evolutionary decision-making steps: Based on the verification score and the initial confidence score in the knowledge proposal, a decision model is used to calculate the composite confidence score and output refined knowledge evolution operation instructions; Dynamic ontology evolution steps: monitor new knowledge patterns, generate ontology evolution proposals, and integrate them into existing ontology based on consistency verification; Trust management steps: Assign and store dynamic trust values ​​for each fact in the knowledge graph, and dynamically adjust the dynamic trust values ​​based on the execution results of the evolution operation instructions; Execution and optimization steps: Execute the evolution operation instructions to update the knowledge graph, analyze the impact of changes, and optimize system performance; Knowledge tracing steps: Record the full lifecycle information of each knowledge unit to form a traceable change chain. The full lifecycle information includes at least: the source of knowledge generation, the results of each verification, evolutionary decision and operation records, the history of trust value changes, and the interaction records with manual review.

2. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 1, characterized in that: The intelligent sensing steps specifically include: Context-enhanced prompt construction sub-step: Based on the text content of the multi-source unstructured data and preset domain knowledge, construct a context-enhanced prompt that includes task instructions, output format examples, and text to be processed; Multi-turn reasoning generation sub-step: The context-enhanced prompts are input into a pre-trained large language model, and a natural language response containing candidate triples and their semantic interpretations is generated through multi-turn reasoning and thought chain technology; Structured parsing and confidence calculation sub-steps: Based on preset parsing rules, candidate triples and corresponding evidence text fragments are extracted from the natural language response, and the initial confidence score is calculated based on the probability distribution output by the large language model; Metadata annotation sub-step: Attach source metadata to each structured knowledge proposal, which includes data source identifier, acquisition timestamp, text location information, and processing batch identifier.

3. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 2, characterized in that: The confidence level calculation includes: Semantic consistency score: The semantic consistency score is calculated based on the logical probability of each candidate triplet output by the large language model in the complete response. Pattern matching score: The extracted candidate triples are matched with a preset relation pattern library, and the pattern matching score is calculated. Evidence Coverage Score: Evaluate the degree of support coverage of the evidence text fragments for the candidate triples and calculate the evidence coverage score; The initial confidence score is a weighted sum of the semantic consistency score, pattern matching score, and evidence coverage score, with the weights dynamically adjusted based on historical extraction accuracy.

4. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 1, characterized in that: The multi-dimensional consistency verification includes factual consistency, logical consistency, temporal consistency, and source reliability; Among them, the fact consistency verification is as follows: check whether the knowledge proposal conflicts with the existing knowledge in the relevant subgraph, calculate the conflict intensity and type, and output the fact consistency score; Logical consistency verification: Based on the ontology constraints and predefined reasoning rules of the knowledge graph, verify the logical rationality of the knowledge proposal and output a logical consistency score; Time consistency verification: For knowledge with time attributes, examine the logical relationship between the knowledge proposal and the event timeline in the relevant subgraph, and output a time consistency score; Source reliability verification: The reliability of the knowledge proposal is evaluated based on the source metadata and historical accuracy, and cross-validation is performed on similar proposals from multiple independent sources to output a source reliability score.

5. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 4, characterized in that: The intelligent evolutionary decision-making steps specifically include: The decision model receives the verification score and the initial confidence score, and calculates the composite confidence score by combining the source metadata of the knowledge proposal with the structured context information provided by the relevant subgraph; Based on the preset set of decision strategies and the composite confidence level, a refined evolutionary operation instruction is output. The evolutionary operation instruction includes at least the operation type and operation parameters. The operation type includes: adding, merging, updating, abolishing, and triggering causal review.

6. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 5, characterized in that: When the operation type is "trigger causal review", the intelligent evolutionary decision-making step further includes a sub-process for deep resolution of causal conflicts: The aforementioned large language model is used to perform counterfactual reasoning analysis on conflicting knowledge, generating a causal analysis report. Based on the causal analysis report and the related subgraph, a composite knowledge structure correction scheme is constructed, and this scheme is output as a new evolutionary operation instruction.

7. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 1, characterized in that: The dynamic ontology evolution steps specifically include: New pattern detection sub-step: Continuously analyze newly added instance knowledge to identify new relational patterns or entity categories that exceed the current ontology definition; Evolutionary proposal generation sub-step: Using the large language model, generate ontology evolutionary proposals for the new relational schema or entity category, including definitions, attributes, and relationships with existing ontology. Ontology verification sub-step: Evaluate the logical consistency, necessity, and scope of impact on existing knowledge of the proposed ontology evolution; Ontology integration sub-step: Integrate the verified new relational patterns or entity categories as new ontology elements into the ontology of the knowledge graph.

8. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration as described in claim 1, characterized in that: The trust management steps specifically include: Dynamic Trust Assessment: When a new fact is adopted according to the evolutionary operation instructions, its dynamic trust value is initialized based on its composite confidence level; Trust propagation mechanism: The positive change of the dynamic trust value is propagated to its associated entities and information sources in a limited positive direction along the relational path in the knowledge graph; when an existing fact is invalidated, its dynamic trust value is cleared to zero and negative trust propagation is triggered, reducing the credibility weight of its associated entities and information sources. Trust decay mechanism: A time decay factor is set for the dynamic trust value, so that the dynamic trust value of knowledge facts that have not been continuously verified decreases over time.

9. The method for dynamic construction and evolution of knowledge graphs based on LLM and KG collaboration according to claim 1, characterized in that: In the knowledge tracing step, the full lifecycle information is stored in a chain log structure organized in chronological order, supporting rapid historical backtracking and audit tracing based on knowledge identifiers, timestamps, or user identifiers.

Images