Business object driving-based knowledge map construction method, system and medium

Abstract

The application relates to the technical field of natural language processing, and discloses a knowledge map construction method and system based on business object driving and a medium. The method comprises the following steps: multi-source knowledge access and preprocessing; multi-modal document structural analysis; construction based on four types of knowledge anchors of entities, events, attributes and scenes; theme aggregation with the knowledge anchors as cores, including rule rough recall, semantic fine arrangement, candidate cluster merging and knowledge type tendency identification; fine-grained knowledge extraction and quality checking according to knowledge types; semantic normalization and multi-dimensional knowledge connection construction; configuration of knowledge management and service release. The knowledge map construction method unifies the mapping of multi-source heterogeneous knowledge into a knowledge map with business objects as cores through a step-by-step convergence construction mode, and improves the retrieval efficiency, question and answer accuracy, source traceability and operability of enterprise knowledge.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, particularly intelligent question answering technology, and more specifically to a knowledge map construction method, system, and medium based on business object-driven approaches. Background Technology

[0002] Currently, in multi-entity collaborative business scenarios such as industrial manufacturing and enterprise management, a core business often manifests as a highly dynamic, long-chain, multi-step cascading process, involving complex temporal dependencies and densely associated with knowledge content contained in multi-source heterogeneous documents, such as paragraphs, tables, images, question-and-answer pairs, and rule entries. Existing technologies for constructing knowledge graphs or knowledge bases typically organize them by document directory or keywords, lacking an organization method centered on business objects, resulting in weak connections between knowledge and difficulties in governance. Traditional knowledge base construction and retrieval methods face numerous challenges in addressing these complex business scenarios characterized by "long chains and high dynamism."

[0003] First, existing knowledge graph solutions are mostly based on static triples of "entity-relationship-entity" or flattened text fragments to organize knowledge. Whether it is a knowledge base built by directly indexing keywords in the original document or a knowledge graph or question-answer base built by extracting FAQs and keywords from the original document, they lack the analysis of document structure and the construction of intermediate semantic layers. When faced with complex cross-document and multimodal corpora, the insufficient semantic decoupling of high-dimensional heterogeneous text leads to low retrieval recall and severe noise diffusion. When existing models process heterogeneous corpora containing tables, illustrations, and multi-level chapters, the lack of a multi-dimensional feature anchoring mechanism between the original text and knowledge nodes often results in extremely high data noise and semantic bias when directly extracted, leading to low retrieval recall and severe ambiguity in synonymous heterogeneous texts in high-dimensional vector spaces. Meanwhile, traditional topology networks are mainly built based on static triples. The extracted content lacks explicit modeling of dynamic elements such as preconditions, post-dependencies, temporal progression, and parallel branches. This results in fragmented and breakpoint-prone knowledge retrieval outputs involving long links and multi-step execution processes, making it difficult to meet the executability requirements of complex business scenarios.

[0004] Secondly, while knowledge content is continuously updated along with business rules, the existing classification and labeling system or single relational topology fails to establish a standardized metadata governance layer for knowledge units. It cannot synchronously represent governance attributes such as source traceability, responsibility attribution, access permissions, valid version, and review and transfer status. This makes it difficult for the system to adaptively identify timeliness and permission boundaries during deep retrieval or multi-round reasoning, which can easily lead to problems such as conflicts between old and new knowledge and unauthorized access. The credibility and compliance of the retrieval results are difficult to guarantee.

[0005] With the expansion of cross-modal and cross-source data, enterprises have accumulated a large amount of multi-source heterogeneous knowledge, including customer lists, work order records, policy documents, operation manuals, FAQs, training materials, case documents, tables, images, and scanned documents. This knowledge is scattered across different business systems, with varying formats and structures. Unstructured text data exhibits high heterogeneity, fragmentation, and long-chain characteristics. Existing knowledge graph construction technologies have structural defects in three dimensions: procedural knowledge representation, dynamic governance assurance, and multi-layer concept decoupling. These limitations make it difficult to effectively support the high-precision retrieval and robust multi-turn question-answering reasoning requirements in complex business scenarios with long chains and multiple steps. Summary of the Invention

[0006] In view of the problems and shortcomings of existing technologies, the purpose of this invention is to provide a knowledge map construction method based on business object-driven approach. This method uses business objects as the main thread, knowledge forms as the carrier, semantic relationships as the connection, and governance attributes as the guarantee, and maps them uniformly into a three-dimensional topology network of business, semantics, and retrieval. It simultaneously realizes the explicit expression of procedural knowledge, the strong constraint guarantee of governance attributes, and the decoupling and accurate mapping of multi-layer concepts at the model level, effectively overcoming the problems and bottlenecks of insufficient retrieval accuracy and easy breakage of inference links in existing technologies in long-link complex business scenarios.

[0007] According to a first aspect of the present invention, a knowledge map construction method based on business object-driven approach is proposed, comprising the following steps:

[0008] Step S1: Access multi-source heterogeneous knowledge data through the data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries;

[0009] Step S2: Perform structured parsing on the standardized text segment, identify the title, paragraph, table and image in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position.

[0010] Step S3: Based on the preset business object system, extract knowledge anchors from the paragraph-level content nodes as the subject intermediate layer connecting the original documents and deep knowledge. The knowledge anchors include entity anchors, event anchors, attribute anchors and scene anchors, and record the source evidence chain corresponding to each knowledge anchor simultaneously.

[0011] Step S4: Based on the extracted knowledge anchors, perform topic aggregation on the paragraph-level content nodes through multi-level recall and clustering, aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and determine the target knowledge type corresponding to each topic cluster;

[0012] Step S5: Based on the target knowledge type, distribute each topic cluster to the corresponding knowledge extraction process, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes;

[0013] Step S6: For each topic cluster, based on the enterprise standard lexicon and the synonym disambiguation model, the knowledge nodes are conceptually and semantically normalized to a unified business object. Using the normalized business object as the core anchor point, a multi-dimensional knowledge connection between the knowledge nodes within the topic cluster is established using graph neural networks or heuristic rules to complete the construction of the knowledge map.

[0014] As an optional implementation, the method further includes the following steps:

[0015] Step S7: Configure governance metadata for the relationship edges of knowledge nodes and multidimensional knowledge connections, and generate differentiated knowledge service views based on the multidimensional matrix of user roles, affiliated organizations, associated business objects, current scenarios and knowledge types, and complete the governance attribute configuration and knowledge map service publication.

[0016] As an optional implementation, the method further includes the following steps:

[0017] Step S8: Capture operational signals during the operation of the knowledge map, including at least one of the following: unmatched query records, incorrect answer records, high-frequency clicks, knowledge reuse rate, low evaluation feedback, and version aging risk. Input the operational signals back into the knowledge map building engine to trigger updates to the business object system, knowledge anchor extraction rules, topic aggregation strategy, concept thesaurus, or extraction template.

[0018] The updates include at least one of the following: updating the business object system, revising the knowledge anchor extraction rules, adjusting the topic aggregation strategy, expanding the concept lexicon, optimizing the extraction template, and activating manual review.

[0019] According to a second aspect of the present invention, a knowledge map construction system based on business object-driven architecture is proposed, comprising:

[0020] The knowledge access module is used to access multi-source heterogeneous knowledge data through a data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries;

[0021] The document parsing module is used to perform structured parsing on the standardized text segments, identify titles, paragraphs, tables and images in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position.

[0022] Anchor point construction module is used to extract knowledge anchor points from the paragraph-level content nodes based on a preset business object system, serving as a thematic intermediate layer connecting original documents and deep knowledge. The knowledge anchor points include entity anchor points, event anchor points, attribute anchor points, and scene anchor points, and synchronously record the source evidence chain corresponding to each knowledge anchor point.

[0023] The topic aggregation module is used to perform topic aggregation on the paragraph-level content nodes based on the extracted knowledge anchors through multi-level recall and clustering, to aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and to determine the target knowledge type corresponding to each topic cluster.

[0024] The knowledge extraction module is used to distribute each topic cluster to the corresponding knowledge extraction process according to the target knowledge type, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes.

[0025] The multidimensional knowledge connection construction module is used to normalize the conceptual semantics of knowledge nodes into unified business objects based on the enterprise standard lexicon and the synonym disambiguation model for each topic cluster. Then, using the normalized business objects as the core anchor points, it establishes multidimensional knowledge connections between knowledge nodes within the topic cluster using graph neural networks or heuristic rules, thus completing the construction of the knowledge map.

[0026] According to a third aspect of the present invention, a computer system is provided, comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the aforementioned knowledge map construction method based on business object-driven architecture.

[0027] In a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the aforementioned knowledge map construction method based on business object-driven principles.

[0028] Combining the business object-driven knowledge map construction method and system design of the above embodiments, based on multi-source heterogeneous knowledge data, and through a step-by-step convergence approach of document parsing, knowledge anchor extraction, knowledge topic aggregation, and fine-grained extraction, this approach changes the traditional practice of directly extracting FAQs and knowledge points from original documents in a coarse manner, making the knowledge generation path more interpretable and stable. By introducing four types of unified knowledge anchors—entities, events, attributes, and scenarios—the content in paragraphs, tables, and images is uniformly mapped to the intermediate semantic layer, enhancing the consistency of knowledge organization and the traceability of its sources.

[0029] In the method of this invention, based on the extraction of knowledge anchors and content nodes, a topic aggregation mechanism is used, combined with rule-based coarse recall, vector similarity cleaning, overlapping candidate cluster merging, and cluster type tendency identification. This allows candidate content to be screened and merged before fine-grained knowledge node extraction, and subsequent concept semantic unification and merging are carried out. Multidimensional knowledge connections are performed under the constraints of whitelist rules, thereby reducing the risk of duplicate knowledge generation and semantic deviation, eliminating the polysemy and ambiguity of proper nouns, and establishing an interpretable and manageable network association between heterogeneous knowledge nodes.

[0030] The knowledge map construction method of this invention can unify heterogeneous documents, chapters, paragraphs, tables, images, videos, attachments, FAQs, knowledge points, processes, cases, terms and relationships from different sources into the knowledge nodes of the same knowledge map. Driven by business objects, a network of relationships is formed, so that any knowledge node simultaneously has business coordinates, morphological coordinates, semantic coordinates and governance coordinates, thus completing the construction of the knowledge map and improving the efficiency of enterprise knowledge retrieval, question and answer accuracy, traceability and operability.

[0031] Furthermore, after the knowledge map is constructed, through governance attribute configuration, decentralized view publication, and operational dynamic feedback loop, the full lifecycle management of the knowledge map can be realized, enabling it to form a closed-loop self-iterative mechanism and dynamic evolution based on actual usage.

[0032] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0033] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0034] Figure 1 The overall flowchart of the knowledge map construction method based on business object driving provided in the embodiments of the present invention is shown.

[0035] Figure 2 This is a schematic diagram of the process for extracting knowledge anchors based on content nodes, provided in an embodiment of the present invention.

[0036] Figure 3 This is a schematic diagram of the multi-source knowledge topic aggregation process provided in an embodiment of the present invention.

[0037] Figure 4 This is a schematic diagram illustrating the process of extracting knowledge nodes from a topic cluster, as provided in an embodiment of the present invention.

[0038] Figure 5 This is a flowchart illustrating a knowledge map construction method based on business object-driven approach, as provided in another embodiment of the present invention.

[0039] Figure 6 This is a schematic diagram of the knowledge governance and publishing process provided in an embodiment of the present invention.

[0040] Figure 7 The structural block diagram of a knowledge map construction system based on business object-driven architecture provided in this embodiment of the invention. Detailed Implementation

[0041] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0042] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0043] Example 1

[0044] Combination Figures 1 to 4 As shown, the knowledge map construction method based on business object-driven approach according to an embodiment of the present invention includes the following steps:

[0045] Step S1: Access multi-source heterogeneous knowledge data through the data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries;

[0046] Step S2: Perform structured parsing on the standardized text segment, identify the title, paragraph, table and image in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position.

[0047] Step S3: Based on the preset business object system, extract knowledge anchors from the paragraph-level content nodes as the subject intermediate layer connecting the original documents and deep knowledge. The knowledge anchors include entity anchors, event anchors, attribute anchors and scene anchors, and record the source evidence chain corresponding to each knowledge anchor simultaneously.

[0048] Step S4: Based on the extracted knowledge anchors, perform topic aggregation on the paragraph-level content nodes through multi-level recall and clustering, aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and determine the target knowledge type corresponding to each topic cluster;

[0049] Step S5: Based on the target knowledge type, distribute each topic cluster to the corresponding knowledge extraction process, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes;

[0050] Step S6: For each topic cluster, based on the enterprise standard lexicon and the synonym disambiguation model, the knowledge nodes are conceptually and semantically normalized to a unified business object. Using the normalized business object as the core anchor point, a multi-dimensional knowledge connection between the knowledge nodes within the topic cluster is established using graph neural networks or heuristic rules to complete the construction of the knowledge map.

[0051] As an optional embodiment, in step S1, multi-source heterogeneous knowledge data from within and outside the enterprise are uniformly accessed through a data interface to build a standardized basic knowledge foundation that is parseable, traceable, and governable.

[0052] As an optional example, the following sub-steps are included:

[0053] Step S11: Unify the access of heterogeneous data from multiple sources, including customer lists, work order records, policy documents, operation manuals, FAQs, training materials, case documents, tables, images, and scanned documents, into the knowledge processing workflow; during the access process, automatically extract and record the basic metadata information of each data source, including the source system, update time, scope of application, usage restrictions, and permission boundaries.

[0054] Step S12: Perform data preprocessing and cleaning on the extracted data, including: performing data deduplication, version identification, and sensitive information identification on multi-source heterogeneous data, and outputting standardized unstructured or semi-structured text segments in a unified format as standardized input for subsequent knowledge extraction and connection construction.

[0055] The aforementioned data deduplication, version identification, and sensitive information identification can be implemented using typical algorithms from existing technologies.

[0056] Further, in step S2, the structured parsing of the multimodal document is performed.

[0057] In this step, the preprocessed original document is subjected to multimodal structured parsing, which transforms the content of different carriers into standardized text segments (i.e., standard content units) that can be extracted later.

[0058] In this example, commercially available multimodal large models (such as GPT-4V, Qwen-VL, or LayoutLM) can be used to identify the text content in standardized text segments, including information such as titles, paragraphs, tables, and images. Tables include table titles, and images include image titles and image descriptions. Finally, paragraph-level content nodes with hierarchical topology and traceability are generated.

[0059] At the same time, each content node is associated with its original document identifier, chapter identifier, and paragraph position, forming a traceable hierarchical topology.

[0060] Furthermore, in step S3, business object-driven knowledge anchors are constructed based on paragraph-level content nodes. As a specific example, business object-driven knowledge anchors are extracted from paragraph-level content nodes based on a preset business object system.

[0061] In this embodiment, the preset business object system includes entity anchors, event anchors, attribute anchors, and scene anchors, combined with... Figure 2 As shown, the specific steps for extracting knowledge anchors are as follows:

[0062] Step S31: Extract entity elements such as equipment, products, organizations, customers, and process objects from the parsed text content of titles, paragraphs, tables, and images to form entity anchors;

[0063] Step S32: Identify the business action elements such as maintenance, fault handling, and disposal contained in the text content to form event anchors used to describe business dynamics;

[0064] Step S33: Identify the feature descriptions such as maintenance cycle, risk points, and operation steps contained in the text content to form attribute anchors to carry the attribute features of objects and events;

[0065] Step S34: Identify the feature descriptions in the text content that represent applicable conditions and scenarios, such as power outage maintenance, routine inspection, and high-load operation in summer, and form scenario anchors to express the boundaries of knowledge application and role context.

[0066] Simultaneously, during the extraction of knowledge anchors, the source evidence chain corresponding to each extracted knowledge anchor is recorded, including the source document ID, chapter number, paragraph number, table number, or image number, to achieve source tracing of knowledge anchors.

[0067] Further, in step S4, based on the extracted knowledge anchors, the paragraph-level content nodes are aggregated into topics through multi-level recall and clustering. Content fragments from multi-source heterogeneous knowledge data are aggregated into the same topic cluster, and the target knowledge type corresponding to each topic cluster is determined. Figure 3 As shown, the specific steps include:

[0068] Step S41: Construct rule query keys based on the knowledge anchors and tag information, perform coarse recall on each paragraph-level content node, and obtain a candidate set;

[0069] Step S42: Perform semantic ranking on the candidate set using a pre-trained text model;

[0070] Step S43: Merge the candidate clusters in the ranked candidate set according to the overlap and semantic distance between the candidate clusters to obtain the topic cluster set;

[0071] Step S44: Identify the knowledge type tendency for each topic cluster within the topic cluster set, and determine the target knowledge type corresponding to each topic cluster.

[0072] Therefore, by aggregating diverse knowledge topics, with knowledge anchors as the core, multi-level recall and clustering algorithms are used to aggregate scattered similar content units into high-quality candidate topic clusters. In cross-document scenarios, content fragments of the same object, the same phenomenon, the same operation, or the same business scenario are merged into the same topic space, thereby improving the completeness and consistency of fine-grained knowledge node extraction such as FAQs, knowledge points, processes, cases, and terms.

[0073] As an optional example, in step S41, a rule query key is constructed based on the knowledge anchors and tag information to perform coarse recall on each paragraph-level content node to obtain a candidate set, specifically including the following steps:

[0074] For any candidate paragraph Based on its knowledge anchor set and tag set, a rule query key q is constructed, and the candidate paragraph p is calculated. i Relative to the rule query key coarse recall relevance R(q,p) i ):

[0075] R(q,p i )=α·I anchor (q,p i )+β·I tag (q,p i );

[0076] Among them, I anchor (q,p i ) represents the anchor matching score, I tag(q,p i The score is the tag matching score, where α and β are the anchor matching weight and the tag matching weight, respectively, and α+β=1.

[0077] Based on the preset first threshold, the coarse recall relevance R(q,p) is... i Candidate paragraphs that meet the preset first threshold condition are added to the candidate set.

[0078] As an optional example, in step S41, the preset first threshold value is 0.8, that is, when a candidate paragraph meets the following conditions, it is included in the coarse recall candidate set C corresponding to the rule query key q. q :

[0079] R(q,p i ≥0.8;

[0080] C q ={p i |R(q,p i )≥0.8}.

[0081] As an optional example, the aforementioned anchor matching score I anchor (q,p i It is set to be calculated in the following manner:

[0082] Let A be the set of anchor points corresponding to the rule query key q. q for:

[0083] A q ={E q V q ,S q U q};

[0084] Among them, E q V q ,S q U q These are the entity set, event set, scenario set, and attribute set corresponding to the rule query key q, respectively.

[0085] Let the candidate set C be... q Each candidate paragraph p in i The corresponding anchor point set A i for:

[0086] A i ={E i V i ,S i U i};

[0087] Among them, E i V i ,S iU i The candidate paragraphs p are respectively i The corresponding entity set, event set, scene set, and attribute set;

[0088] Define the set overlap function J(X,Y) as follows:

[0089] J(X,Y)=|X∩Y| / (|X∩Y|+ε);

[0090] In the formula, ε is a local constant used to prevent the denominator from being 0;

[0091] Then calculate J(E) according to the set overlap function J(X,Y). q E i ), J(V q V i ), J(S q ,S i ), J(U q U i ), and the anchor point matching score I anchor (q,p i ) is defined as:

[0092] I anchor (q,p i )=w e ·J(E q E i )+w v J(V q V i )·+w s J(S q ,S i )·+w u ·J(U q U i );

[0093] Among them, w e ,w v ,w s ,w u Let w represent the weights of the entity, event, scenario, and attribute dimensions, respectively, and satisfy: w e +w v +w s +w u =1.

[0094] As an optional example, the aforementioned label matching score I tag (q,p i It is set to be calculated in the following manner:

[0095] Let T be the set of tags corresponding to the rule query key q. q Candidate paragraph p iThe corresponding tag set is T i Then the label matching score I tag (q,p i ) is defined as:

[0096] ;

[0097] In the formula, the label similarity Sim tag (t a ,t b The calculation rules for ) are as follows:

[0098] ;

[0099] Among them, t a Representing the label set T respectively q The a-th label in the set T has a value of 1. i The bth tag;

[0100] , The labels are t respectively a ,t b The semantic vector, where ρ is the scaling factor.

[0101] As an optional example, in step S42, semantic ranking of the candidate set is performed using a pre-trained text model, including:

[0102] The candidate set obtained from coarse recall is vectorized using a pre-trained text model. Noisy data that deviates from the topic is then removed by calculating semantic similarity.

[0103] Sim(v q ,v i The calculation method is as follows: ;

[0104] Among them, v q v is the semantic vector of the rule query key q. i For candidate paragraph p i semantic vector;

[0105] For candidate set C q Each candidate paragraph p in i If the following expression is satisfied, retain it:

[0106] Sim(v q ,v i )≥τ;

[0107] Obtain the filtered candidate set C q ’ :

[0108] Cq ’ ={p i ∈C q |Sim(v q ,v i )≥τ};

[0109] In the formula, τ is the semantic filtering threshold. In this example, the semantic filtering threshold is taken as 0.85. When the candidate paragraph p i Semantic similarity Sim(v q ,v i If the following conditions are met, retain it:

[0110] C q ’ ={p i ∈C q |Sim(v q ,v i )≥τ};

[0111] That is, candidate paragraphs with a similarity of 0.85 or higher are retained to obtain the filtered candidate set C. q ’ .

[0112] As an optional implementation, in step S43, overlapping candidate clusters are merged and merged, and candidate clusters that have overlap, semantic similarity or cross sources are merged.

[0113] Specifically, by calculating the overlap and semantic distance between candidate clusters retrieved by different rule query keys, and merging candidate clusters in the refined candidate set based on the overlap and semantic distance between candidate clusters, a topic cluster set is obtained, including the following steps:

[0114] For set C q ’ Any two candidate clusters C a With C b Calculate its member overlap coefficient J(C) a C b ), and calculate the center vector μ of each candidate cluster respectively. a μ b Calculate the semantic similarity Sim(C) between two candidate clusters. a C b ),in:

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] Then, based on the merge decision function Merge(C) a C b The value of ) is determined by the merge decision function Merge(C). a C b When the value of ) meets the preset second threshold condition, the candidate cluster C is selected. a With C b Merging them yields a new topic cluster C. k ;

[0120] Among them, the merge decision function Merge(C) a C b )=f(J(C a C b Sim(C) a C b Specifically, it means:

[0121] ;

[0122] Here, λ is the balance parameter between the overlap coefficient and semantic similarity, and η is the candidate cluster merging threshold. For example, η = 0.8.

[0123] When the merge decision function Merge(C) is used a C b When the value of ) is 1, candidate cluster C is selected. a With C b Merging them yields a new topic cluster C. k .

[0124] Finally, all the merged topic clusters are recorded as a topic cluster set C:

[0125] C = {C1, C2, ..., C} K}

[0126] Where K represents the total number of all topic clusters in the topic cluster set.

[0127] Furthermore, in step S44, the target knowledge type tendency of the topic cluster is identified: based on the content form, trigger anchor type and semantic structure features within the candidate cluster, the target knowledge type of the topic cluster is predicted by a probability classification model.

[0128] Specifically, by extracting anchor features, structural features, and semantic features of each topic cluster, and then using a probabilistic classification model (such as random forest or BERT classifier) ​​to predict the probability distribution of the topic cluster on the knowledge type set, the knowledge type corresponding to the maximum probability is determined as the target knowledge type of the topic cluster.

[0129] As an example, the knowledge type set in this embodiment of the invention includes FAQs, knowledge points, processes, cases, and terms, specifically represented as: T={FAQ, knowledge points, processes, cases, terms}, where T represents the knowledge type set.

[0130] For any topic cluster C k It can be extracted by its anchor point features, structural features, and semantic features, denoted as f respectively. anchor (C k ), f structure (C k ), f semantic (C k If the topic cluster C is... k The target knowledge type is defined as:

[0131] Type*(C k =argmaxP(t|f anchor (C k ),f structure (C k ),f semantic (C k ));

[0132] For each candidate knowledge type t, the BERT classifier P() outputs a probability value; argmaxP(t| ) represents the knowledge type t corresponding to the maximum probability value returned by the BERT classifier output, and the topic cluster C k The target knowledge type, t∈T.

[0133] In embodiments of the present invention, for topic clusters C k Anchor point features f anchor ( C k This is used to characterize the type composition of knowledge anchors within a topic cluster. The topic cluster can be analyzed first. C k The text fragments within are segmented and tagged with parts of speech. After retaining noun and verb phrases, the TF-IDF algorithm is used to calculate the weight of each phrase, and the top-weighted phrases are selected. N The top 10 (e.g.) are used as the candidate anchor set. Sentence-BERT is used to encode each candidate anchor to obtain a sentence vector. Finally, the sentence vectors of the candidate anchors are weighted and averaged according to TF-IDF weights to obtain the anchor features of the topic cluster.

[0134] Structural features f structure (C k ) is used to characterize the topic cluster C kThe organization of internal content nodes can be obtained based on the hierarchical topology and source location information of paragraph-level content nodes. For example, statistical analysis can be performed on at least one of the following: title hierarchy distribution, proportion of paragraphs / tables / images, density of step numbers or list items, and number of question-answer pairs. The statistical results can then be normalized to obtain a structural statistical vector.

[0135] Semantic feature f semantic (C k The semantic vector is used to represent the overall semantic content of the topic cluster. The text, title, anchor name and / or attribute content of each content node in the topic cluster can be vectorized and encoded using a pre-trained text model to obtain the semantic vector of each node. Then, the semantic vector of each node in the topic cluster is averaged, weighted averaged or the center vector is taken to obtain the cluster-level semantic vector.

[0136] f anchor (C k ), f structure (C k ), f semantic (C k After concatenation, the data is input into the BERT classifier, which predicts the output topic cluster C. k The probability distribution on the knowledge type set T is used to determine the knowledge type corresponding to the maximum probability and identify it as the target knowledge type for the topic cluster.

[0137] For example, question-and-answer topics with a clear structure and semantically corresponding question-and-answer pairs tend to be more like FAQs; topic clusters with high step number density and dense event anchors tend to be more like processes; topic clusters with a set of terms, aliases, abbreviations, and definitional descriptions tend to be more like terms; topic clusters with dense attribute anchors such as definitions, rules, conditions, and parameters tend to be more like knowledge points; and topic clusters with a background-question-process-summary narrative structure tend to be more like case studies.

[0138] Therefore, when topic cluster C k Target knowledge type Type*(C k Once identified, the corresponding topic clusters can be distributed to the corresponding knowledge decomposition processes. If determined to be a FAQ, the process proceeds to the question and answer extraction process; if determined to be a knowledge point, the process proceeds to the rule and fact extraction process; if determined to be a process, the process proceeds to the operation step extraction process; if determined to be a case study, the process proceeds to the event review extraction process; and if determined to be a term, the process proceeds to the terminology definition extraction process.

[0139] Furthermore, in step S5, based on the target knowledge type, each topic cluster is distributed to the corresponding knowledge extraction process, and atomic-level knowledge nodes are deeply extracted from the aggregated topic clusters, combined with... Figure 4 As shown, the specific steps include:

[0140] When the target knowledge type is FAQ, execute the question and answer extraction process to extract standard questions, standard answers, and similar words and phrases;

[0141] When the target knowledge type is a knowledge point, the rule-based fact extraction process is executed to extract the core definition, triggering reason, judgment criteria and handling suggestions.

[0142] When the target knowledge type is a process, the operation steps are executed to extract the process, including the preconditions, execution steps, responsible roles, and output results.

[0143] When the target knowledge type is a case study, the event review and extraction process is executed to extract the business background, core issues, technical processes, and experience summaries.

[0144] When the target knowledge type is a term, the term definition extraction process is executed to extract standard terms, aliases, abbreviations, and superordinate and subordinate descriptions.

[0145] For each type of target knowledge, the extracted results are used as knowledge nodes of the topic cluster.

[0146] Furthermore, in step S5, the extracted knowledge nodes are also subjected to multi-dimensional quality verification. For example, a comprehensive quality assessment is performed on the extracted knowledge nodes, based on the source reliability score, extraction confidence score, and consistency verification score. When any score is lower than the corresponding preset threshold, the corresponding fine-grained knowledge node is marked as a risk node, triggering a manual review process. The source reliability score is calculated based on the metadata information of the data source to which the knowledge node belongs.

[0147] Furthermore, in step S6, for each topic cluster, based on the enterprise standard lexicon and the synonym disambiguation model, the knowledge nodes are semantically normalized to a unified business object. Using the normalized business object as the core anchor point, a multi-dimensional knowledge connection is established between the knowledge nodes within the topic cluster using graph neural networks or heuristic rules. Specifically, this includes the following processes:

[0148] Step S61: Based on the enterprise standard thesaurus and the synonym disambiguation model, identify the terms, abbreviations, aliases and industry terms in the knowledge nodes;

[0149] Construct a synonym merging mapping table and a hierarchical concept tree;

[0150] Based on the synonym merging mapping table and hierarchical concept tree, knowledge nodes are merged into a unified business object, and differentiated alarm descriptions are merged into a unified intent node.

[0151] Step S62: Using the normalized business objects as the core anchor points, and employing graph neural networks or heuristic rules, under the constraints of a pre-defined whitelist, establish structural, business, semantic, rule, and causal relationships between knowledge nodes, constructing a mesh-like connection between knowledge nodes within a topic cluster. The aforementioned whitelist predefines a set of allowed relationship types to ensure the interpretability and governability of the knowledge network.

[0152] Relationship networks are constructed based on graph neural networks or heuristic rules, and multidimensional connection networks are established only within the scope of relationship types limited by a whitelist.

[0153] In embodiments of the present invention, the purpose of graph neural networks or heuristic rules is to form a connection network between knowledge nodes based on structural relationships, business relationships, semantic relationships, rule relationships, and causal relationships. That is, to form a connection between one knowledge node and another or more knowledge nodes that have structural relationships, business relationships, semantic relationships, rule relationships, and causal relationships, which is to form a one-hop relationship between nodes.

[0154] The following example uses a graph neural network for illustration.

[0155] First, to prevent the generation of illegal redundant edges during the knowledge connection construction process and to avoid unordered expansion of the graph topology, a whitelist set is predefined in the construction engine. The whitelist explicitly limits the range of legal relation types that are allowed to be established, including the following five categories:

[0156] Structural relationships: used to express the physical hierarchy of a document or the hierarchy of objects, including "contains" and "belongs to".

[0157] Business relationships: used to express the process associations and collaborative dependencies between business entities, including "relates_to_equipment", "operated_by", and "depends_on".

[0158] Semantic relations: used to express equivalence or hierarchical abstraction at the conceptual level, including "synonym_of", "hypernym_of", and "hyponym_of";

[0159] Rule relationships: used to express the preconditions and parameter threshold constraints for action execution, including "preconditions (precondition_of)" and "parameter threshold constraints (constrained_by)";

[0160] Causal relationship: used to express state evolution and event logic triggering, including "causes", "triggers_alert", and "resolved_by".

[0161] In the subsequent generation of multidimensional knowledge connections, the graph neural network only performs topological edge connections within the relation types defined by the aforementioned preset whitelist. Relationship types outside the whitelist are not allowed to establish connections between nodes, thereby ensuring the interpretability and governance compliance of the knowledge network topology.

[0162] Then, using explicit rules, high-confidence deterministic relation edges are directly extracted and solidified from the text structure and business logic to initialize the knowledge topology skeleton:

[0163] Structural relationship rule construction: Based on the hierarchical topology tags of paragraph-level content nodes obtained in step S2, if node A (such as a chapter title) directly contains node B (such as a clause or table under the chapter) in the topology tree, then establish a directed edge with the relationship type "containment" between node A and node B.

[0164] Semantic relationship rule construction: Based on the synonym merging mapping table constructed in step S61, if node A (e.g., "main transformer") and node B (e.g., "transformer") are mapped to the same standard main business object, then a bidirectional "synonymous" relationship edge is established between node A and node B.

[0165] Rule construction for rules and temporal relationships: For process-type or knowledge-point-type nodes, if node A contains prerequisite markers such as "power must be disconnected first", and node B is "maintenance and disassembly operation", then establish a "precondition" relationship edge between node A and node B; for two consecutive action nodes in the same process step chain, establish a "temporal progression" edge according to the step sequence.

[0166] Furthermore, for cross-document, long-link, or deeply related node pairs that cannot be directly matched through explicit rules, the system uses a graph neural network as an example model to mine implicit connections and predict relationships, specifically including the following processes:

[0167] Let the set of knowledge nodes contained within the topic cluster to be connected be {v}. 1, v2, …, v N For any knowledge node v iLet i = 1, 2, 3, ..., N. Its initial feature vector consists of two parts: first, the text, anchor name, and attribute content of the knowledge node are vectorized and encoded using the BERT text pre-training model to extract the text semantic representation vector; second, the knowledge form tag (such as FAQ, process, knowledge point, case, term) to which the node belongs is categorically embedded, performing an embedding operation. The text semantic vector and the categorical embedding vector are concatenated to obtain the knowledge node v. i The initial feature vector;

[0168] The prior topological edges constructed by heuristic rules are used as the input graph structure, and a graph neural network is used for multi-layer neighborhood feature message passing. The graph neural network can be one of the general graph neural network structures such as graph convolutional network (GCN), graph attention network (GAT), or relational graph convolutional network (R-GCN). Under the graph attention mechanism, each node aggregates the feature representations of its neighboring nodes and assigns different attention transformation weights according to different relation types, so that the vector representation of the node can dynamically incorporate its local business context and temporal correlation information, and output the updated high-order topological node representation.

[0169] Finally, for any pair of candidate nodes that are not directly connected (v i ,v j The updated high-order topological representations are then input into the relation classifier. Taking a fully connected bilinear hierarchical model as an example, the prediction of node v is performed. i With v j The probability score of the relationship between nodes v under each relation type r in the whitelist. When the predicted probability value is greater than or equal to the preset connection confidence threshold, such as 0.75, and the relation type r belongs to the aforementioned whitelist set, the system determines that node v... i With v j A multidimensional relationship edge of type r is established between them; if the probability value is lower than the threshold or the predicted relationship is not in the whitelist, no connection is established.

[0170] After completing the rule-based edge construction and neural network inference edge construction, each fine-grained knowledge node, the normalized business object anchor point, and the structure, business, semantics, rules, and causal connections (edges between nodes) between them are uniformly written into the graph database, thereby completing the construction of a business object-driven multidimensional network knowledge map.

[0171] This leads to the construction of a complete knowledge map driven by business objects. By introducing four unified knowledge anchors—entities, events, attributes, and scenarios—the content in paragraphs, tables, and images is uniformly mapped to the intermediate semantic layer. Furthermore, through thematic aggregation of heterogeneous knowledge, heterogeneous documents, chapters, paragraphs, tables, images, videos, attachments, FAQs, knowledge points, processes, cases, terms, and relationships from different sources are uniformly incorporated into the knowledge nodes of the same knowledge map. Driven by business objects, a network of relationships is formed, enabling any knowledge node to simultaneously possess business coordinates, morphological coordinates, semantic coordinates, and governance coordinates. This completes the construction of the knowledge map, thereby improving the efficiency of enterprise knowledge retrieval, question-and-answer accuracy, traceability, and operability.

[0172] Based on the above embodiments of the present invention, the knowledge map construction method driven by business objects explicitly constructs four types of high-dimensional knowledge anchors—entities, events, attributes, and scenarios—between the original text and deep knowledge. Combined with a two-stage spatial convergence algorithm of rule-based coarse recall and semantic fine ranking, it achieves the alignment of heterogeneous text features, eliminates data noise, and improves the topic aggregation effect of cross-document heterogeneous knowledge fragments. Simultaneously, by reconstructing discrete knowledge nodes into a knowledge network with a temporal graph topology, it utilizes graph neural networks to perform directional gradient propagation and multi-hop convergence along causal and temporal paths during message passing, eliminating semantic breaks and fragmented output problems in multi-step, long-link, and complex graph reasoning.

[0173] Example 2

[0174] like Figure 5 The process of the knowledge map construction method based on business object-driven approach of another embodiment of the present invention shown is based on steps S1-S6 of the aforementioned embodiment, and further includes the configuration of governance attributes and service publication.

[0175] Specifically, in combination Figure 6 As shown, in step S7, governance metadata is configured for the relationship edges of knowledge nodes and multidimensional knowledge connections, and a differentiated knowledge service view is generated based on the multidimensional matrix of user roles, affiliated organizations, associated business objects, current scenarios and knowledge types. This completes the configuration of governance attributes and the publication of knowledge map services, such as the decentralization and on-demand publication of maintenance views, customer service views and management views.

[0176] like Figure 5 The process of the knowledge map construction method based on business object-driven approach of another embodiment of the present invention, which is based on steps S1-S6 of the foregoing embodiment, further includes knowledge map service update operation based on operational dynamic feedback.

[0177] Specifically, in step S8, operational signals during the operation of the knowledge map are captured, such as unmatched query records, incorrect answer records, high-frequency clicks, knowledge reuse rate, low evaluation feedback, and version aging risk. These operational signals are then fed back into the knowledge map's construction engine to trigger updates to the business object system, knowledge anchor extraction rules, topic aggregation strategies, concept thesaurus, or extraction templates. Examples of updates include updating the business object system, correcting knowledge anchor extraction rules, adjusting topic aggregation strategies, expanding the concept thesaurus, optimizing extraction templates, and triggering manual review.

[0178] Example 3

[0179] In this embodiment, taking equipment maintenance knowledge in the power industry as an example, an exemplary implementation process of the technical solution of the foregoing embodiment of the present invention is described in detail.

[0180] S1: Multi-source knowledge access and standardized preprocessing.

[0181] The data sources accessed include: equipment ledger database (customer list), maintenance work order system (work order records), power maintenance procedures (system documents), equipment operation manual (operation manual), common fault FAQ, maintenance training materials, historical fault case documents, equipment parameter tables, on-site photos, scanned drawings, etc.

[0182] Multi-source heterogeneous data is accessed through data interfaces, using REST APIs, database connections, and file upload interfaces to unify the integration of multi-source heterogeneous data into the knowledge processing workflow. During the access process, metadata information from each data source is automatically extracted and recorded, including: source system identifier (e.g., "maintenance work order system"), update timestamp (e.g., "2024-05-10 14:30:00"), scope of application (e.g., "110kV transformer"), usage restrictions (e.g., "internal maintenance personnel only"), and permission boundaries (e.g., "maintenance teams can read, administrators can write").

[0183] The collected data undergoes deduplication (based on content hashing and similarity calculation), version identification (based on timestamps and version numbers), and sensitive information identification (based on regular expressions and named entity recognition, identifying ID card numbers, mobile phone numbers, etc.). The resulting unstructured or semi-structured text segments are output in a standardized format as input for subsequent knowledge extraction and connection construction. In this embodiment, the aforementioned data preprocessing is performed using existing algorithms embedded within the data.

[0184] S2: Structured parsing of multimodal documents.

[0185] In this example, a multimodal large model (using GPT-4V as an example) is used to perform structured parsing of standardized text segments. The model identifies information such as titles, paragraphs, tables, and images in the text. For tables, it identifies the table title, header, and cell content; for images, it identifies the image title and description. The final result is a traceable paragraph-level content node with a hierarchical topology.

[0186] For example, after parsing a "110kV Transformer Maintenance Procedure", the following content nodes are generated:

[0187] Node 1: Title "Chapter 1 General Provisions", Level 1, Source Document ID=DOC001, Page Number=1;

[0188] Node 2: Paragraph “Transformer maintenance cycle is once every three years”, Level 2, Source Document ID=DOC001, Page Number=1, Paragraph Number=2;

[0189] Node 3: Table "Table 1 Maintenance Items and Standards", Level 2, Source Document ID=DOC001, Page Number=2, Table Number=1;

[0190] Node 4: Image " Figure 1 "Transformer structural schematic diagram", Level 2, Source document ID=DOC001, Page number=3, Image number=1, Image description="Internal structure of transformer".

[0191] Each content node is associated with its original document identifier, chapter identifier, and paragraph position, forming a traceable hierarchical topology.

[0192] S3: Construct knowledge anchors driven by business objects.

[0193] In this embodiment, targeting a power maintenance scenario, the pre-defined business object system includes equipment, lines, substations, maintenance teams, and customers. Four types of knowledge anchors are extracted from paragraph-level content nodes:

[0194] Entity anchors: Extract equipment name (e.g., "110kV transformer"), product model (e.g., "SZ11-50000 / 110"), organization (e.g., "Maintenance Team 1"), customer (e.g., "XX Power Supply Company"), and process object (e.g., "Power Outage Maintenance Process") from the content.

[0195] Event anchors: Identify business actions such as "maintenance", "fault handling", "defect elimination", and "preventive testing";

[0196] Attribute anchor points: Identify maintenance cycles (e.g., "every three years"), risk points (e.g., "high voltage hazard"), operating procedures (e.g., "disconnect the circuit breaker"), and applicable conditions (e.g., "ambient temperature below 40°C").

[0197] Scene anchor points: Identify power outage maintenance, routine inspections, high-load operation in summer, thunderstorms, etc.

[0198] Simultaneously, the source evidence chain corresponding to each knowledge anchor is recorded, including the source document ID, chapter number, paragraph number, table number, or image number. For example, the source evidence chain for the anchor "110kV transformer" is: document DOC001, Chapter 1, Paragraph 2.

[0199] S4: Multi-source knowledge topic aggregation.

[0200] Using knowledge anchors as the core, content nodes are aggregated by theme.

[0201] S41: For any candidate paragraph p i The rule query key q is constructed based on its knowledge anchor set and tag set.

[0202] For example, if the rule query key q is "110kV transformer + maintenance + power outage maintenance", the corresponding anchor point set E q Represented as: E q ={110kV transformer};V q ={Inspection}, S q ={Power outage for maintenance}, U q ={}, that is U q It is an empty set. Candidate paragraph p i The set of anchor points is E i Represented as: E i ={110kV transformer}, V i ={Inspection},S i ={Power outage for maintenance}, U i ={maintenance cycle}.

[0203] The coarse recall relevance R(q,p) was calculated based on the weight settings of α=0.7 and β=0.3. i ).

[0204] Among them, configuration w e =0.4,w v =0.3,w s =0.2,w u =0.1, calculate the anchor point matching score.

[0205] The minimum constant ε = 0.01 is used to prevent the denominator from being zero.

[0206] Then, based on a preset first threshold, it is determined whether the candidate paragraph should be included in the candidate set.

[0207] S42: Utilize the BERT text pre-trained model to vectorize the candidate set and calculate semantic similarity:

[0208] Sim(v q ,v i )= v q ⋅v i ∥v q ∥∥v i ∥);

[0209] The preset semantic filtering threshold τ=0.85 retains candidate paragraphs with a similarity greater than or equal to 0.85.

[0210] S43: For candidate clusters recalled with different rule keys, calculate the membership overlap coefficient J(C). a C b ), and calculate the cluster center vector μ of each cluster. a μ b and inter-cluster semantic similarity Sim(C a C b ).

[0211] Configure λ=0.5, η=0.8, and calculate Merge(C). a C b The value of ) and in Merge(C a C b When )=1, merge the two candidate clusters C. a With C b To obtain the topic cluster C k .

[0212] The final set of topic clusters is C = {C1, C2, ..., C}. K}

[0213] S44: Extract anchor features, structural features, and semantic features of topic clusters, and use the BERT classifier as a probabilistic classification model to predict knowledge types. The preset knowledge type set T = {FAQ, knowledge points, processes, cases, terms}.

[0214] For example, topic cluster C1 contains multiple paragraphs about "transformer maintenance procedures," with structural features showing step-by-step descriptions and anchor features primarily consisting of events and attributes. The classification model outputs probabilities: process 0.85, knowledge point 0.10, and case 0.05. Therefore, the target knowledge type is "process."

[0215] S5: Extracting fine-grained knowledge nodes and quality verification

[0216] Based on the type of target knowledge, it is distributed to the corresponding extraction process. For example, for the aforementioned topic cluster C1 (process), the operation step extraction process is executed to extract the preconditions (such as "power outage"), execution steps (such as "disconnect circuit breaker → test for voltage → grounding → maintenance"), responsible roles (such as "maintenance team"), and output results (such as "maintenance report").

[0217] Furthermore, the extracted knowledge nodes undergo quality verification.

[0218] The source reliability score can be calculated based on the metadata information of the data source to which the knowledge node belongs. Specifically, it can be determined based on the authority and timeliness of the data source from which the knowledge node originates. Authority levels are pre-assigned to different source systems; for example, enterprise standard system libraries are assigned 1.0, officially released operation manuals or maintenance procedures 0.9, training materials 0.6, historical work order records 0.5, and customer service dialogue records 0.4. Simultaneously, the update time of the source documents is considered, and source documents exceeding a preset time (e.g., one year) are proportionally attenuated. The source reliability score is obtained by weighted averaging of the authority level score and the timeliness score. If a knowledge node has only one source, the authority level score of that source is used directly.

[0219] The confidence score is derived from the model's output probability, specifically determined by the probability value output by the knowledge node model or the degree of rule matching. For knowledge nodes extracted based on a probability model, the probability value output by the model is directly used as the extraction confidence score; for knowledge nodes extracted based on rules, the ratio of the number of rule hits to the total number of rules is used as the extraction confidence score. If a knowledge node is composed of multiple content fragments, the average confidence score of each fragment can be used.

[0220] The consistency check score can be determined based on the degree of consistency between a knowledge node and similar knowledge nodes. The similarity between the semantic vector of the knowledge node and the center vector of similar knowledge nodes can be calculated as the consistency check score; if there are no similar knowledge nodes in the knowledge base, the default consistency check score is 1. For knowledge nodes containing numerical fields, the deviation of their values ​​from the average of similar nodes can also be checked; the smaller the deviation, the higher the score.

[0221] When any score falls below the corresponding preset threshold, the corresponding fine-grained knowledge node is marked as a risk node, triggering a manual review process.

[0222] S6: Conceptual semantic unification and construction of multidimensional knowledge connections.

[0223] Based on the enterprise standard thesaurus and synonym disambiguation model, it identifies terms, abbreviations, and aliases. For example, it merges "transformer," "main transformer," and "TM" into a unified main business object "transformer." It constructs a hierarchical concept tree, such as "power equipment → transformer → 110kV transformer."

[0224] Using normalized business objects as the core anchor points, multidimensional connections between knowledge nodes are established using graph neural networks or heuristic rules.

[0225] Connection relationship types include structural relationships (such as "a chapter contains a paragraph"), business relationships (such as "equipment belongs to a line"), semantic relationships (such as "synonyms"), rule relationships (such as "maintenance cycle rules"), and causal relationships (such as "faults cause power outages").

[0226] All connections are constrained by a whitelist, which predefines allowed relationship types. The graph neural network only establishes connections within the whitelist.

[0227] S7: Knowledge Governance Configuration and Deployment.

[0228] In this step, governance metadata is configured for each knowledge node and the relationship edges of the established mesh connection, including source tracking, responsible person tags, audit status, effective / expired time, security permission scope, version number, and lifecycle status.

[0229] Furthermore, based on a multi-dimensional matrix of user roles, affiliated organizations, associated business objects, current scenarios, and knowledge types, differentiated knowledge service views are generated. For example, maintenance personnel see a "Maintenance View," customer service personnel see a "Customer Service View," and managers see a "Management View."

[0230] S8: Knowledge map service updated.

[0231] By capturing operational signals in real time, such as unmatched queries, incorrect answers, high-frequency clicks, reuse rates, low rating feedback, and version aging risks, these signals are fed back into the build engine to automatically trigger object system updates, anchor rule corrections, topic aggregation strategy adjustments, concept thesaurus expansion, extraction template optimization, and manual review activation.

[0232] For example, if a large number of queries for "abnormal transformer oil temperature" are found to have no matching knowledge, a new scenario anchor point "abnormal oil temperature" will be added, and the topic aggregation strategy will be adjusted to re-aggregate the relevant paragraphs.

[0233] Example 4

[0234] Combination Figure 7 As shown, in conjunction with the business object-driven knowledge map construction method of the above embodiments, the present invention also proposes a business object-driven knowledge map construction system, including:

[0235] The knowledge access module is used to access multi-source heterogeneous knowledge data through a data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries;

[0236] The document parsing module is used to perform structured parsing on the standardized text segments, identify titles, paragraphs, tables and images in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position.

[0237] Anchor point construction module is used to extract knowledge anchor points from the paragraph-level content nodes based on a preset business object system, serving as a thematic intermediate layer connecting original documents and deep knowledge. The knowledge anchor points include entity anchor points, event anchor points, attribute anchor points, and scene anchor points, and synchronously record the source evidence chain corresponding to each knowledge anchor point.

[0238] The topic aggregation module is used to perform topic aggregation on the paragraph-level content nodes based on the extracted knowledge anchors through multi-level recall and clustering, to aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and to determine the target knowledge type corresponding to each topic cluster.

[0239] The knowledge extraction module is used to distribute each topic cluster to the corresponding knowledge extraction process according to the target knowledge type, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes.

[0240] The multidimensional knowledge connection construction module is used to normalize the conceptual semantics of knowledge nodes into unified business objects based on the enterprise standard lexicon and the synonym disambiguation model for each topic cluster. Then, using the normalized business objects as the core anchor points, it establishes multidimensional knowledge connections between knowledge nodes within the topic cluster using graph neural networks or heuristic rules, thus completing the construction of the knowledge map.

[0241] The functions and specific implementation principles and methods of the aforementioned knowledge access module, document parsing module, anchor point construction module, topic aggregation module, knowledge extraction module, and multidimensional knowledge connection construction module can be implemented according to the corresponding process of the knowledge map construction method in the aforementioned embodiments.

[0242] Example 5

[0243] Based on the implementation of the business object-driven knowledge map construction method in the above embodiments, a computer system including a processor and a memory is also proposed according to the present invention.

[0244] The memory stores a computer program, and when the processor executes the computer program, it implements the business object-driven knowledge map construction method of the aforementioned embodiment.

[0245] Example 6

[0246] Based on the implementation of the knowledge map construction method driven by business objects in the above embodiments, the present invention also proposes a computer-readable storage medium on which a computer program is stored.

[0247] When the computer program is executed by the processor, it implements the process of constructing a knowledge map based on business object-driven methods as described in the aforementioned embodiments.

[0248] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A knowledge map construction method based on business object-driven approach, characterized in that, Includes the following steps: Step S1: Access multi-source heterogeneous knowledge data through the data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries; Step S2: Perform structured parsing on the standardized text segment, identify the title, paragraph, table and image in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position. Step S3: Based on the preset business object system, extract knowledge anchors from the paragraph-level content nodes as the subject intermediate layer connecting the original documents and deep knowledge. The knowledge anchors include entity anchors, event anchors, attribute anchors and scene anchors, and record the source evidence chain corresponding to each knowledge anchor simultaneously. Step S4: Based on the extracted knowledge anchors, perform topic aggregation on the paragraph-level content nodes through multi-level recall and clustering, aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and determine the target knowledge type corresponding to each topic cluster; Step S5: Based on the target knowledge type, distribute each topic cluster to the corresponding knowledge extraction process, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes; Step S6: For each topic cluster, based on the enterprise standard lexicon and the synonym disambiguation model, the knowledge nodes are conceptually and semantically normalized to a unified business object. Using the normalized business object as the core anchor point, a multi-dimensional knowledge connection between the knowledge nodes within the topic cluster is established using graph neural networks or heuristic rules to complete the construction of the knowledge map.

2. The knowledge map construction method based on business object-driven approach according to claim 1, characterized in that, In step S3, knowledge anchors are extracted from the paragraph-level content nodes, specifically including the following steps: Step S31: Extract entity elements from the parsed text content of titles, paragraphs, tables, and images to form entity anchors; Step S32: Identify the business action elements contained in the text content and form event anchors to describe business dynamics; Step S33: Identify the feature descriptions contained in the text content to form attribute anchors that carry the attribute features of objects and events; Step S34: Identify the applicable conditions and scene feature descriptions contained in the text content to form scene anchors used to express the boundaries of knowledge application and role context.

3. The knowledge map construction method based on business object-driven approach according to claim 2, characterized in that, In step S3, during the extraction of knowledge anchors, the positions of the source documents, chapters, paragraphs, tables, or images corresponding to the extracted knowledge anchors are recorded simultaneously to construct the source evidence chain corresponding to the knowledge anchors.

4. The knowledge map construction method based on business object-driven approach according to claim 1, characterized in that, In step S4, based on the extracted knowledge anchors, the paragraph-level content nodes are aggregated into topics through multi-level recall and clustering. Content fragments from multi-source heterogeneous knowledge data are aggregated into the same topic cluster, and the target knowledge type corresponding to each topic cluster is determined. Specifically, this includes the following steps: Step S41: Construct rule query keys based on the knowledge anchors and tag information, perform coarse recall on each paragraph-level content node, and obtain a candidate set; Step S42: Perform semantic ranking on the candidate set using a pre-trained text model; Step S43: Merge the candidate clusters in the ranked candidate set according to the overlap and semantic distance between the candidate clusters to obtain the topic cluster set; Step S44: Identify the knowledge type tendency for each topic cluster within the topic cluster set, and determine the target knowledge type corresponding to each topic cluster.

5. The knowledge map construction method based on business object-driven approach according to claim 4, characterized in that, In step S41, a rule query key is constructed based on the knowledge anchors and tag information to perform coarse recall on each paragraph-level content node, resulting in a candidate set. This specifically includes the following steps: For any candidate paragraph Based on its knowledge anchor set and tag set, a rule query key q is constructed, and the candidate paragraph p is calculated. i Relative to the rule query key coarse recall relevance R(q,p) i ): R(q,p i )=α·I anchor (q,p i )+β·I tag (q,p i ); Among them, I anchor (q,p i ) represents the anchor matching score, I tag (q,p i The score is the tag matching score, where α and β are the anchor matching weight and the tag matching weight, respectively, and α+β=1. Based on the preset first threshold, the coarse recall relevance R(q,p) is... i Candidate paragraphs that meet the preset first threshold condition are added to the candidate set.

6. The knowledge map construction method based on business object-driven approach according to claim 5, characterized in that, The anchor point matching score I anchor (q,p i It is set to be calculated in the following manner: Let A be the set of anchor points corresponding to the rule query key q. q for: A q ={E q ,V q ,S q ,U q }; Among them, E q V q ,S q U q These are the entity set, event set, scenario set, and attribute set corresponding to the rule query key q, respectively. Let the candidate set C be... q Each candidate paragraph p in i The corresponding anchor point set A i for: A i ={E i ,V i ,S i ,U i }; Among them, E i V i ,S i U i The candidate paragraphs p are respectively i The corresponding entity set, event set, scene set, and attribute set; Define the set overlap function J(X,Y) as follows: J(X,Y)=|X∩Y| / (|X∩Y|+ε); In the formula, ε is a local constant used to prevent the denominator from being 0; Then calculate J(E) according to the set overlap function J(X,Y). q E i ), J(V q V i ), J(S q ,S i ), J(U q U i ), and the anchor point matching score I anchor (q,p i ) is defined as: I anchor (q,p i )=w e ·J(E q ,E i )+w v J(V q ,V i )·+w s J(S q ,S i )·+w u ·J(U q ,U i ); Among them, w e ,w v ,w s ,w u Let w represent the weights of the entity, event, scenario, and attribute dimensions, respectively, and satisfy: w e +w v +w s +w u =1.

7. The knowledge map construction method based on business object-driven approach according to claim 5, characterized in that, The tag matching score I tag (q,p i It is set to be calculated in the following manner: Let T be the set of tags corresponding to the rule query key q. q Candidate paragraph p i The corresponding tag set is T i Then the label matching score I tag (q,p i ) is defined as: ； In the formula, the label similarity Sim tag (t a ,t b The calculation rules for ) are as follows: ； Among them, t a Representing the label set T respectively q The a-th label in the set T has a value of 1. i The bth tag; , The labels are t respectively a ,t b The semantic vector, where ρ is the scaling factor.

8. The knowledge map construction method based on business object-driven approach according to claim 5, characterized in that, In step S42, the candidate set is semantically ranked using a pre-trained text model, including: The candidate set obtained from coarse recall is vectorized using a pre-trained text model, and noisy data that deviates from the topic is removed by calculating semantic similarity, where the semantic similarity Sim(v) is... q ,v i The calculation method is as follows: ； Among them, v q v is the semantic vector of the rule query key q. i For candidate paragraph p i semantic vector; For candidate set C q Each candidate paragraph p in i If the following expression is satisfied, retain it: Sim(v q ,v i )≥τ; Obtain the filtered candidate set C q ’ : C q ’ ={p i ∈C q |Sim(in q ,in i )≥τ}; In the formula, τ is the semantic filtering threshold.

9. The knowledge map construction method based on business object-driven approach according to claim 8, characterized in that, In step S43, the candidate clusters in the refined candidate set are merged based on the overlap and semantic distance between the candidate clusters to obtain the topic cluster set. This specifically includes the following steps: For set C q ’ Any two candidate clusters C a With C b Calculate its member overlap coefficient J(C) a C b ), and calculate the center vector μ of each candidate cluster respectively. a μ b Calculate the semantic similarity Sim(C) between two candidate clusters. a C b ),in: ； ； ； ； Then, based on the merge decision function Merge(C) a C b )=f(J(C a C b Sim(C) a C b The value of the merging determination function is such that when the value satisfies the preset second threshold condition, the candidate cluster C is... a With C b Merging them yields a new topic cluster C. k ; Record all the merged topic clusters as a topic cluster set C: C={C1,C2,…,C K } Where K represents the total number of all topic clusters in the topic cluster set.

10. The knowledge map construction method based on business object-driven approach according to claim 5, characterized in that, In step S44, knowledge type preference identification is performed on each topic cluster within the topic cluster set to determine the target knowledge type corresponding to each topic cluster. This specifically includes the following steps: Extract anchor features, structural features, and semantic features of each topic cluster; The probability distribution of the topic cluster on the knowledge type set is predicted by a probability classification model, and the knowledge type corresponding to the maximum probability is determined as the target knowledge type of the topic cluster; the preset knowledge type set includes FAQ, knowledge points, processes, cases and terms.

11. The knowledge map construction method based on business object-driven approach according to claim 10, characterized in that, In step S5, based on the target knowledge type, each topic cluster is distributed to the corresponding knowledge extraction process, and knowledge nodes are extracted from each topic cluster. This specifically includes the following steps: When the target knowledge type is FAQ, execute the question and answer extraction process to extract standard questions, standard answers, and similar words and phrases; When the target knowledge type is a knowledge point, the rule-based fact extraction process is executed to extract the core definition, triggering reason, judgment criteria and handling suggestions. When the target knowledge type is a process, the operation steps are executed to extract the process, including the preconditions, execution steps, responsible roles, and output results. When the target knowledge type is a case study, the event review and extraction process is executed to extract the business background, core issues, technical processes, and experience summaries. When the target knowledge type is a term, execute the term definition extraction process to extract standard terms, aliases, abbreviations, and hypernyms and hyponyms. For each type of target knowledge, the extracted results are used as knowledge nodes of the topic cluster.

12. The knowledge map construction method based on business object-driven approach according to claim 10, characterized in that, In step S6, for each topic cluster, based on the enterprise standard lexicon and the synonym disambiguation model, the knowledge nodes are semantically normalized to a unified business object. Using the normalized business object as the core anchor point, a multi-dimensional knowledge connection between the knowledge nodes within the topic cluster is established using graph neural networks or heuristic rules. Specifically, this includes the following process: Step S61: Based on the enterprise standard thesaurus and the synonym disambiguation model, identify the terms, abbreviations, aliases and industry terms in the knowledge nodes; Construct a synonym merging mapping table and a hierarchical concept tree; Based on the synonym merging mapping table and hierarchical concept tree, knowledge nodes are merged into a unified business object, and differentiated alarm descriptions are merged into a unified intent node. Step S62: Using the normalized business object as the core anchor point, and using graph neural networks or heuristic rules, under the constraint control of a preset whitelist, establish structural relationships, business relationships, semantic relationships, rule relationships, and causal relationships between knowledge nodes. The whitelist predefines a set of relationship types that are allowed to be established, and the graph neural network or heuristic rules only establish connections within the relationship types limited by the whitelist.

13. The knowledge map construction method based on business object-driven approach according to any one of claims 1-12, characterized in that, The method further includes the following steps: Step S7: Configure governance metadata for the relationship edges of knowledge nodes and multidimensional knowledge connections, and generate differentiated knowledge service views based on the multidimensional matrix of user roles, affiliated organizations, associated business objects, current scenarios and knowledge types, and complete the governance attribute configuration and knowledge map service publication.

14. The knowledge map construction method based on business object-driven approach according to any one of claims 1-12, characterized in that, The method further includes the following steps: Step S8: Capture operational signals during the operation of the knowledge map, including at least one of the following: unmatched query records, incorrect answer records, high-frequency clicks, knowledge reuse rate, low evaluation feedback, and version aging risk. Input the operational signals back into the knowledge map building engine to trigger updates to the business object system, knowledge anchor extraction rules, topic aggregation strategy, concept thesaurus, or extraction template. The updates include at least one of the following: updating the business object system, revising the knowledge anchor extraction rules, adjusting the topic aggregation strategy, expanding the concept lexicon, optimizing the extraction template, and activating manual review.

15. A knowledge map construction system based on business object-driven architecture, characterized in that, include: The knowledge access module is used to access multi-source heterogeneous knowledge data through a data interface, extract metadata information from each data source, and output standardized text segments; the metadata information includes source system identifier, update timestamp, scope of application, usage restrictions, and permission boundaries; The document parsing module is used to perform structured parsing on the standardized text segments, identify titles, paragraphs, tables and images in the text, and generate paragraph-level content nodes with hierarchical topology and source location information. Each content node is associated with its corresponding original document identifier, chapter identifier and paragraph position. Anchor point construction module is used to extract knowledge anchor points from the paragraph-level content nodes based on a preset business object system, serving as a thematic intermediate layer connecting original documents and deep knowledge. The knowledge anchor points include entity anchor points, event anchor points, attribute anchor points, and scene anchor points, and synchronously record the source evidence chain corresponding to each knowledge anchor point. The topic aggregation module is used to perform topic aggregation on the paragraph-level content nodes based on the extracted knowledge anchors through multi-level recall and clustering, to aggregate content fragments from multi-source heterogeneous knowledge data into the same topic cluster, and to determine the target knowledge type corresponding to each topic cluster. The knowledge extraction module is used to distribute each topic cluster to the corresponding knowledge extraction process according to the target knowledge type, extract knowledge nodes from each topic cluster, and perform quality verification on the knowledge nodes. The multidimensional knowledge connection construction module is used to normalize the conceptual semantics of knowledge nodes into unified business objects based on the enterprise standard lexicon and the synonym disambiguation model for each topic cluster. Then, using the normalized business objects as the core anchor points, it establishes multidimensional knowledge connections between knowledge nodes within the topic cluster using graph neural networks or heuristic rules, thus completing the construction of the knowledge map.

16. A computer system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method of any one of claims 1 to 14.

17. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 14.

Images