Incremental knowledge graph dynamic updating method and system based on influence domain recognition

Abstract

The invention discloses an incremental knowledge graph dynamic updating method and system based on influence domain recognition, and belongs to the technical field of artificial intelligence and knowledge engineering. The method comprises the following steps: acquiring an incremental knowledge unit containing a new entity and a relationship; determining an influence domain of the new entity in an existing knowledge graph by taking the new entity as a center and utilizing a multi-strategy mixed neighborhood retrieval algorithm, and constructing a local relation sub-graph; and performing iterative enhancement and integration processing on the local relation sub-graph, and circularly performing context enhancement, intelligent deduplication detection, entity aggregation and relation prediction operation in each round of iteration until the graph quality index meets a preset convergence condition. According to the method, the updating complexity is reduced to a local range through an influence domain recognition mechanism, and the problems that the incremental updating efficiency of the large-scale knowledge graph is low, the data consistency is poor and the technical evolution version is missing are effectively solved by combining iterative convergence and a multi-dimensional conflict detection technology.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to an incremental knowledge graph dynamic update method and system based on influence domain recognition. Background Technology

[0002] Knowledge graphs, as a structured semantic knowledge base, organize and represent domain knowledge through triples of entities, relations, and attributes, and have become a core infrastructure for artificial intelligence, search engines, and intelligent question-answering systems. Since Google proposed the concept of knowledge graphs in 2012, its construction technology has gone through three development stages: rule-based extraction (such as YAGO), supervised learning-based extraction (such as DeepDive), and the current generation based on large language models (LLM).

[0003] With the rapid development of knowledge-intensive fields such as artificial intelligence, the rate of knowledge generation is increasing exponentially. Taking the AI ​​field as an example, the arXiv platform adds hundreds of new papers daily, and GitHub projects iterate frequently. This high-frequency dynamic update requirement poses a significant challenge to the maintenance of knowledge graphs. However, the inventors have discovered that existing knowledge graph construction and update technologies suffer from the following architectural flaws when facing rapid iteration scenarios: 1. An irreconcilable contradiction exists between efficiency and quality in update mechanisms: Existing mainstream solutions typically employ two extreme strategies: The first is a full reconstruction strategy (such as Chinese patent CN111984643B), which reconstructs the entire graph each time new data is received. While this ensures the integrity of the graph structure, when the graph reaches tens of thousands of entities, the update time increases linearly or even superlinearly, failing to meet real-time requirements. The second is a simple incremental insertion strategy (such as Chinese patent application CN116542327A), which directly inserts new knowledge using a streaming update method. Although the update speed is fast, it lacks in-depth analysis of the impact of new and old knowledge relationships, resulting in the inability to maintain local consistency. Currently, there is a lack of a "precise patching" mechanism that can both avoid full graph reconstruction and accurately control the scope of local updates.

[0004] 2. Lagging conflict detection technology struggles to handle semantic-level logical contradictions: In the process of multi-source knowledge fusion, data from different sources often conflict. Existing systems (such as Wikidata) mainly use string matching or simple vector similarity methods for detection, which can only identify basic data type errors or explicit duplications. However, in complex scientific literature, conflicts often manifest at the semantic-logical level. For example, the performance indicators of the same technical model may differ under different experimental conditions, or the definition of the same concept may drift at different times. Existing technologies lack quantitative assessment of the credibility of evidence sources and cannot identify implicit conflicts of "conditional dependence" based on context, leading to the frequent long-term coexistence of self-contradictory knowledge in the knowledge graph, seriously affecting the credibility of the knowledge graph.

[0005] 3. Lack of native support for knowledge evolution and temporal versioning: Traditional knowledge graph construction methods typically treat entities as static concepts. When technology iterates (such as model version updates), existing methods often simply treat them as duplicate entities and merge them, or directly overwrite old data. This approach loses valuable context for technological evolution, making it impossible for users to query the "optimal technical solution" at a specific point in time, or to track the evolution of technical parameters. Although some systems have introduced timestamps, these are mainly used to record editing history and have not established a cross-entity version evolution semantic network.

[0006] In summary, existing technologies lack an integrated dynamic update architecture that combines precise local updates, intelligent conflict resolution, and end-to-end version management. How to achieve low-cost incremental expansion in large-scale knowledge graphs while ensuring logical consistency and evolutionary traceability of knowledge is a pressing technical challenge in this field. Summary of the Invention

[0007] The purpose of this application is to provide a knowledge graph incremental update method and system based on influence domain identification and conflict detection, aiming to solve the technical problems of low efficiency of full reconstruction, lack of consistency guarantee for simple incremental updates, and lack of technical evolution version management capabilities in existing knowledge graph update technologies.

[0008] To achieve the above objectives, this application provides an incremental knowledge graph dynamic update method based on influence domain identification, comprising the following steps: Step S1: Obtain the incremental knowledge unit to be updated; the incremental knowledge unit includes entity information, relation triples and corresponding metadata extracted and structured from the new document; Step S2: Determine the influence domain of the incremental knowledge unit in the existing knowledge graph; take the entity in the incremental knowledge unit as the center, perform multi-strategy hybrid neighborhood retrieval in the existing knowledge graph, obtain the set of neighboring nodes that are related to the entity, and construct a local relation subgraph based on the set of neighboring nodes; Step S3: Perform iterative enhancement and integration processing on the local relation subgraph; in each iteration, perform context enhancement, deduplication detection, entity merging and relation prediction operations on the entities and relations in the local relation subgraph, and calculate the rate of change of the graph quality index until the rate of change meets the preset convergence condition or reaches the maximum number of iterations; Step S4: Perform multidimensional conflict detection and resolution on the local relation subgraph after iterative processing; identify conflicts based on multidimensional rules of entity attribute layer and relation structure layer, and generate the resolved knowledge representation using multi-source evidence fusion mechanism, and persist the final result into the existing knowledge graph.

[0009] Optionally, the multi-strategy hybrid neighborhood retrieval in step S2 includes: retrieving neighbor nodes based on three strategies: text co-occurrence window, semantic vector similarity, and entity type matching; wherein, the text co-occurrence window strategy is used to identify nodes that co-occur with the entity within a preset distance in the newly added document; the semantic vector similarity strategy is used to identify nodes in the existing knowledge graph whose embedding vector cosine similarity with the entity is greater than a first preset threshold; and the entity type matching strategy is used to identify nodes with the same type definition as the entity; the set of neighbor nodes is a weighted union of the retrieval results of the above three strategies.

[0010] Optionally, the context enhancement operation in step S3 specifically includes: constructing a structured context object based on the retrieved neighbor nodes for the new entity in the local relational subgraph; constructing prompt words containing role definition, task objective and the structured context object; and calling a large language model to generate enhanced descriptive text for the new entity; the enhanced descriptive text is generated by an optimization function, which aims to maximize the similarity between the descriptive text and the original text, the relevance with the local context and the degree of satisfaction of preset constraints.

[0011] Optionally, the deduplication detection operation in step S3 specifically includes: performing a five-layer progressive duplicate entity determination process, wherein the five layers are as follows: Name layer matching: calculating the string exact match degree or edit distance similarity of entity names; Semantic layer matching: calculating the vector cosine similarity of entity enhanced description text; Local role layer analysis: comparing the differences in the distribution of relation types of entities in the graph and the overlap rate of neighbors; Temporal layer verification: verifying whether there is an evolutionary relationship based on the differences in timestamps and reference relationships of entities; Large model inference confirmation: when the above four-layer determination results are in the critical interval, calling the large language model for final semantic confirmation.

[0012] Optionally, the entity merging operation in step S3 specifically includes: when a duplicate entity is detected, determining the main node to be retained based on a preset main entity selection strategy; the main entity selection strategy includes: prioritizing the selection of entities with more complete descriptive text, prioritizing the selection of entities with higher source credibility, or prioritizing the selection of entities with the latest timestamp in the context of technological evolution; migrating the alias list, attribute information and relationship network of secondary entities to the main node, and marking the secondary entities as non-core entities pointing to the main node.

[0013] Optionally, the relationship prediction operation in step S3 specifically includes: using semantic embedding vectors to filter candidate entity pairs in the local relationship subgraph; constructing prompt words containing attribute information and domain knowledge of candidate entity pairs, calling a large language model to predict whether there is an implicit relationship between entities and its relationship type; calculating the confidence score of the predicted relationship, retaining only relationships with confidence scores higher than a second preset threshold, and performing logical verification of transitivity and symmetry on the newly generated relationships.

[0014] Optionally, the convergence condition in step S3 shall satisfy at least one of the following conditions: Search result stability condition: In two consecutive iterations, the rate of change of the set of neighbor nodes obtained by the multi-strategy hybrid neighborhood search is less than the preset convergence threshold; Quality convergence condition: In two consecutive iterations, the improvement in the graph quality score is less than the preset quality threshold; the graph quality score is calculated based on entity integrity, deduplication accuracy and relation confidence.

[0015] Optionally, the multidimensional conflict detection and resolution in step S4 specifically includes: Entity layer detection: Detects whether the performance index deviation of the same entity exceeds a preset ratio, or detects whether the description of the same concept at different time points has semantic drift; Relationship layer detection: Detects whether there are abnormal structures in the relationship network that violate predefined symmetry constraints, lack transitive closures, or are logically mutually exclusive; For detected conflicts, if the fusion confidence is higher than the automatic resolution threshold, the map is automatically updated based on the fusion result; if it is lower than the automatic resolution threshold, a task to be reviewed is generated and pushed to the manual review queue.

[0016] Optionally, the multi-source evidence fusion mechanism calculates the credibility score of knowledge based on the following formula. :

[0017] in, The scores represent different dimensions, including: scores based on source authority, timeliness scores based on time decay factors, influence scores based on citation counts, scores based on attribute completeness, and confidence scores output by large language models. These are the adaptive weighting coefficients for each dimension.

[0018] An incremental knowledge graph dynamic update system based on influence domain identification includes: a memory for storing computer programs and knowledge graph data; and a processor for executing the computer programs.

[0019] Compared with the prior art, this application has the following beneficial effects: 1. This invention proposes a local update mechanism based on influence domain identification. By accurately locating the ripple effect boundary of new knowledge using the Ms-khop algorithm, the O(N) complexity of traditional full reconstruction is reduced to O(k) complexity, which is related to the size of the influence domain. This makes it possible to process massive incremental data (such as daily new AI papers) in large-scale knowledge graphs, significantly shortening the update latency.

[0020] 2. This invention constructs a comprehensive conflict detection framework covering dimensions such as entity values, concept definitions, and topological structures. Combined with an intelligent multi-source evidence fusion algorithm (TrustScore), the system can automatically distinguish between "measurement errors" and "real evolution," effectively solving the knowledge base pollution problem caused by simple incremental insertion strategies and ensuring that the quality of the graph does not degrade with the number of updates.

[0021] 3. Unlike existing technologies that simply overwrite old data, this invention introduces a "version-aware" processing logic. It incorporates temporal verification during the deduplication detection phase, adopts a time-priority strategy during the entity merging phase, and retains historical versions during conflict resolution. This ensures that the updated knowledge graph not only records the current state but also fully preserves the evolution of technical concepts, supporting users in querying technical solutions at specific points in time or analyzing technological development trends.

[0022] 4. By introducing an iterative convergence mechanism and a hierarchical conflict resolution strategy, this invention achieves highly automated closed-loop maintenance. The vast majority of routine updates and high-confidence conflicts are handled automatically by the system, with only a very small number of complex conflicts (low confidence, logical mutual exclusion) being pushed to manual review, significantly reducing the manpower cost of knowledge graph operation and maintenance. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is an overall flowchart of an incremental knowledge graph dynamic update method based on influence domain identification provided by an embodiment of the present invention. Figure 2 This is a detailed flowchart of the first stage (initial construction and incremental extraction) provided by the embodiments of the present invention; Figure 3 This is a detailed flowchart of the second-stage core step S2 (knowledge graph enhancement and expansion integration) provided in the embodiments of the present invention; Figure 4 This is a detailed flowchart of the second-stage core step S3 (conflict detection and evolution tracking) provided in the embodiments of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be noted that the specific embodiments described herein are only for explaining this application and are not intended to limit the scope of protection of this application.

[0026] 1. Overall Process Architecture like Figure 1 As shown in the figure, the incremental knowledge graph dynamic update method based on influence domain identification provided by the embodiments of the present invention is logically divided into two main stages and six core steps.

[0027] Phase 1 (S0-S1): Focuses on knowledge acquisition and structuring. This includes initializing and building an explicit knowledge graph (S0) and incremental knowledge extraction and structuring (S1). For building a completely new domain, the entire S0-S1 process is executed; for maintaining an existing graph, S1 and subsequent steps are mainly executed cyclically.

[0028] Phase Two (S2-S4): Focuses on knowledge integration, governance, and storage. This includes knowledge graph enhancement and expansion integration (S2), conflict detection and evolution tracking (S3), and persistent graph writing and indexing (S4).

[0029] The operating environment of this embodiment can be based on a single high-performance server or a distributed computing cluster. For calls to large language models (LLMs), the system supports accessing general commercial models (such as GPT-4, Claude 3) via API, or deploying open-source fine-tuning models locally (such as Llama 3, Qwen). This embodiment uses the API call of a general large model as an example for explanation.

[0030] 2. Step S0: Initialize and construct the explicit knowledge graph This step aims to build a high-quality "baseline map" for subsequent incremental updates. For example... Figure 2 As shown, it specifically includes the following sub-steps: S01: Document Hierarchical Segmentation. This step is performed by the document hierarchical segmentation device. The device first employs different segmentation strategies based on the document type: For structured documents (such as textbooks and standard books): The table of contents page location algorithm is executed, identifying table of contents items through text block detection, and extracting chapter titles and page number ranges based on font size, indentation level, and line spacing features. Regular expressions are used to match chapter numbering patterns (such as "1.1", "Chapter 2") to parse the hierarchy depth.

[0031] For unstructured documents (such as papers and reports without a table of contents): chapter boundaries are identified by detecting heading style tags (such as HTML H1-H3 or Markdown # symbols). For plain text, a pre-trained language model is used to perform semantic analysis on candidate headings, filtering out non-content chapters and constructing a virtual table of contents structure. Finally, the document is parsed into a chapter tree data structure, with each node containing attributes such as {id, type, title, level, summary, entities[], parent}.

[0032] S02: Large Model Generates Summary To reduce the contextual overhead of subsequent processing, this step utilizes a large language model to generate a summary of 200-300 words for each smallest granular chapter.

[0033] Prompt word engineering: Utilizes structured prompt word templates, including: Role: Knowledge graph construction expert in XX field; Task: Generate a summary for the input chapter containing core concepts, main knowledge points, and key theorems; Constraints: Abstract length is limited; it must include a description of the actual application. Output format: JSON format {summary, key_concepts[], key_theorems[]}.

[0034] Quality Control: A batch processing strategy is used to call LLM, and a quality verification mechanism is set up. The similarity between the generated summary and the original text's embedding vector is calculated. If the similarity is lower than a preset threshold or JSON parsing fails, a retry mechanism is automatically triggered.

[0035] S03: Initial entity and relation extraction is performed based on the summary generated in S02, performing domain-specific extraction.

[0036] Entity node construction: Convert the extracted results into an Entity_Node structure and generate a unique identifier entity_id; initialize the core entity identifier is_core=false; initialize the embedding vector field embedding=null (to be filled in later in the S2 stage). Few-shot learning strategy: Provide 2-3 domain-specific extraction examples in the prompt words to guide the model to learn extraction patterns. For example, in the field of AI, predefined entity types include: neural network architecture, optimization algorithm, pre-trained model, evaluation metric, etc.

[0037] Relation extraction: Identify predefined relation types, including definition relations (defines), application relations (applies_to), dependency relations (depends_on), etc., and record the model's confidence score.

[0038] S04: The three-layer graph structure organizes the above information into a hierarchical graph structure, containing three types of edges: 1. Vertical edge: Used to represent the parent-child relationship between chapters; 2. Belonging Edge: Connects an entity to the chapter node where it was first introduced; 3. Horizontal edges: These represent the semantic relationships between entities. Finally, they are written to the graph database via transaction batch processing using a graph database (such as Neo4j), completing the initialization and construction.

[0039] 3. Step S1: Incremental Knowledge Extraction and Structuring This step is triggered when the system detects a new document (such as a newly published arXiv paper or technical report).

[0040] S11: The document source classification and metadata annotation system first identifies the document source (e.g., top conference papers, technical blogs, official documents) based on file metadata, URL patterns, and text features, and assigns an initial credibility score. It constructs a Document_Metadata node to store the document name, publication date, author list, and source type, and links it to all subsequently extracted new knowledge units, providing the basis for calculating evidence weights for conflict resolution in stage S3. S12: Incremental entity information extraction reuses the prompt word template of S03, but pays special attention to technical version and performance parameter information in incremental extraction.

[0041] Parameter extraction: Specially designed parameter extraction rules extract structured fields such as model size, training data volume, and hardware requirements from text.

[0042] Performance Recording: Identify the results tables in the documents, extract precise performance metrics (such as accuracy and latency), and record the evaluation conditions (dataset, evaluation metrics) to construct the Performance_Record node. This is the key data foundation for tracking technology evolution.

[0043] S13: Incremental Relation Triple Extraction. Building upon S03, this step focuses on expanding technology evolution-specific relations: Improvements: Specific performance improvement metrics should be specified. replaces (alternatives): Applicable scenarios must be specified; "Uses" is further subdivided into subtypes such as "calling APIs," "adopting ideas," and "integrating components." In addition, this embodiment introduces a conditional relation extraction mechanism. It identifies the preconditions for a relation to hold (such as "on long sequence tasks"), stores them as the relation's `condition` attribute, and extracts key sentences from the original text as evidence.

[0044] S14: Incremental knowledge unit construction assembles the scattered information extracted from S11-S13 into standardized units to be processed.

[0045] Define a unified JSON data schema: containing an array of entities, an array of relations, a metadata object, and a quality_metrics object.

[0046] Preprocessing verification: Establish a temporary mapping table for entity IDs, resolve entity references in relation triples, and ensure that all relations point to existing entity IDs. Perform preliminary quality checks on knowledge units, including checks on the completeness of required fields, the correctness of data types, and the consistency of references.

[0047] The output incremental knowledge units will serve as the standard input for the subsequent step S2 (knowledge graph enhancement and expansion integration).

[0048] 4. Step S2: Knowledge Graph Enhancement and Expansion Integration like Figure 3 As shown, this step is the core of the incremental update. The system receives the incremental knowledge unit output from step S1 and, through the iterative convergence execution mechanism (S26), cyclically calls the five functional modules from S21 to S25 until the graph state reaches stability. This design ensures that after the introduction of new knowledge, its associated effects (such as attribute completion and relation reasoning) can fully propagate and converge within the local region.

[0049] S21: Local Context Enhancement Device (Influence Domain Recognition) This embodiment first uses a multi-strategy hybrid neighborhood retrieval (Ms-khop) algorithm to accurately locate new entities. Influence domain, construct local relation subgraph The retrieval strategy uses a weighted union approach, defined as follows: ; in, k The number of hops (in this embodiment, it is set to) k =2), N i Representing neighbor sets in different dimensions: Physical co-occurrence neighbors N i Text co-occurrence window Identify entities that appear in the same incremental document fragment.

[0050] Semantic similar neighbors N 2 : Calculate the entity vectors and their corresponding vectors in the existing graph. Cosine similarity is used to filter samples with similarity greater than a preset threshold. The entity.

[0051] Type Ontology Neighbors N 3 : Filtering and Nodes belonging to the same type (e.g., all are "Transformer architecture").

[0052] Weighting coefficient Adaptive calculation mechanism: ; Based on the retrieved neighbor nodes and their relationships, the device constructs a structured context object and uses a large language model to generate context-enhanced descriptions. This process is a constrained optimization problem, aiming to maximize the similarity between the descriptive text and the original text, its relevance to the local context, and the degree to which it meets pre-defined constraints (such as including technical positioning and innovative points).

[0053] S22: To address the issue that simple string matching cannot identify semantic duplicates, this embodiment employs a five-layer progressive detection mechanism: 1. Name layer: Perform exact matching, Levenshtein fuzzy matching (threshold 0.9), and abbreviation dictionary matching. If the similarity is <0.85, they are directly determined to be different entities.

[0054] 2. Semantic layer: Enhanced description of entities Perform vector encoding and calculate cosine similarity. .like A value greater than 0.90 is considered a high-susceptibility duplicate.

[0055] 3. Local Role Layer: For highly suspected pairs, analyze their structural role differences. By comparing the distribution of entity relation types The overlap rate with neighbors distinguishes entities at different levels, such as "abstract theory" and "concrete implementation".

[0056] 4. Time-series layer: Validating the timestamp of the entity t And the relationship of reference. If there is a time difference If the time difference is less than 3 months and there is a citation relationship, it is determined to be a technological evolution rather than a duplication; if the time difference is less than 3 months and the authors overlap, it is determined to be the same technology.

[0057] 5. LLM final confirmation: For boundary cases (confidence level 0.8-0.9) of the first four layers of decision, construct a prompt word containing the complete context, and call LLM to perform a binary decision (true / false).

[0058] S23: Entity aggregation and merging device. When S22 determines that two entities are duplicates or aliases, it performs an aggregation operation.

[0059] Main entity selection: The main node is selected using a weighted scoring formula. :

[0060] In the context of technological evolution, by adjusting... The entity with the latest version is retained as the master node based on weight.

[0061] Information fusion: The unique attributes and alias list of the secondary entity are appended to the main entity; the description text of the two are merged using LLM; the relationship network of the secondary entity is migrated to the main entity, and the secondary entity is marked as is_core=false, pointing to the main entity through the merged_into field to retain the ability to trace history.

[0062] Feature concatenation: Concatenates the entity's name, alias, type, description, and key parameters into a comprehensive text.

[0063] Encoding and Indexing: Normalized vectors are generated using a pre-trained model (such as all-MiniLM-L6-v2). An incremental indexing strategy is adopted, where new vectors are first written to a memory buffer, and after accumulating to a batch threshold (such as 1000), they are inserted into the HNSW or IVF index structure using a local graph update algorithm, avoiding a full index reconstruction.

[0064] S25: Relation prediction constructs a device to mine potential implicit relations in local subgraphs.

[0065] Candidate generation: Filter semantically related entity pairs based on vector similarity (similarity > 0.6); LLM Prediction: Construct cue words containing details of entity pairs, and ask LLM to predict the relationship type and provide a strength score.

[0066] Validation filtering: Only retain high-confidence relationships with an intensity score ≥ 7.0. Simultaneously perform logical validation: Loop detection: to prevent generation A logical dead loop.

[0067] Transitivity verification: If it exists Then verify Should it also have an improves relationship?

[0068] S26: Iterative Convergence Execution Mechanism. The system iteratively executes steps S21-S25, calculating the convergence metric after each iteration. Iteration stops when any of the following conditions are met: Retrieval stability: Rate of change of the node set in a local subgraph .

[0069] Quality convergence: The improvement in map quality score (based on completeness, conflict rate, etc.) .

[0070] Forced termination: The number of iterations reaches a preset upper limit (e.g., max_iter = 5). After the convergence condition is met, the enhanced and integrated intermediate knowledge graph is output, and the process enters the S3 stage.

[0071] 5. Step S3: Conflict Detection and Evolution Tracking like Figure 4 As shown, after the local enhancements and integration are completed in stage S2, logical inconsistencies may still exist in the graph. This step performs comprehensive quality governance through modules S31-S34.

[0072] S31: Entity Layer Conflict Detection Device. This device focuses on identifying numerical contradictions and conceptual evolution conflicts, employing the following detection operators: Performance Deviation Detection D 1) For the Performance_Record node of the AI ​​model, calculate the relative deviation of the same metric in different sources.

[0073]

[0074] in, The deviation threshold is set to 0.15 in this embodiment. Used to verify whether test conditions (such as datasets and hardware) are consistent. If the conditions do not match, it is marked as "condition dependency difference" rather than conflict.

[0075] Semantic drift detection ( D 2) Identify changes in the definition of technical concepts over time. By comparing the semantic vectors of entity descriptions at different points in time, if the similarity is below a drift threshold... If so, it is marked as "concept drift".

[0076] S32: Relationship layer conflict detection device detects the logical consistency of the graph topology: Symmetry violation: For relations defined as symmetric (such as similarity relations), check for the existence of one-way edges.

[0077] Transitive closure missing: if present Check if it exists .

[0078] Logical mutual exclusion: Based on domain ontology constraints, detect whether there are mutually exclusive type connections (such as an entity that is both an "optimization algorithm" and a "dataset").

[0079] S33: When a conflict is detected, the multi-source evidence fusion device aggregates evidence from different sources and calculates a comprehensive trust score to determine which side's data to accept. The scoring formula is as follows: ; in, Normalized scores representing five dimensions: Source authority (S1): Based on the source type in the metadata (e.g., top conference vs. blog).

[0080] Timeliness (S2): Using an exponential decay function This gives higher weight to new evidence.

[0081] Reference Influence (S3): .

[0082] Evidence integrity (S4): Calculated based on field missing rate.

[0083] LLM confidence score (S5): The logical consistency score of the evidence content by the large language model.

[0084] S34: The conflict resolution and marking mechanism implements a hierarchical processing strategy based on TrustScore. Automatic resolution (high confidence): If TrustScore > 0.85, the system automatically adopts high-scoring evidence, updates the graph nodes, and archives the conflict records to the historical version.

[0085] Manual review (low confidence level): If TrustScore < 0.7 or involves core entities (degree > 50), the system generates a review task and pushes it to the manual review queue.

[0086] Evolution Preservation: For conflicts identified as "technical evolution" (not errors), the old version node is preserved, and a new one is created. The relationship points to a new node, enabling version management.

[0087] 6. Step S4: Persistent writing and indexing of the map This step is responsible for writing the processed, high-quality data into the graph database.

[0088] S41 Transaction Batch Processing: Encapsulates update operations for entities, relationships, and attributes into atomic transactions. It employs a Write-Ahead Log (WAL) mechanism to prevent data loss and performs topology sorting based on operation relevance to avoid deadlocks.

[0089] S43 Multi-modal Index Construction: The system maintains multiple index structures simultaneously to support efficient queries. Structural index: Primary key and attribute index based on B+ tree.

[0090] Vector Index: A semantic vector index based on the HNSW algorithm, supporting "image search by meaning".

[0091] Full-text index: The inverted index supports keyword retrieval.

[0092] 7. Hardware Implementation Environment This invention also provides a computer system for performing the above-described method. The system includes: Processor: Used to execute program instructions in memory, implementing all steps S1-S4.

[0093] Storage: Stores knowledge graph data, vector index files, and computer programs.

[0094] Network interface: Used for receiving incremental document streams and communication with large model APIs.

[0095] The technical solution adopted in this embodiment reduces the update complexity of large-scale knowledge graphs from O(N) to O(k) through a local update mechanism based on influence domain identification, significantly improving real-time performance. By using multi-dimensional conflict detection and TrustScore evidence fusion, the data consistency problem in incremental updates is effectively solved. At the same time, the version management function based on temporal logic enables the knowledge graph to fully record the technological evolution, which has extremely high application value.

[0096] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An incremental knowledge graph dynamic update method based on influence domain identification, characterized in that, Includes the following steps: Step S1: Obtain the incremental knowledge unit to be updated; the incremental knowledge unit includes entity information, relation triples and corresponding metadata extracted and structured from the new document; Step S2: Determine the influence domain of the incremental knowledge unit in the existing knowledge graph; take the entity in the incremental knowledge unit as the center, perform multi-strategy hybrid neighborhood retrieval in the existing knowledge graph, obtain the set of neighboring nodes that are related to the entity, and construct a local relation subgraph based on the set of neighboring nodes; Step S3: Perform iterative enhancement and integration processing on the local relation subgraph; In each iteration, context enhancement, deduplication detection, entity merging and relation prediction operations are performed on the entities and relations within the local relation subgraph, and the rate of change of the graph quality index is calculated until the rate of change meets the preset convergence condition or reaches the maximum number of iterations. Step S4: Perform multidimensional conflict detection and resolution on the local relation subgraph after iterative processing; identify conflicts based on multidimensional rules of entity attribute layer and relation structure layer, and generate the resolved knowledge representation using multi-source evidence fusion mechanism, and persist the final result into the existing knowledge graph.

2. The method according to claim 1, characterized in that, The multi-strategy hybrid neighborhood retrieval in step S2 includes: retrieving neighbor nodes based on three strategies: text co-occurrence window, semantic vector similarity, and entity type matching; wherein, the text co-occurrence window strategy is used to identify nodes that co-occur with the entity within a preset distance in the newly added document; the semantic vector similarity strategy is used to identify nodes in the existing knowledge graph whose embedding vector cosine similarity with the entity is greater than a first preset threshold; and the entity type matching strategy is used to identify nodes with the same type definition as the entity; the set of neighbor nodes is the weighted union of the retrieval results of the above three strategies.

3. The method according to claim 2, characterized in that, The context enhancement operation in step S3 specifically includes: for a new entity in the local relational subgraph, constructing a structured context object based on its retrieved neighbor nodes; constructing prompt words containing role definitions, task objectives, and the structured context object; and calling a large language model to generate enhanced descriptive text for the new entity; the enhanced descriptive text is generated through an optimization function, which aims to maximize the similarity between the descriptive text and the original text, the relevance to the local context, and the degree of satisfaction of preset constraints.

4. The method according to claim 1, characterized in that, The deduplication detection operation in step S3 specifically includes: executing a five-layer progressive duplicate entity determination process, wherein the five layers are as follows: Name layer matching: calculating the string exact match degree or edit distance similarity of entity names; Semantic layer matching: calculating the vector cosine similarity of entity augmented description text; Local role layer analysis: comparing the differences in the distribution of relation types of entities in the graph and the overlap rate of neighbors; Temporal layer verification: verifying the existence of evolutionary relationships based on the differences in timestamps and reference relationships of entities; Large model inference confirmation: when the above four-layer determination results are in the critical interval, calling the large language model for final semantic confirmation.

5. The method according to claim 4, characterized in that, The entity merging operation in step S3 specifically includes: when a duplicate entity is detected, determining the main node to be retained based on a preset main entity selection strategy; the main entity selection strategy includes: prioritizing entities with more complete descriptive text, prioritizing entities with higher source credibility, or prioritizing entities with the latest timestamp in the context of technological evolution; migrating the alias list, attribute information and relationship network of secondary entities to the main node, and marking the secondary entities as non-core entities pointing to the main node.

6. The method according to claim 1, characterized in that, The relationship prediction operation in step S3 specifically includes: using semantic embedding vectors to filter candidate entity pairs in the local relationship subgraph; constructing prompt words containing attribute information and domain knowledge of candidate entity pairs, calling a large language model to predict whether there is an implicit relationship between entities and its relationship type; calculating the confidence score of the predicted relationship, retaining only relationships with a confidence score higher than a second preset threshold, and performing logical verification of transitivity and symmetry on the newly generated relationship.

7. The method according to claim 1, characterized in that, The convergence condition in step S3 must satisfy at least one of the following conditions: Search result stability condition: In two consecutive iterations, the rate of change of the set of neighbor nodes obtained by the multi-strategy hybrid neighborhood search is less than the preset convergence threshold; Quality convergence condition: In two consecutive iterations, the improvement in the graph quality score is less than the preset quality threshold; the graph quality score is calculated based on entity integrity, deduplication accuracy and relation confidence.

8. The method according to claim 1, characterized in that, The multidimensional conflict detection and resolution in step S4 specifically includes: Entity layer detection: Detects whether the performance index deviation of the same entity exceeds a preset ratio, or detects whether the description of the same concept at different time points has semantic drift; Relationship layer detection: Detects whether there are abnormal structures in the relationship network that violate predefined symmetry constraints, lack transitive closures, or are logically mutually exclusive; For detected conflicts, if the fusion confidence is higher than the automatic resolution threshold, the map is automatically updated based on the fusion result; if it is lower than the automatic resolution threshold, a task to be reviewed is generated and pushed to the manual review queue.

9. The method according to claim 1, characterized in that, The multi-source evidence fusion mechanism calculates the credibility score of knowledge based on the following formula. : ； in, The scores represent different dimensions, including: scores based on source authority, timeliness scores based on time decay factors, influence scores based on citation counts, scores based on attribute completeness, and confidence scores output by large language models. These are the adaptive weighting coefficients for each dimension.

10. An incremental knowledge graph dynamic update system based on influence domain identification, characterized in that, include: Memory, used to store computer programs and knowledge graph data; A processor for executing the computer program to implement the steps of the incremental knowledge graph dynamic update method based on influence domain identification as described in any one of claims 1 to 9.

Images