An enterprise AI-powered data document federated knowledge base construction and management system

Abstract

This invention discloses an enterprise AI data document joint knowledge base construction and management system, relating to the fields of enterprise knowledge management and AI data processing technology. It addresses specific technical problems in existing technologies, such as the inability to accurately identify conflicts between multiple versions of documents on the same topic, insufficient mining of semantic relationships between documents, and asynchronous updates of related documents. This invention comprises five modules. The improved semantic relationship clustering module introduces version time sequence factors and business relationship weights, designs an improved semantic relationship clustering algorithm, and integrates semantic similarity, version time sequence, and business relationship to achieve accurate clustering and mine deep semantic relationships between documents. The remaining modules work in conjunction with the improved semantic relationship clustering module to complete document collection, conflict arbitration, knowledge storage, and dynamic updates, forming a closed-loop management system.

Description

Technical Field

[0001] This invention relates to the field of enterprise knowledge management and AI data processing, specifically to an enterprise AI data document joint knowledge base construction and management system. Background Technology

[0002] Currently, enterprises generally face a specific pain point when building AI-powered data document joint knowledge bases: similar business documents within the enterprise (such as compliance systems, product specifications, operation manuals, etc.) will generate multiple versions as the business iterates. Documents uploaded by different departments may have content conflicts, and the relationships between documents are not effectively identified, resulting in a large number of "zombie documents" and "conflicting documents" in the knowledge base. AI is prone to bias when retrieving and calling knowledge, and may even mislead business decisions.

[0003] In existing technologies, enterprise knowledge base construction often adopts a traditional document classification and storage + keyword retrieval model. The modules used (such as document parsing, keyword extraction, and storage modules) are all common industry designs, and no specific solutions have been designed for the specific problem of "multi-version document conflicts and untimely updates of related documents." Some improved solutions attempt to use simple similarity comparison for document deduplication, but they do not consider the semantic relationships, version sequence, and business relevance of documents. They cannot accurately identify conflict points among multiple versions of documents on the same topic, nor can they achieve synchronous updates of related documents, making it difficult to guarantee the accuracy and timeliness of the knowledge base, thus affecting the reliability of AI applications. In addition, existing clustering algorithms (such as K-means and traditional semantic clustering) are prone to low clustering accuracy and ignoring version relationships when processing multi-version and multi-domain enterprise documents, failing to meet the needs of precise management of enterprise knowledge bases. Summary of the Invention

[0004] The purpose of this invention is to provide an enterprise AI data document joint knowledge base construction and management system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an enterprise AI data document joint knowledge base construction and management system, comprising a document multimodal acquisition module, an improved semantic association clustering module, a document version conflict arbitration module, an AI knowledge fusion storage module, and a knowledge dynamic update module; the output end of the document multimodal acquisition module is connected to the input end of the improved semantic association clustering module, the output end of the improved semantic association clustering module is connected to the input ends of the document version conflict arbitration module and the AI ​​knowledge fusion storage module respectively, the output end of the document version conflict arbitration module is connected to the input end of the AI ​​knowledge fusion storage module, the output end of the AI ​​knowledge fusion storage module is connected to the input end of the knowledge dynamic update module, and the output end of the knowledge dynamic update module feeds back to the improved semantic association clustering module, forming a closed-loop management; The document multimodal acquisition module is used to collect various AI data documents within the enterprise, including text documents (Word, PDF, TXT), table documents (Excel, CSV), and image documents (PNG, JPG containing text). The improved semantic association clustering module receives standardized text data output by the document multimodal acquisition module. Through the improved semantic association clustering algorithm, it clusters the text data according to topic category, semantic association degree, and version time sequence, and outputs the clustering results of multiple versions of documents with the same topic, the document semantic association matrix, and the version time sequence ranking results. This provides accurate clustering basis for the document version conflict arbitration module and semantic association clues for the knowledge dynamic update module. The document version conflict arbitration module receives the clustering results of multiple versions of documents on the same topic output by the improved semantic association clustering module, identifies content conflicts between documents, and arbitrates them. The AI ​​knowledge fusion and storage module is used to extract and fuse knowledge from clustered and arbitrated documents for storage. The knowledge dynamic update module monitors the collection and modification of documents in real time, feeds back the relevant documents to the improved semantic association clustering module for re-clustering, triggers synchronous update prompts for related documents, updates the knowledge index and knowledge graph, and deletes outdated version information.

[0006] Preferably, the document multimodal acquisition module works as follows: it acquires documents through three methods: enterprise internal office system interface, local folder scanning, and employee upload; it parses the acquired documents to extract text information from text, tables, and images, and converts non-text information into a processable text format; it performs preliminary cleaning on the parsed text to remove invalid characters and blank paragraphs, unifies text encoding, and outputs standardized text data to the improved semantic association clustering module.

[0007] Preferably, the improved semantic association clustering module operates as follows: Step 1, Data Input and Preprocessing: Receive standardized text data output from the document multimodal acquisition module, and process each document... Extracting semantic feature vectors using a BERT pre-trained model Simultaneously extract the document's version timestamp. Business Area Tags Establish a basic document information database; Step 2, Parameter Initialization: Statistics of Document Collection Timestamp range, calculation and Based on the enterprise's business architecture, preset the correlation weights of each business area. (Can be manually adjusted by the enterprise administrator); Initialize weight coefficients Set an iteration termination condition, i.e., the number of iterations ≤ 50. The difference is less than 0.001; Step 3, Initial Clustering and Grouping: Document-based Business Domain Tags Documents in the same business domain are grouped into the same initial cluster, resulting in an initial cluster set. Calculate the temporary central semantic vector of each initial cluster (the mean of the semantic feature vectors of all documents within the cluster). Step 4: Iterative optimization of clustering: 4.1 Calculate the overall similarity: For each initial cluster Any two documents within , Calculate according to the formula (Cosine similarity) , And thus obtain ; 4.2 Document Reassignment: Calculate each document The overall similarity with each initial cluster temporary center will Assign to the cluster with the highest overall similarity and update the cluster set; 4.3 Update cluster centers: For each updated cluster... Calculate the semantic vector of the central document within the cluster according to the formula. As the new cluster center; 4.4 Validation of the Objective Function: Calculate the objective function for the current cluster set. If the iteration termination condition is met, then stop the iteration; otherwise, repeat steps 4.1-4.3 until the condition is met. Step 5, Cluster Optimization and Adjustment: Merge and split the iterated cluster set: Cluster Merging: Traverse all clusters. If the number of documents in a cluster is less than 3 and the average comprehensive similarity with a neighboring cluster (with similar business domain) is greater than or equal to 0.7, merge the two clusters and recalculate the center document of the merged cluster; Cluster Splitting: Traverse all clusters. If the average comprehensive similarity of all documents in a cluster is less than 0.5, split the cluster into two sub-clusters. Use the two documents with the highest comprehensive similarity in the cluster as the initial center of the sub-cluster. Repeat the iterative process of Step 4 until the sub-clusters meet the clustering requirements. Step 6, Output Results: After cluster optimization and adjustment, the final set of clusters is obtained. Simultaneously, it generates two core outputs: multi-version document clusters on the same topic, with each cluster containing all versions of documents on the same topic, and labeled with the central document (baseline document) within the cluster and the version chronological order; and a document semantic association matrix, where the matrix elements are any two documents. The value is used to reflect the degree of semantic relationship between documents; the above outputs are sent to the document version conflict arbitration module and the AI ​​knowledge fusion storage module respectively to provide support for subsequent modules; Step 7, Dynamic Feedback Adaptation: Receive new and modified document data sent by the knowledge dynamic update module, repeat steps 1-6, and update the clustering clusters and semantic association matrix in real time to ensure that the clustering results are synchronized with the document updates, thus ensuring the timeliness of subsequent conflict arbitration and knowledge storage.

[0008] Preferably, the document version conflict arbitration module is implemented as follows: extract the core semantic elements and version timestamps of each document in the cluster; compare the core semantic elements of each version of the document to identify conflict points (such as different expressions of the same rule or different values ​​of the same parameter); determine the optimal version as the benchmark document according to the preset arbitration rules (business priority, version sequence, departmental authority), mark the conflicting documents and provide modification suggestions; and output the arbitrated benchmark document and the marked conflicting documents to the AI ​​knowledge fusion storage module.

[0009] Preferably, the specific implementation logic of the AI ​​knowledge fusion storage module is as follows: extract the core knowledge (concepts, rules, parameters, processes) of the benchmark document and construct a knowledge graph; associate and store the relationship between the conflicting document and the benchmark document, as well as the modification suggestions, to form a version traceability chain; adopt a hybrid storage architecture (relational database to store structured knowledge, graph database to store the knowledge graph, and distributed file to store the original document) to achieve hierarchical storage of knowledge; establish a knowledge index and associate it with the semantic tags output by the improved semantic association clustering module to provide support for AI retrieval.

[0010] Preferably, the specific implementation logic of the knowledge dynamic update module is as follows: real-time monitoring of newly acquired documents from the document multimodal acquisition module, as well as modifications to existing documents; sending new and modified documents to the improved semantic association clustering module for re-clustering and semantic association mining; if a modification to the baseline document is detected, triggering a synchronous update prompt for related documents based on the association relationships in the knowledge graph, and notifying relevant departments for review; updating the knowledge index and knowledge graph, deleting outdated version tracking information, and ensuring the timeliness of the knowledge base.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention, through an improved semantic association clustering module combined with an improved ISAC algorithm, accurately identifies multiple versions of documents on the same topic, mines semantic and version associations between documents, effectively solves the problems of inaccurate identification of conflicts between multiple versions of documents on the same topic and untimely updates of associations, avoids conflicting and outdated documents in the knowledge base, and improves the accuracy and relevance of the knowledge base.

[0012] This invention introduces version time-series factors and business association weights into an improved semantic association clustering module. It designs an ISAC comprehensive similarity calculation method and an optimized clustering objective function, overcoming the shortcomings of traditional clustering algorithms that ignore version associations and insufficient semantic mining. The clustering accuracy is improved compared to traditional algorithms, and it can accurately match the clustering needs of enterprise documents with multiple versions and multiple domains. It is significantly different from existing common clustering modules and algorithms and has outstanding creativity.

[0013] The various modules of this invention work collaboratively, with the core module focusing on creativity and the auxiliary modules being simple and efficient. It does not require significant modifications to the existing document management processes of enterprises and can quickly adapt to the knowledge base construction needs of enterprises of different sizes. The knowledge dynamic update module and the core clustering module form a closed loop, realizing the synchronous update of related documents, reducing the maintenance cost of the enterprise knowledge base, ensuring the reliability of AI retrieval and knowledge retrieval, and thus improving the efficiency of enterprise business decision-making.

[0014] The improved ISAC algorithm of this invention can adaptively adjust the weight coefficients and clustering parameters according to the business needs of enterprises, and adapt to the document characteristics of different industries (such as manufacturing, finance and Internet). The auxiliary module can flexibly add new document collection methods and arbitration rules according to the needs of enterprises, and has strong scalability. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the system structure of the present invention; Figure 2 This is a schematic diagram of the workflow of the improved semantic association clustering module of the present invention. Detailed Implementation

[0016] 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.

[0017] Please see Figure 1-2This invention provides a technical solution: an enterprise AI data document joint knowledge base construction and management system, including a document multimodal acquisition module, an improved semantic association clustering module, a document version conflict arbitration module, an AI knowledge fusion storage module, and a knowledge dynamic update module; the output end of the document multimodal acquisition module is connected to the input end of the improved semantic association clustering module, the output end of the improved semantic association clustering module is connected to the input ends of the document version conflict arbitration module and the AI ​​knowledge fusion storage module respectively, the output end of the document version conflict arbitration module is connected to the input end of the AI ​​knowledge fusion storage module, the output end of the AI ​​knowledge fusion storage module is connected to the input end of the knowledge dynamic update module, and the output end of the knowledge dynamic update module feeds back to the improved semantic association clustering module, forming a closed-loop management; The document multimodal acquisition module is used to collect various AI data documents within the enterprise, including text documents (Word, PDF, TXT), table documents (Excel, CSV), and image documents (PNG, JPG containing text). The improved semantic association clustering module receives standardized text data output by the document multimodal acquisition module. Through the improved semantic association clustering algorithm, it clusters the text data according to topic category, semantic association degree, and version time sequence, and outputs clusters of documents with multiple versions of the same topic, document semantic association matrix, and version time sequence ranking results. This provides accurate clustering basis for the document version conflict arbitration module and semantic association clues for the knowledge dynamic update module. The document version conflict arbitration module receives the clustering results of multiple versions of documents on the same topic from the improved semantic association clustering module, identifies content conflicts between documents, and arbitrates them. The AI ​​knowledge fusion and storage module is used to extract and fuse knowledge from clustered and arbitrated documents for storage. The knowledge dynamic update module monitors the collection and modification of documents in real time, and feeds back the relevant documents to the improved semantic association clustering module for re-clustering; it triggers synchronous update prompts for related documents, updates the knowledge index and knowledge graph, and deletes outdated version information.

[0018] Furthermore, the specific working logic of the document multimodal acquisition module is as follows: Documents are collected through three methods: enterprise internal office system interface, local folder scanning, and employee upload; the collected documents are parsed to extract text information from text, tables, and images, and non-text information is converted into a processable text format; the parsed text is preliminarily cleaned to remove invalid characters and blank paragraphs, unify text encoding, and output standardized text data to the improved semantic association clustering module.

[0019] Furthermore, the specific working steps of the improved semantic association clustering module are as follows: Step 1, Data Input and Preprocessing: Receive standardized text data output from the document multimodal acquisition module, and process each document... Extracting semantic feature vectors using a BERT pre-trained model Simultaneously extract the document's version timestamp. Business Area Tags Establish a basic document information database; Step 2, Parameter Initialization: Statistics of Document Collection Timestamp range, calculation and Based on the enterprise's business architecture, preset the correlation weights of each business area. (Can be manually adjusted by the enterprise administrator); Initialize weight coefficients Set an iteration termination condition, i.e., the number of iterations ≤ 50. The difference is less than 0.001; Step 3, Initial Clustering and Grouping: Document-based Business Domain Tags Documents in the same business domain are grouped into the same initial cluster, resulting in an initial cluster set. Calculate the temporary central semantic vector of each initial cluster (the mean of the semantic feature vectors of all documents within the cluster). Step 4: Iterative optimization of clustering: 4.1 Calculate the overall similarity: For each initial cluster Any two documents within , Calculate according to the formula (Cosine similarity) , And thus obtain ; 4.2 Document Reassignment: Calculate each document The overall similarity with each initial cluster temporary center will Assign to the cluster with the highest overall similarity and update the cluster set; 4.3 Update cluster centers: For each updated cluster... Calculate the semantic vector of the central document within the cluster according to the formula. As the new cluster center; 4.4 Validation of the Objective Function: Calculate the objective function for the current cluster set. If the iteration termination condition is met, then stop the iteration; otherwise, repeat steps 4.1-4.3 until the condition is met. Step 5, Cluster Optimization and Adjustment: Merge and split the iterated cluster set: Cluster Merging: Traverse all clusters. If the number of documents in a cluster is less than 3 and the average comprehensive similarity with a neighboring cluster (with similar business domain) is greater than or equal to 0.7, merge the two clusters and recalculate the center document of the merged cluster; Cluster Splitting: Traverse all clusters. If the average comprehensive similarity of all documents in a cluster is less than 0.5, split the cluster into two sub-clusters. Use the two documents with the highest comprehensive similarity in the cluster as the initial center of the sub-cluster. Repeat the iterative process of Step 4 until the sub-clusters meet the clustering requirements. Step 6, Output Results: After cluster optimization and adjustment, the final set of clusters is obtained. Simultaneously, it generates two core outputs: multi-version document clusters on the same topic, with each cluster containing all versions of documents on the same topic, and labeled with the central document (baseline document) within the cluster and the version chronological order; and a document semantic association matrix, where the matrix elements are any two documents. The value is used to reflect the degree of semantic relationship between documents; the above outputs are sent to the document version conflict arbitration module and the AI ​​knowledge fusion storage module respectively to provide support for subsequent modules; Step 7, Dynamic Feedback Adaptation: Receive new and modified document data sent by the knowledge dynamic update module, repeat steps 1-6, and update the clustering clusters and semantic association matrix in real time to ensure that the clustering results are synchronized with the document updates, thus ensuring the timeliness of subsequent conflict arbitration and knowledge storage.

[0020] The improved semantic association clustering module will be explained further below: The improved semantic association clustering module, based on traditional semantic clustering, introduces document version time factors and business association weights to design an improved semantic association clustering algorithm (ISAC). This algorithm not only accurately identifies multiple versions of documents on the same topic but also uncovers deep semantic relationships between documents, providing precise support for subsequent conflict arbitration and dynamic updates. It overcomes the shortcomings of existing clustering algorithms that ignore version relationships and have insufficient semantic mining. Specifically: Improved Algorithm Design (ISAC Algorithm): Existing semantic clustering algorithms only consider the semantic similarity of documents, failing to incorporate the version characteristics and business relevance of enterprise documents. This results in low clustering accuracy and insufficient association mining for multiple versions of documents on the same topic. The ISAC algorithm designed in this invention introduces a version time factor (…). ), business association weight ( This paper improves upon traditional semantic similarity calculation and optimizes the clustering objective function to achieve accurate clustering and semantic association mining of multiple versions of documents on the same topic.

[0021] Algorithm core parameter definition Let the collected standardized text document set be... ,in Total number of documents; Let each document The semantic feature vector is ,in The semantic feature dimension (extracted through the BERT model, avoiding the limitations of traditional TF-IDF features); Set document The version timestamp is (The format is a timestamp value; the larger the value, the newer the version.) The earliest version timestamp for documents on the same topic is... The latest version timestamp is ; Set document The relevant business area tag is (e.g., "compliance", "product", "R&D"), business relevance weighting Document and The degree of relevance in business areas, with a value range of [value range]. The value is 1 for the same business area, 0.6-0.9 for similar areas (such as "R&D" and "testing"), and 0 for unrelated areas. Set document and Traditional semantic similarity is (Calculated using cosine similarity), the improved overall similarity is: ; Let the set of clusters be ,in The number of clusters is adaptively determined, requiring no manual preset, and the semantic vector of the central document within each cluster is... ( ).

[0022] Key formula derivation: Version time series factor calculation: This reflects the temporal correlation of document versions. The closer the versions are, the larger the time series factor. The formula is as follows: ; when This indicates that all documents are of the same version. .

[0023] Comprehensive similarity calculation: This involves integrating semantic similarity, version time factor, and business association weights to form an improved comprehensive similarity score, which serves as the core basis for clustering. The formula is as follows: ,in For the weighting coefficients, satisfying It adaptively adjusts according to the enterprise's knowledge base management needs, with the default value being... (Semantic similarity is the core) (Version sequence) (Business relevance).

[0024] Clustering objective function: The goal is to maximize the sum of the overall similarity of documents within a cluster and minimize the sum of the overall similarity of documents between clusters, optimizing the clustering results. The formula is as follows: .

[0025] Cluster center document determination: Within each cluster, the document with the highest overall similarity and the most recent version (largest timestamp) is selected as the cluster center document, serving as the benchmark for subsequent conflict arbitration. The formula is as follows: .

[0026] Algorithm optimization strategies: To avoid clustering getting trapped in local optima, a strategy of initial clustering followed by iterative optimization is adopted: 1) Initial clustering: Based on business domain tags, documents are initially grouped to obtain initial clusters; 2) Iterative optimization: Calculate the comprehensive similarity between each document and each cluster center, assign the document to the cluster with the highest comprehensive similarity, update the cluster centers, and repeat the iteration until the objective function is achieved. It tends to stabilize (the number of iterations does not exceed 50, or the number of consecutive iterations is close to 50). (The difference is less than 0.001). 3) Cluster merging and splitting: For clusters with too few documents (less than 3) after clustering, if the average comprehensive similarity with other clusters is greater than 0.7, they are merged; for clusters with an average comprehensive similarity of less than 0.5, they are split to ensure clustering accuracy.

[0027] Furthermore, the specific implementation logic of the document version conflict arbitration module is as follows: extract the core semantic elements and version timestamps of each document in the cluster; compare the core semantic elements of each version of the document to identify conflict points (such as different expressions of the same rule or different values ​​of the same parameter); determine the optimal version as the benchmark document according to the preset arbitration rules (business priority, version sequence, departmental authority), mark the conflicting documents and provide modification suggestions; output the arbitrated benchmark document and the marked conflicting documents to the AI ​​knowledge fusion storage module.

[0028] Furthermore, the specific implementation logic of the AI ​​knowledge fusion storage module is as follows: extract the core knowledge (concepts, rules, parameters, processes) from the benchmark document and construct a knowledge graph; associate and store the relationship between conflicting documents and benchmark documents, as well as modification suggestions, to form a version traceability chain; adopt a hybrid storage architecture (relational database to store structured knowledge, graph database to store knowledge graph, and distributed file to store original documents) to achieve hierarchical storage of knowledge; establish a knowledge index and associate it with the semantic tags output by the improved semantic association clustering module to provide support for AI retrieval.

[0029] Furthermore, the specific implementation logic of the knowledge dynamic update module is as follows: Real-time monitoring of newly acquired documents from the document multimodal acquisition module, as well as modifications to existing documents; sending new and modified documents to the improved semantic association clustering module for re-clustering and semantic association mining; if a modification to the baseline document is detected, triggering a synchronous update prompt for related documents based on the association relationships in the knowledge graph, and notifying relevant departments for review; updating the knowledge index and knowledge graph, deleting outdated version tracking information, and ensuring the timeliness of the knowledge base.

[0030] This invention discloses an enterprise AI data document joint knowledge base construction and management system, relating to the fields of enterprise knowledge management and AI data processing technology. It addresses specific technical problems in existing technologies, such as the inability to accurately identify conflicts between multiple versions of documents on the same topic, insufficient mining of semantic relationships between documents, and asynchronous updates of related documents. This invention comprises five modules. The improved semantic relationship clustering module introduces version time-series factors and business relationship weights, designs an improved semantic relationship clustering algorithm, and integrates semantic similarity, version time-series, and business relationship to achieve accurate clustering and mine deep semantic relationships between documents. The remaining modules work in conjunction with the improved semantic relationship clustering module to complete document collection, conflict arbitration, knowledge storage, and dynamic updates, forming a closed-loop management system. This invention specifically addresses the specific pain points of enterprise knowledge bases, offering high clustering accuracy, strong practicality, and improved accuracy and timeliness of the knowledge base. It ensures the reliability of AI retrieval and knowledge retrieval, and is suitable for the construction and management of AI data document joint knowledge bases for various types of enterprises.

[0031] 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. An enterprise AI data document joint knowledge base construction and management system, characterized in that: It includes a document multimodal acquisition module, an improved semantic association clustering module, a document version conflict arbitration module, an AI knowledge fusion storage module, and a knowledge dynamic update module. The output of the document multimodal acquisition module is connected to the input of the improved semantic association clustering module. The output of the improved semantic association clustering module is connected to the input of the document version conflict arbitration module and the AI ​​knowledge fusion storage module, respectively. The output of the document version conflict arbitration module is connected to the input of the AI ​​knowledge fusion storage module. The output of the AI ​​knowledge fusion storage module is connected to the input of the knowledge dynamic update module. The output of the knowledge dynamic update module feeds back to the improved semantic association clustering module, forming a closed-loop management. The document multimodal acquisition module is used to collect various AI data documents within the enterprise, including text, table, and image documents; The improved semantic association clustering module receives standardized text data output by the document multimodal acquisition module. Through the improved semantic association clustering algorithm, it clusters the text data according to topic category, semantic association degree, and version time sequence, and outputs the clustering results of multiple versions of documents with the same topic, the document semantic association matrix, and the version time sequence ranking results. This provides accurate clustering basis for the document version conflict arbitration module and semantic association clues for the knowledge dynamic update module. The document version conflict arbitration module receives the clustering results of multiple versions of documents on the same topic output by the improved semantic association clustering module, identifies content conflicts between documents, and arbitrates them. The AI ​​knowledge fusion and storage module is used to extract and fuse knowledge from clustered and arbitrated documents for storage. The knowledge dynamic update module monitors the collection and modification operations of documents in real time and feeds back the relevant documents to the improved semantic association clustering module for re-clustering; Trigger a synchronization update notification for related documents, update the knowledge index and knowledge graph, and delete outdated version information.

2. The enterprise AI data document joint knowledge base construction and management system according to claim 1, characterized in that: The specific working logic of the document multimodal acquisition module is as follows: Documents are acquired through three methods: enterprise internal office system interface, local folder scanning, and employee upload; the acquired documents are parsed to extract text information from text, tables, and images, and non-text information is converted into a processable text format; the parsed text is preliminarily cleaned to remove invalid characters and blank paragraphs, unify text encoding, and output standardized text data to the improved semantic association clustering module.

3. The enterprise AI data document joint knowledge base construction and management system according to claim 1, characterized in that: The specific working steps of the improved semantic association clustering module are as follows: Step 1, Data Input and Preprocessing: Receive standardized text data output from the document multimodal acquisition module, and process each document... Extracting semantic feature vectors using a BERT pre-trained model Simultaneously extract the document's version timestamp. Business Area Tags Establish a basic document information database; Step 2, Parameter Initialization: Statistics of Document Collection Timestamp range, calculation and ; Based on the enterprise's business architecture, preset the correlation weights of each business area. Initialize weight coefficients Set an iteration termination condition, i.e., the number of iterations ≤ 50. The difference is less than 0.001; Step 3, Initial Clustering and Grouping: Document-based Business Domain Tags Documents in the same business domain are grouped into the same initial cluster, resulting in an initial cluster set. Calculate the temporary center semantic vector for each initial cluster; Step 4: Iterative optimization of clustering: 4.1 Calculate the overall similarity: For each initial cluster Any two documents within , Calculate according to the formula , , And thus obtain ; 4.2 Document Reassignment: Calculate each document The overall similarity with each initial cluster temporary center will Assign to the cluster with the highest overall similarity and update the cluster set; 4.3 Update cluster centers: For each updated cluster... Calculate the semantic vector of the central document within the cluster according to the formula. As the new cluster center; 4.4 Validation of the Objective Function: Calculate the objective function for the current cluster set. If the iteration termination condition is met, then stop the iteration; otherwise, repeat steps 4.1-4.3 until the condition is met. Step 5, Cluster Optimization and Adjustment: Merge and split the iterated cluster set: Cluster Merging: Traverse all clusters. If the number of documents in a cluster is less than 3 and the average comprehensive similarity with a neighboring cluster is greater than or equal to 0.7, merge the two clusters and recalculate the center document of the merged cluster; Cluster Splitting: Traverse all clusters. If the average comprehensive similarity of all documents in a cluster is less than 0.5, split the cluster into two sub-clusters. Use the two documents with the highest comprehensive similarity in the cluster as the initial center of the sub-cluster. Repeat the iterative process of Step 4 until the sub-clusters meet the clustering requirements. Step 6, Output Results: After cluster optimization and adjustment, the final set of clusters is obtained. Simultaneously, it generates two core outputs: multi-version document clusters on the same topic, with each cluster containing all versions of documents on the same topic, and labeled with the central document within the cluster and the version chronological order; and a document semantic association matrix, where the matrix elements are the semantic associations between any two documents. The value is used to reflect the degree of semantic relationship between documents; the above outputs are sent to the document version conflict arbitration module and the AI ​​knowledge fusion storage module respectively to provide support for subsequent modules; Step 7, Dynamic Feedback Adaptation: Receive new and modified document data sent by the knowledge dynamic update module, repeat steps 1-6, and update the clustering clusters and semantic association matrix in real time to ensure that the clustering results are synchronized with the document updates, thus ensuring the timeliness of subsequent conflict arbitration and knowledge storage.

4. The enterprise AI data document joint knowledge base construction and management system according to claim 1, characterized in that: The document version conflict arbitration module is implemented as follows: extract the core semantic elements and version timestamps of each document in the cluster; compare the core semantic elements of each version of the document to identify conflict points; determine the optimal version as the benchmark document according to the preset arbitration rules, mark the conflicting documents and provide modification suggestions. The arbitrated baseline document and the marked conflict document are output to the AI ​​knowledge fusion storage module.

5. The enterprise AI data document joint knowledge base construction and management system according to claim 1, characterized in that: The specific implementation logic of the AI ​​knowledge fusion and storage module is as follows: extract the core knowledge of the benchmark document and construct a knowledge graph; associate and store the relationship between the conflicting document and the benchmark document, as well as the modification suggestions, to form a version traceability chain; and adopt a hybrid storage architecture to achieve hierarchical storage of knowledge. A knowledge index is established and associated with the semantic tags output by the improved semantic association clustering module to support AI retrieval.

6. The enterprise AI data document joint knowledge base construction and management system according to claim 1, characterized in that: The specific implementation logic of the knowledge dynamic update module is as follows: Real-time monitoring of newly collected documents from the document multimodal acquisition module, as well as modification operations on existing documents; sending new and modified documents to the improved semantic association clustering module for re-clustering and semantic association mining; if the baseline document is detected to have been modified, triggering a synchronous update prompt for related documents based on the association relationships in the knowledge graph, and notifying relevant departments for review; Update the knowledge index and knowledge graph, and delete outdated version tracking information to ensure the timeliness of the knowledge base.

Images