An organization content construction system based on a knowledge graph

Abstract

The application belongs to the technical field of knowledge graph, and discloses a kind of organization content construction system based on knowledge graph, and dynamic updating module is captured data change by three ways of log monitoring, incremental interface, timing scanning, and the associated business chain is combed in combination with organization business process topology structure.Introduce update influence degree evaluation model, determine update granularity according to business importance and urgency, for example, personnel resignation can adjust project responsible relationship, document permission and other associated information synchronously in this kind of scene, avoid isolated update;Rely on graph database transaction mechanism and update in order, also support rollback and multi-subsidiary cooperation, both reduce the resource consumption of full update, and also guarantee that knowledge graph and business scene are real-time synchronized, solve the key problems of knowledge consistency deficiency and poor real-time performance;Build three-level dynamic terminology library of basic layer, industry layer and organization layer, fuse business semantic pre-training model and graph neural network, accurately extract organization exclusive entity, relationship and attribute.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph technology, specifically a knowledge graph-based content organization and construction system. Background Technology

[0002] In the process of digital transformation, knowledge graphs, as a core technological support for organizational digitalization, can integrate scattered multi-source data, establish semantic relationships and logical links between data, and uncover hidden knowledge value, providing intelligent support for business collaboration and decision-making. They have become a key tool for improving organizational operational efficiency. However, current knowledge graph construction solutions still face the following technical challenges:

[0003] Existing technologies mostly adopt static ontology modeling and batch update mode, which makes it difficult to perceive the internal relationship of organizational business processes. For example, when "personnel leave", it is necessary to adjust related information such as "project responsibility relationship", "document permission" and "approval chain" at the same time. However, existing solutions can only realize isolated updates of single data, resulting in the disconnect between the map and the actual business scenario. At the same time, the full update mode consumes a lot of resources, while incremental updates lack systematic analysis of business relationship logic, making it difficult to accurately locate the update scope, resulting in insufficient knowledge consistency and difficulty in ensuring real-time performance.

[0004] Existing knowledge graph construction relies heavily on general domain ontology, failing to adequately adapt to the semantics specific to an organization's business. For example, there are semantic differences between "agile projects" for technology companies and "production line projects" for manufacturing companies, and between the definition of "OKR cycle" for internet companies and "KPI cycle" for traditional companies. This makes semantic biases prone to occur during knowledge extraction and rule conflicts easy to arise during the knowledge fusion stage, thus requiring a large amount of manual adjustment of ontology and rules, resulting in high adaptation costs. Summary of the Invention

[0005] The purpose of this invention is to provide a knowledge graph-based content organization and construction system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a knowledge graph-based content organization and construction system, comprising a collection module, a knowledge extraction module, a knowledge fusion module, a graph construction module, a dynamic update module, an intelligent application module, and a quality control module;

[0007] Preferably, the acquisition module supports access to data sources including databases, API interfaces, enterprise collaboration platforms, and business systems. For micro and small enterprises, it supports rapid access via lightweight API protocols; for large enterprises, it supports parallel access via distributed data sources; and for government agencies, it supports access via dedicated lines for confidential data.

[0008] It automatically identifies data source types and matches transmission protocols based on deep learning models, extracts data source business attributes, and generates collection configurations with business tags by combining the characteristics of the organization's industry.

[0009] Complete routine data cleaning operations such as deduplication and format standardization, and label data with business semantic tags according to organizational business rules; in the processing of classified data of government agencies, retain non-sensitive semantic tags such as approval number and processing status, and hide specific approval content; the labeling results are associated with the organization's basic semantic library and directly connected to the extraction engine of the knowledge extraction module.

[0010] The method for hiding classified data is a combination of "field-level de-identification + transmission encryption":

[0011] 1) Desensitization rules: High-secret fields (such as approval content, customer ID number) use "character replacement", medium-secret fields (such as project budget) use "range fuzzing", and low-secret fields (such as the name of the person in charge) use "surname retention + name desensitization";

[0012] 2) Encryption of transmission: Classified data is encrypted using HTTPS+SM4 symmetric encryption from the data source to the acquisition module. The key is dynamically allocated by the organization's key management system and is updated every 24 hours. After encryption, the data transmission packet must carry "encryption identifier-key version number-decryption permission level". Only processes in the acquisition module with the corresponding permissions can decrypt it. The permission level corresponds one-to-one with the classification level. Level 3 classification requires two-person authorization for decryption.

[0013] Preferably, the knowledge extraction module relies on the labeled data from the acquisition module and combines it with seed data, including organizational business manuals and process specifications, to mine organizational-specific business terms and semantic variations, and establish a three-level dynamic terminology library: a basic layer, an industry layer, and an organizational layer. The basic layer contains general terms, the industry layer contains industry-specific terms, and the organizational layer contains personalized terms. The terminology library is associated with business rules and is automatically updated as organizational business rules change, while also providing core terminology resources for the knowledge fusion module.

[0014] The automatic update mechanism for the Level 3 dynamic terminology database is as follows:

[0015] 1) Triggering conditions: Changes in organizational business rules, a cumulative total of ≥10 low-confidence review results from the knowledge extraction module, or the addition of a new data source containing unmatched terms;

[0016] 2) Update steps:

[0017] ① Based on newly added business rules / reviewed data / new data source text, candidate terms are extracted using the TF-IDF algorithm;

[0018] ② Use cosine similarity (threshold 0.85) to remove duplicates and exclude candidate words that are repeated with existing terms;

[0019] ③ The terminology level (basic level / industry level / organizational level) is matched by business rules. Organizational level terms need to be manually reviewed (approval pass rate ≥ 80% before they can be added to the database).

[0020] ④ Update terminology relationships: Calculate the degree of association between new terms and existing terms using a GNN model to supplement semantic variations of terms;

[0021] 3) Storage and Synchronization: The terminology database uses key-value pairs for storage (terminology-hierarchy-semantic tag-association rule ID). After updates, it is synchronized to the knowledge extraction, knowledge fusion, and dynamic update modules in real time, with a synchronization delay of ≤5 minutes.

[0022] By integrating BERT-business semantic pre-trained model and GNN, entities, relationships and attributes are extracted from business-labeled data collected by the acquisition module; when extracting entities, organization-specific semantics are distinguished; when extracting relationships, cross-business line semantic correspondences are associated; and when extracting attributes, industry characteristics are combined; in cross-departmental project collaboration data processing, cross-business line related information is automatically extracted.

[0023] The BERT-Business Semantic Pre-trained Model adopts a domain fine-tuning scheme based on a general Chinese pre-trained model. The training data includes three types of seed data: organizational business manuals, historical project documents, and industry standard texts. The fine-tuning steps are as follows:

[0024] 1) Data preprocessing: The seed data is segmented, part-of-speech tagging is performed, and entity boundary annotation is performed to generate input samples in the format of [CLS] text sequence [SEP];

[0025] 2) Model tuning: Freeze the parameters of the first 8 layers of the pre-trained model, fine-tune the last 4 layers and the classification head, and set the output dimension of the classification head to 3;

[0026] 3) Training parameters: The optimizer used is AdamW, the learning rate is set to 2e-5, the batch size is 16, the number of iterations is 10, and the early stopping strategy is used to prevent overfitting.

[0027] The GNN adopts the graph SAGE model, using the text vectors output by BERT as the initial features of the nodes, and the edge features are defined as the business label relevance. Cross-business line relevance features are learned through three aggregation layers.

[0028] The model input consists of text fragments with business labels and corresponding business association matrices. The output consists of entity type labels, relation triples, and attribute key-value pairs. Entity labels are annotated in BIO format, and relations and attributes are associated through entity location mapping.

[0029] Based on a three-level dynamic terminology database and business rules, the system automatically detects semantic deviations in the extraction results; sets differentiated evaluation thresholds for different business scenarios; labels low-confidence results and supports manual review; the review results update the extraction model parameters in reverse; and transmits the accurate extraction results to the knowledge fusion module.

[0030] Specific steps of the manual review process:

[0031] 1) Triggering condition: If the confidence level of the extraction result is <0.8, the system will automatically push the low-confidence results to the "Knowledge Review Workbench";

[0032] 2) Permission allocation: Review permissions are allocated according to business lines, and the workbench only displays review tasks with the corresponding permissions;

[0033] 3) Review process: Reviewers can choose "Confirm Correct", "Correct Error", or "Mark for Confirmation". When "Correct Error" is selected, the correct entity / relationship and the reason for the correction must be filled in. When "Mark for Confirmation" is selected, collaborating reviewers must be specified.

[0034] 4) Result feedback: After the review is completed, the system will automatically save the "review result - correction content - operator - timestamp" into the log. At the same time, the corrected labeled data will be used as training samples to update the parameters of the BERT-business semantic pre-trained model. Low confidence results that have not been reviewed will be retained for 7 days. If they are not processed within the time limit, they will be automatically marked as "to be optimized" and pushed to the quality control module.

[0035] Preferably, the knowledge fusion module calls the three-level terminology database of the knowledge extraction module and automatically matches synonymous entities from different data sources using a semantic similarity algorithm; in the scenario of a group enterprise with multiple subsidiaries, the semantic relationship between the subsidiaries and the group's semantics is established to ensure that the alignment results conform to the group's unified business definition; the alignment process takes into account the degree of business relevance.

[0036] The semantic similarity algorithm adopts a hybrid algorithm of "BERT embedding + cosine similarity + business relevance weighting", specifically implemented as follows:

[0037] 1) Entity embedding generation: Input the text description of the entity to be matched into the fine-tuned BERT model of the knowledge extraction module, and output a 768-dimensional entity vector;

[0038] 2) Basic similarity calculation: The vector similarity S1 is calculated using the cosine similarity formula, where S1 = (vector a) / (vector a) Vector b) / (||vector a|| × ||vector b||), with a value range of 0-1;

[0039] 3) Business relevance weighting: Introduce business relevance weight W (defined by organizational business rules, such as "entities of the same project type W=0.3, cross-departmental collaboration entities W=0.2"), and finally the similarity S=S1×(1+W), with the threshold set to 0.7 (the threshold for group enterprise scenarios can be customized to 0.65-0.8). If S≥ the threshold, it is determined to be a synonym entity.

[0040] The algorithm takes entity text and business tags from different data sources as input and outputs synonymous entity matching pairs and similarity scores. The matching results must meet the constraints defined by the organization's business.

[0041] It integrates organizational business rules and data source authority to resolve semantic conflicts, generates collaborative computing templates to incorporate both types of results into the calculation of project health indicators, supports customized business rules, and generates conflict handling logs that are pushed to the quality control module.

[0042] When monitoring the compatibility of newly added knowledge with existing ontology in the knowledge extraction module, analyze its relevance to the organization's business processes; combine the high-frequency related data types synchronized in real time by the dynamic update module, and prioritize recommend ontology extension schemes related to high-frequency entities; the ontology evolution of micro and small enterprises adopts a lightweight extension strategy, while the ontology evolution of group enterprises adopts a multi-sub-ontology collaborative evolution strategy, and the fused knowledge after evolution is transferred to the graph construction module.

[0043] Preferably, the graph construction module generates an initial ontology with business-related logic based on the fused knowledge of the knowledge fusion module and the organizational business process diagram; it provides an organization-specific ontology template import function with preset nodes; and the ontology generation process refers to scenario-based requirements.

[0044] By integrating community discovery and business association analysis algorithms, the graph hierarchy and relationship weights are dynamically adjusted; in cross-departmental business scenarios, the cross-departmental node association paths are optimized to improve query efficiency; the graph structure can be visually adjusted, allowing users to manually fine-tune relationship weights based on business needs, and the adjustment results are updated to the knowledge recommendation engine of the intelligent application module.

[0045] It adopts a hybrid storage architecture of Neo4j and HBase, which supports efficient storage and parallel querying of entities and business relationships, and provides transactional storage guarantee for dynamic updates of the dynamic update module. Small and micro enterprises adopt a single-machine hybrid storage architecture, group enterprises adopt a distributed cluster storage architecture, and government agencies adopt a confidential partition storage architecture. It has a built-in storage performance monitoring mechanism that automatically adjusts the storage sharding strategy according to the real-time data volume, and the monitoring data is transmitted to the quality control module.

[0046] Specific implementation of the hybrid storage interaction mechanism:

[0047] 1) Data allocation rules: The "ID-name-type-core business tag" of entity nodes is stored in Neo4j, and the detailed attributes of the entity are stored in HBase. The HBase table structure is designed as "rowkey=entity_id+timestamp, column family=attr (including fields such as budget and duration), log (including modification record fields)".

[0048] 2) Query interaction process: When the intelligent application module initiates a query request, Neo4j first queries the entity ID and associated personnel node ID of "Project A", then queries the budget data through HBase by the entity ID, and finally integrates "node relationship + attribute data" into JSON format and returns it.

[0049] 3) Synchronization mechanism: When the dynamic update module updates data, it first updates the node relationship of Neo4j, and then synchronously updates the attribute data of HBase. It uses "transaction lock" to ensure consistency, and the synchronization frequency is consistent with the update granularity.

[0050] Preferably, the dynamic update module captures data changes with business tags in the acquisition link of the acquisition module in real time through three methods: log monitoring, incremental interface of business system, and timed scanning; the core business data monitoring frequency is set to minute level, and the non-core data monitoring frequency is set to hour level; the monitoring results are labeled with the data change type and associated business tags;

[0051] Specific parameters for scenario-based monitoring frequency: For micro and small enterprises, the monitoring frequency for core business (such as order data) is 5 minutes / time, and for non-core business (such as administrative notices) it is 2 hours / time; for group enterprises, the monitoring frequency for core business (such as revenue data, cross-subsidiary projects) is 1 minute / time, and for local non-core business of subsidiaries it is 1 hour / time; for government agencies, the monitoring frequency for confidential business (such as approval data) is 3 minutes / time, and for public business (such as service guides) it is 4 hours / time.

[0052] Multi-subsidiary collaboration rules: When a subsidiary initiates an update request, it must include "Subsidiary ID - Update content summary - Group semantic rule version number". After receiving the request, the group first verifies the consistency of semantic rules (by calling the group semantic rule library of the knowledge fusion module). If the verification is successful, a "collaborative update token" is generated (valid for 30 minutes). The subsidiary executes the update with the token. The update result must be synchronized to the group graph library within 10 minutes. If the synchronization fails, a rollback is triggered and an alarm is pushed to the group administrator.

[0053] The system invokes the three-level dynamic terminology base of the knowledge extraction module and the ontology evolution rules of the knowledge fusion module to automatically analyze the business chain related to data changes based on the organization's business process topology. It introduces an update impact assessment model, which combines historical update logs and the importance weight of business data to predict the degree of impact of updates on business processes. Based on the impact and business urgency, it dynamically generates incremental update granularity. High-impact and high-urgency updates adopt a full-linkage coverage strategy, while low-impact and low-urgency updates adopt a core node priority strategy.

[0054] The updated impact assessment model is a gradient boosting decision tree model, specifically implemented as follows:

[0055] 1) Input feature dimensions: Includes 8 core features, namely business data importance weight (1-5 points, preset by the organization, 5 points for core business and 1 point for non-core business), frequency of historical update conflicts, length of related business chain, business process dependency, data update frequency, confidentiality level, user access frequency, and number of cross-departmental associations.

[0056] 2) Model training: The training data is the update log of the past year (including the above 8 types of features and manually labeled influence labels). The labels are divided into three levels: high (3 points), medium (2 points), and low (1 point). Training parameters: learning rate 0.05, number of decision trees 100, depth of each tree ≤ 8, minimum number of sample splits 20.

[0057] 3) Output and Thresholds: The model outputs an impact score (1-3 points), with high impact threshold ≥2.5 points, medium impact 1.5-2.4 points, and low impact ≤1.4 points; business urgency is determined by the organization's preset rules (e.g., "updates within 7 days before project delivery are high urgency, with an urgency score of 1-3 points"). The final update granularity is determined by the impact score × urgency score. A product ≥6 points indicates high priority, using full association chain coverage; 3-5 points indicates medium priority, using core + association node coverage; ≤2 points indicates low priority, using core node priority.

[0058] Based on the graph database transaction mechanism of the graph construction module, relevant knowledge is updated in the order of business association chains to ensure that the update process does not disrupt business logic; update logs are recorded to support update rollback and recovery, and the update results are transmitted to the application interface of the intelligent application module and the quality control module; in the scenario of multiple subsidiaries of a group enterprise, subsidiary updates are coordinated with group updates, and the group semantic rules of the knowledge fusion module are verified during the subsidiary update process.

[0059] Preferably, the intelligent application module relies on the graph structure data of the graph construction module and the real-time update results of the dynamic update module, combined with user roles and business scenarios, to automatically identify knowledge needs; in the scenario of simplified roles for micro and small enterprises, it supports the adaptation of one person to multiple roles and automatically aggregates knowledge related to multiple roles;

[0060] Based on the optimized business association graph built by the graph construction module, corresponding knowledge views and knowledge chains are pushed to different roles; in cross-departmental project collaboration scenarios, customer feedback knowledge from the marketing department is pushed to R&D personnel, and technical bottleneck knowledge from the R&D department is pushed to marketing personnel; the recommendation results are dynamically optimized by combining user interaction data, and the optimized data is transmitted to the quality control module.

[0061] It provides interfaces for natural language query, business correlation analysis, and visualization, supporting integration with business systems. Interface data directly calls the storage architecture of the graph construction module and the update results of the dynamic update module. In OA system integration, a graph query entry is embedded, allowing users to query project-related personnel and historical approval records when initiating project approvals. It supports multi-terminal adaptation, with optimized knowledge display format for mobile interfaces and support for complex graph analysis for PC interfaces. The government agency interface is configured with access control logic for classified data, allowing only authorized personnel to query sensitive knowledge, and the permission rules are consistent with the classified data processing strategy of the collection module.

[0062] Preferably, the quality control module sets quality indicators at each stage, including completeness and accuracy at the data level and semantic consistency and logical correlation at the business level; it sets differentiated quality thresholds for different organizational types; and it receives monitoring data from each module in real time and generates quality reports.

[0063] It calls upon the three-level dynamic terminology database of the knowledge extraction module and the conflict handling rules of the knowledge fusion module, and automatically identifies semantic errors and related logical errors based on organizational business rules and anomaly detection models; it generates error type statistical reports for high-frequency error scenarios; it supports graded error handling, automatically corrects minor errors, and pushes serious errors to manual review, recording the error cause, handling plan, and executor information in the review results;

[0064] The anomaly detection model is a hybrid model of "rule base + isolated forest", specifically implemented as follows:

[0065] 1) Rule base input: Semantic rules and association logic rules are constructed based on a three-level dynamic terminology database, which includes 50 basic rules, 30 industry rules, and 20 organization-defined rules. The rule matching priority is: organization rules > industry rules > basic rules.

[0066] 2) Isolation Forest Model Configuration: Used to detect anomalies not covered by the rule base. Input features are data integrity score (0-1), semantic annotation consistency score (0-1), and association reasonableness score (0-1). Model parameters: 100 decision trees, 256 samples per tree, and anomaly detection threshold set to 0.3. Anomaly scores ≥ 0.3 are judged as logical errors.

[0067] 3) Model input and output: The input is the monitoring data (including data-level indicators and business-level indicators) transmitted by each module, and the output is the anomaly type, anomaly confidence (0-1), and associated module identifier. Semantic errors include entity classification errors, label confusion, etc., and logical errors include contradictory relationships, conflicting business rules, etc.

[0068] Correctable errors are handled automatically, and the handling solutions are updated in reverse to update the business rule base; the results of complex error handling are fed back to the corresponding front-end modules; a closed loop of quality control and business rule optimization is established, quality data is analyzed regularly, control strategies are optimized, and the optimization strategies are transmitted to the update evaluation model of the knowledge fusion module and the dynamic update module.

[0069] The beneficial effects of this invention are as follows:

[0070] 1. The dynamic update module of this invention captures data changes through three methods: log monitoring, incremental interface, and scheduled scanning, and combines this with the organizational business process topology to sort out related business chains. It introduces an update impact assessment model to determine the update granularity based on business importance and urgency. For example, in scenarios such as employee departures, it can synchronously adjust related information such as project responsibility relationships and document permissions, avoiding isolated updates. Relying on the graph database transaction mechanism for sequential updates, it also supports rollback and collaboration with multiple subsidiaries, reducing the resource consumption of full updates while ensuring real-time synchronization between the knowledge graph and business scenarios, thus solving the key problems of insufficient knowledge consistency and poor real-time performance.

[0071] 2. This invention constructs a three-level dynamic terminology library consisting of a foundational layer, an industry layer, and an organizational layer. It integrates a business semantic pre-trained model with a graph neural network to accurately extract organizations-specific entities, relationships, and attributes. In the knowledge fusion stage, it aligns multi-source data through a semantic similarity algorithm, resolves conflicts by combining organizational rules, provides differentiated ontology evolution strategies, significantly reduces semantic bias, lowers the cost of manually adjusting ontology and rules, and improves graph adaptability.

[0072] 3. The quality control module of this invention sets dual quality indicators in each core link. At the data level, it focuses on completeness and accuracy, while at the business level, it emphasizes semantic consistency and logical correlation. It combines a terminology database and anomaly detection model to identify and process errors in a tiered manner, and can also optimize business rules in reverse. The intelligent application module relies on the optimized knowledge graph, combined with user roles and actual business scenarios, to accurately push knowledge views and knowledge chains. It supports natural language queries, multi-business system integration, and multi-terminal adaptation, and can strictly control access permissions in confidential scenarios. It not only ensures the long-term quality of the knowledge graph, but also realizes cross-departmental and multi-role knowledge flow, providing support for organizational decision-making and process optimization, and effectively improving operational efficiency. Attached Figure Description

[0073] Figure 1This is a flowchart of the knowledge graph-based content organization and construction system of the present invention. Detailed Implementation

[0074] 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.

[0075] like Figure 1 As shown, this embodiment of the invention provides a knowledge graph-based content organization and construction system, including a collection module, a knowledge extraction module, a knowledge fusion module, a graph construction module, a dynamic update module, an intelligent application module, and a quality control module;

[0076] The acquisition module supports access to data sources such as databases, API interfaces, enterprise collaboration platforms, and business systems without the need for customized development. It supports rapid access via lightweight API protocols for micro and small enterprises, parallel access via distributed data sources for group enterprises, and access via dedicated lines for confidential data in government agencies.

[0077] It leverages deep learning models to automatically identify data source types and match transmission protocols (HTTP / JDBC), extracts business attributes from data sources, such as "ERP data - production line projects" and "CRM data - customer projects," and generates collection configurations with business tags based on the characteristics of the organization's industry. For example, ERP data from manufacturing enterprises is automatically labeled with industry-specific business tags such as "production line debugging" and "equipment maintenance."

[0078] Complete routine cleaning operations such as data deduplication and format standardization, and label the data with business semantic tags according to the organization's business rules; in the processing of classified data of government agencies, retain non-sensitive semantic tags such as approval number and processing status, and hide the specific approval content; the labeling results are associated with the organization's basic semantic library, which is the foundation layer of the three-level dynamic terminology library established by the subsequent knowledge extraction module, containing general terms such as personnel and departments, and directly connecting to the extraction engine of the knowledge extraction module.

[0079] The labeled data is encapsulated in JSON format, and the specific fields include: {"Data Source ID": "D-20240501-001", "Business Tag": ["Production Line Project", "Equipment Debugging"], "Semantic Tag": {"Entity Candidate": ["Production Line A", "Debugging Cycle"], "Relationship Candidate": ["Attribution", "Duration"]}, "Data Cleaning Status": "Deduplicated / Standardized", "Confidentiality Level": "0 (Non-Confidential) / 1 (Low Confidentiality) / 2 (Medium Confidentiality) / 3 (High Confidentiality)", "Associated Basic Semantic Library ID": "B-Term-0032", "Timestamp": "2024-05-01 14:30:00"};

[0080] After receiving the JSON data, the knowledge extraction module automatically parses the "entity candidate" and "relationship candidate" fields as the starting basis for extraction. If a field is missing, the collection module's retransmission mechanism is triggered (retransmission interval is 5 minutes, up to 3 times).

[0081] The knowledge extraction module relies on the labeled data from the acquisition module, combined with seed data such as organizational business manuals and process specifications, to mine organization-specific business terms and semantic variations, establishing a three-level dynamic terminology library: a basic layer, an industry layer, and an organizational layer. The basic layer contains general terms such as "project" and "personnel," the industry layer contains industry-specific terms such as "agile sprint" for technology companies and "production line debugging cycle" for manufacturing companies, and the organizational layer contains personalized terms such as "OKR iteration cycle = bi-weekly" for internet companies and "subsidiary project coding rules = region + year + serial number" for group companies. The terminology library is associated with business rules, such as "iteration version number format = V + year + serial number," and is automatically updated as organizational business rules change, while also providing core terminology resources for the knowledge fusion module.

[0082] By integrating BERT-business semantic pre-trained model and GNN, entities, relationships, and attributes are extracted from business-labeled data collected by the module. When extracting entities, organization-specific semantics are distinguished, such as "Agile Project - Sprint Cycle" and "Production Line Project - Equipment Debugging Duration". When extracting relationships, cross-business line semantic correspondences are associated, such as the relationship between "Requirement Delivery" in the R&D line and "Feature Launch" in the product line. When extracting attributes, industry characteristics are taken into account, such as "Project Iteration Count" for technology companies and "Production Line Compliance Days" for manufacturing companies. In cross-departmental project collaboration data processing, cross-business line related information such as "Technical Solution - Customer Requirements" and "R&D Progress - Marketing Promotion Plan" is automatically extracted.

[0083] Based on the three-tier dynamic terminology database and business rules of the organization, the system automatically detects semantic deviations in the extraction results, such as classifying "temporary project team" as "formal department". Differentiated evaluation thresholds are set for different business scenarios, such as the evaluation threshold for "project review" being higher than that for "daily document management". Low-confidence results are marked and manual review is supported. The review results are synchronized to the quality control module for filing, and the extraction model parameters are updated in reverse. At the same time, the accurate extraction results are transmitted to the knowledge fusion module.

[0084] The knowledge fusion module calls the three-level terminology library of the knowledge extraction module and automatically matches synonymous entities from different data sources using a semantic similarity algorithm, such as "production line project - manufacturing project" and "demand delivery rate - function launch rate". In the scenario of multiple subsidiaries of a group enterprise, the semantic relationship between the subsidiary and the group is established, such as the relationship between the subsidiary's "regional project" and the group's "cross-regional collaborative project", to ensure that the alignment result conforms to the group's unified business definition. The alignment process refers to the business relevance, such as "the correlation between monthly KPI and project sprint cycle is 85%", to improve the alignment accuracy.

[0085] It integrates organizational business rules and data source authority to resolve semantic conflicts. For example, in scenarios such as "project completion rate = number of requirements delivered / total number of requirements" for the R&D line and "project completion rate = number of features launched / planned number of features launched" for the product line, it generates collaborative calculation templates to incorporate both types of results into the calculation of project health indicators; it supports customized business rules, such as government agencies can configure a "prioritize retaining the approval rules at this level for conflicts involving classified data" resolution strategy; it generates conflict handling logs containing conflict type, resolution rule ID, and data source identifier, and pushes them to the quality control module;

[0086] When monitoring the compatibility of newly added knowledge with existing ontology in the knowledge extraction module, analyze its relevance to organizational business processes. For example, when adding the "Agile Sprint" entity, associate it with "OKR Tasks," "R&D Personnel," and "Test Reports." Based on the high-frequency associated data types synchronized in real time by the dynamic update module, prioritize recommending ontology extension schemes related to high-frequency entities. The ontology evolution of micro and small enterprises adopts a lightweight extension strategy, while the ontology evolution of group enterprises adopts a multi-sub-ontology collaborative evolution strategy. The fused knowledge after evolution is transferred to the graph construction module.

[0087] The graph construction module relies on the fused knowledge and organizational business process diagrams of the knowledge fusion module to generate an initial ontology with business-related logic, such as the hierarchical relationship of "project-sprint-OKR-R&D personnel" and "procurement-approval-supplier-equipment". It provides an organization-specific ontology template import function. The template for technology companies includes preset nodes of "iteration version" and "requirement pool", while the template for government agencies includes preset nodes of "approval process" and "processing time limit". The ontology generation process takes into account scenario-based requirements, such as strengthening the association logic of "project progress-problem record-improvement measures" in the "project review" scenario.

[0088] By integrating community discovery and business association analysis algorithms, the graph hierarchy and relationship weights are dynamically adjusted. For example, the association weight between "R&D Projects" and "OKR" is higher than that between "R&D Projects" and "Administrative Documents," and the association weight between "Production Line Projects" and "Equipment Debugging" is higher than that between "Production Line Projects" and "Meeting Minutes." In cross-departmental business scenarios, the cross-departmental node association paths are optimized to improve query efficiency. The graph structure can be visually adjusted, allowing users to manually fine-tune relationship weights based on business needs. The adjustment results are then updated to the knowledge recommendation engine of the intelligent application module.

[0089] It adopts a hybrid storage architecture of Neo4j (graph structure) and HBase (attribute data), which supports efficient storage and parallel querying of entities and business relationships, and provides transactional storage guarantee for dynamic updates of the dynamic update module. Small and micro enterprises adopt a single-machine hybrid storage architecture, group enterprises adopt a distributed cluster storage architecture, and government agencies adopt a confidential partition storage architecture. It has a built-in storage performance monitoring mechanism that automatically adjusts the storage sharding strategy according to the real-time data volume, and the monitoring data is transmitted to the quality control module.

[0090] The dynamic update module captures data changes with business tags in the acquisition chain in real time through three methods: log monitoring, incremental interfaces of business systems, and timed scanning. These tags include "employee departure - R&D department", "project cycle modification - production line project", and "approval process change - government services". The monitoring frequency for core business data (project progress, revenue data) is set to minute-level, while the monitoring frequency for non-core data (administrative documents, meeting minutes) is set to hour-level. The monitoring results are labeled with the data change type (addition, modification, deletion) and related business tags to provide data input for the association chain analysis.

[0091] The system utilizes the three-level dynamic terminology database of the knowledge extraction module and the ontology evolution rules of the knowledge fusion module. Based on the organizational business process topology, it breaks down the data into three layers: "node-association rules-constraints." It automatically analyzes the business chains associated with data changes. For example, the association chain for "employee departure" is "departure application → project unbinding → document permission revocation → approval link redirection," and the association chain for "project cycle modification" is "cycle adjustment → budget allocation update → supplier performance deadline adjustment → customer acceptance node update." An update impact assessment model is introduced, which combines historical update logs and the importance weight of business data to predict the degree of impact of updates on business processes. Incremental update granularity is dynamically generated based on impact and business urgency. High-impact, high-urgency updates adopt a full association chain coverage strategy, while low-impact, low-urgency updates adopt a core node priority strategy.

[0092] Based on the graph database transaction mechanism of the graph construction module, relevant knowledge is updated in the order of business association chains to ensure that the update process does not disrupt business logic. For example, after "project unbinding" is completed, "document permission revocation" is performed. Update logs are recorded, including update content, scope of business impact, and execution time, supporting update rollback and recovery. At the same time, the update results are transmitted to the application interface of the intelligent application module and the quality control module. In the scenario of multiple subsidiaries in a group enterprise, subsidiary updates are coordinated with group updates. The subsidiary update process verifies the group semantic rules of the knowledge fusion module to avoid inconsistencies in knowledge at the group level.

[0093] The intelligent application module relies on the graph structure data of the graph construction module and the real-time update results of the dynamic update module. Combined with user roles (R&D manager, production director, government approval personnel) and business scenarios (project review, production line optimization, approval processing), it automatically identifies knowledge requirements. For example, the R&D manager - "Agile project sprint review" scenario requires knowledge related to "sprint cycle - requirement delivery rate - OKR completion rate", the production director - "production line optimization" scenario requires knowledge related to "equipment debugging time - production line compliance days - product qualification rate", and the government approval personnel - "matter processing" scenario requires knowledge related to "approval process - processing time - prerequisite materials". In the scenario of simplified roles for micro and small enterprises, it supports the adaptation of one person to multiple roles and automatically aggregates knowledge related to multiple roles.

[0094] Based on the optimized business relationship graph built by the graph construction module, corresponding knowledge views and knowledge chains are pushed to different roles. For example, R&D personnel are pushed a knowledge chain of "iteration version - requirement document - test report", marketing personnel are pushed a knowledge chain of "customer needs - product function - promotion plan", and government service personnel are pushed a knowledge chain of "item guide - approval rules - common problems". In cross-departmental project collaboration scenarios, "customer feedback" knowledge from the marketing department is pushed to R&D personnel, and "technical bottleneck" knowledge from the R&D department is pushed to marketing personnel. The recommendation results are dynamically optimized by combining user interaction data to improve the accuracy of the recommendation model and optimize the data transmission to the quality control module.

[0095] It provides interfaces for natural language query, business correlation analysis, and visualization, supporting integration with business systems such as ERP / CRM / OA. Interface data directly calls the storage architecture of the graph construction module and the update results of the dynamic update module. In OA system integration, a graph query entry is embedded, allowing users to query project-related personnel and historical approval records when initiating project approvals. It supports multi-terminal adaptation, with optimized knowledge display format for mobile interfaces and support for complex graph analysis for PC interfaces. The government agency interface is configured with access control logic for classified data, allowing only authorized personnel to query sensitive knowledge, and the permission rules are consistent with the classified data processing strategy of the collection module.

[0096] The quality control module sets quality indicators at each stage. At the data level, these include completeness (field missing rate < 5%) and accuracy (semantic annotation accuracy > 95%). At the business level, they include semantic consistency (entity classification conforms to organizational semantic rules) and logical association ("project cycle ≥ equipment debugging time"). Differentiated quality thresholds are set for different organizational types, such as higher data accuracy thresholds for government agencies than for ordinary enterprises. The module receives monitoring data from each module in real time and generates quality reports, including indicator values, anomaly markers, and business impact analysis, such as "semantic annotation accuracy 92%, affecting the accuracy of the knowledge recommendation model in the 'project review' scenario."

[0097] The system invokes the three-level dynamic terminology database of the knowledge extraction module and the conflict handling rules of the knowledge fusion module. Based on organizational business rules and anomaly detection models, it automatically identifies semantic errors and related logical errors. For high-frequency error scenarios (semantic tag confusion, missing relationships), it generates error type statistical reports, including error frequency, involved modules, and typical cases. It supports graded error handling, automatically correcting minor errors (non-core data format deviations). For serious errors such as core semantic conflicts and business logic contradictions (distinguished from low-confidence data verification in the knowledge extraction module), it pushes them for manual review. The review results record the error cause, handling plan, and executor information.

[0098] Correctable errors are automatically handled, and the handling solutions are used to update the business rule base, optimize the semantic annotation rules of the data collection module, and adjust the extraction model parameters of the knowledge extraction module. The results of complex error handling are fed back to the corresponding core modules (such as the data collection module, knowledge extraction module, knowledge fusion module, etc.). For example, the problem of "misclassification by temporary project team" is fed back to the knowledge extraction module to optimize the entity classification model, and the problem of "semantic conflict after update" is fed back to the data collection module to adjust the business semantic labeling rules. A closed loop of quality control and business rule optimization is established, quality data is analyzed regularly, control strategies are optimized, and the long-term quality stability indicators of the knowledge graph are improved. At the same time, the optimization strategies are passed to the update evaluation models of the knowledge fusion module and the dynamic update module.

[0099] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0100] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A knowledge graph-based content organization and construction system, characterized in that, It includes a data acquisition module, a knowledge extraction module, a knowledge fusion module, a knowledge graph construction module, a dynamic update module, an intelligent application module, and a quality control module; Data Acquisition Module: Adapts to access from multiple data sources, including databases, API interfaces, enterprise collaboration platforms, and business systems. The data source is text, providing organizational scenario-based access methods, intelligently identifying data source attributes and generating business tags, and completing data cleaning and semantic annotation. Knowledge extraction module: Constructs a three-level dynamic terminology database and updates it in real time; integrates pre-trained models and graph neural network technology to extract knowledge elements; detects semantic biases and optimizes the data before passing it to the knowledge fusion module. The knowledge fusion module calls the terminology library to complete the semantic alignment of multi-source data, resolves conflicts by combining organizational rules and records logs, monitors the compatibility of newly added knowledge with the ontology and recommends ontology evolution schemes, and transfers the fused knowledge to the graph construction module. The graph construction module generates an initial ontology and optimizes the graph structure based on integrated knowledge and business processes, provides a hybrid storage architecture that adapts to different organizations, and outputs graph data and adjustment results. Dynamic update module: Real-time capture of changes in data collection links, calls terminology base and ontology rules to analyze related business chains, assesses the impact of updates and determines update strategies, and updates knowledge sequentially based on storage transaction mechanism; Intelligent application module: Based on the graph structure and update results, it identifies users' knowledge needs, pushes knowledge content and optimizes recommendation accuracy, provides multiple types of interfaces to achieve business system integration, and adapts to multiple terminals and confidential data access rules; Quality control module: Set dual-dimensional quality indicators for each core link, monitor and generate quality reports in real time, identify semantic and logical errors and handle them in a graded manner, and push the optimization plan back to the aforementioned module.

2. The knowledge graph-based content organization and construction system according to claim 1, characterized in that, The data acquisition module supports access to data sources including databases, API interfaces, enterprise collaboration platforms, and business systems, and provides differentiated access solutions adapted to micro and small enterprises, group enterprises, and government agencies. It automatically identifies data source types and matches transmission protocols based on deep learning models, extracts data source business attributes, and generates collection configurations with business tags by combining the extracted business attributes with industry characteristics; Complete data deduplication and format standardization cleaning operations, and label data with business semantic tags according to organizational business rules; retain non-sensitive tags and hide sensitive content for classified data in accordance with the rules. The annotation results are associated with the organization's basic semantic library and connected to the extraction engine of the knowledge extraction module.

3. The knowledge graph-based content organization and construction system according to claim 2, characterized in that, The knowledge extraction module relies on the labeled data from the acquisition module and combines it with seed data, including organizational business manuals and process specifications, to mine organizational-specific business terms and semantic variations, and establish a three-level dynamic terminology library of basic, industry, and organizational layers. The terminology library is associated with business rules and is automatically updated as organizational business rules change. By integrating pre-trained models and graph neural networks, entities, relationships, and attributes are extracted from data with business labels; when extracting entities, organization-specific semantics and cross-business line related information are distinguished. Semantic bias in the extraction results is detected based on a three-level dynamic terminology database and business rules; differentiated evaluation thresholds are set for different business scenarios; low-confidence results are labeled and manual review is supported; the review results are used to update the extraction model parameters; and the extraction results are passed to the knowledge fusion module.

4. The knowledge graph-based content organization and construction system according to claim 3, characterized in that, The knowledge fusion module calls a three-level terminology database and combines a semantic similarity algorithm to automatically match synonymous entities from different data sources, while taking into account business relevance to ensure that the alignment results conform to the organization's business definition. It integrates organizational business rules and data source authority to resolve semantic conflicts, supports customized business rules, generates conflict handling logs, and pushes them to the quality control module; Monitor the compatibility of newly added knowledge with existing ontology, recommend ontology extension schemes based on high-frequency associated data types, provide ontology evolution strategies adapted to different organizational types, and transfer the evolved fused knowledge to the graph construction module.

5. The knowledge graph-based content organization and construction system according to claim 4, characterized in that, The graph construction module generates an initial ontology with business-related logic based on integrated knowledge and organizational business process diagrams; it supports the import of organization-specific ontology templates with preset nodes. The graph hierarchy and relationship weights are dynamically adjusted through a fusion algorithm; cross-scene node association paths are optimized; and the graph structure can be manually adjusted visually, with the adjustment results updated to the intelligent application module. It adopts a hybrid storage architecture of Neo4j and HBase, supports efficient storage and parallel querying of entities and business relationships, and provides storage deployment solutions that are adaptable to different organizations; it has a built-in storage performance monitoring mechanism that automatically adjusts the storage sharding strategy based on the amount of data and transmits the monitoring data to the quality control module.

6. The knowledge graph-based content organization and construction system according to claim 5, characterized in that, The dynamic update module captures data changes with business tags in the collection link through three methods: log monitoring, incremental interface, and timed scanning. It sets differentiated monitoring frequencies according to business importance and marks the data change type with associated business tags. It calls upon a three-level dynamic terminology library and ontology evolution rules, and analyzes the business chain related to data changes based on the organization's business process topology; it introduces an update impact assessment model, combines historical update logs and business data importance weights to predict the degree of impact of updates on business processes; and generates incremental update granularity based on impact and business urgency. Based on the graph database transaction mechanism, relevant knowledge is updated according to the order of business relationship chains; It records update logs and supports update rollback and recovery, while also transmitting update results to the intelligent application module and the quality control module; it enables update collaboration between subsidiaries and the group in multi-subsidiary scenarios.

7. The knowledge graph-based content organization and construction system according to claim 6, characterized in that, The intelligent application module relies on graph structure data and real-time update results to identify knowledge needs by combining user roles and business scenarios, and supports multi-role requirement adaptation and knowledge aggregation. Based on the optimized business association graph constructed by the graph construction module, corresponding knowledge views and knowledge chains are pushed to different roles. Recommendation results are dynamically optimized based on user interaction data, and the optimized data is then transmitted to the quality control module. It provides interfaces for natural language query, business correlation analysis, and visualization, supports integration with business systems, is compatible with multiple terminals, and allows configuration of access control logic in confidential scenarios, with permission rules consistent with the confidential data processing strategy of the collection module.

8. The knowledge graph-based content organization and construction system according to claim 7, characterized in that, The quality control module sets quality indicators at each stage, including completeness and accuracy at the data level and semantic consistency and logical correlation at the business level; it sets differentiated quality thresholds for different organizational types and generates quality reports in real time. It calls upon a three-level dynamic terminology database and conflict handling rules, and combines organizational business rules and anomaly detection models to identify semantic errors and related logical errors. Supports error classification and statistical analysis; Correctable errors are handled automatically, and the handling solutions are updated in reverse to update the business rule base; the results of complex error handling are fed back to the corresponding front-end modules; a closed loop of quality control and business rule optimization is established, and control strategies are optimized regularly and transmitted to relevant modules.

Images