Artificial intelligence-based enterprise middle platform management method and system, and storage medium

Abstract

The present application relates to the technical field of enterprise digital management and artificial intelligence, and provides an enterprise middle platform management method and system based on artificial intelligence and a storage medium, which comprises the following steps: obtaining cross-department original data and processing the same to obtain a pre-processed data set; extracting entity information and constructing a dynamic enterprise knowledge graph; using a multi-modal embedding model to generate a high-dimensional semantic vector and storing the same; mapping the graph nodes and the vector through a unified identification system; receiving a business request and calling an AI engine layer to execute a processing flow; collecting interaction behavior data and business decision results, converting the same into implicit reward signals and explicit labeled data, and accordingly fine-tuning model parameters to complete self-evolution. The present application realizes semantic alignment of structured and unstructured data, and continuously optimizes the model through a double feedback mechanism to drive the model to adapt, thereby improving the intelligent level and service efficiency of the enterprise middle platform.

Description

Technical Field

[0001] This application relates to the field of enterprise digital management, and in particular to enterprise middleware management methods, systems and storage media based on artificial intelligence. Background Technology

[0002] As enterprises deepen their digital transformation, enterprise middleware, as the core hub connecting front-end applications and back-end data, undertakes the crucial functions of data aggregation, service reuse, and business empowerment. However, in existing technologies, enterprise middleware still faces significant challenges in handling massive amounts of heterogeneous data. On the one hand, enterprises accumulate a large amount of structured business data and unstructured document data, such as contracts, invoices, and resumes. Traditional middleware often adopts a separate storage and processing architecture, resulting in a disconnect between the logical relationships of structured data and the semantic information of unstructured data. This makes it difficult to form a unified enterprise-level knowledge view and limits the ability to perform intelligent correlation analysis across business scenarios. On the other hand, the AI ​​models integrated into existing middleware typically adopt a static deployment mode of offline training and online inference. After the model goes live, it cannot effectively utilize user behavioral feedback data generated in actual business interactions for self-optimization. This mode, lacking a closed-loop feedback mechanism, makes it difficult for AI models to adapt to the dynamic changes and personalized needs of enterprise business rules. As a result, the accuracy of scenarios such as intelligent recruitment and expense verification declines over time, severely restricting the intelligence level and service efficiency of enterprise middleware.

[0003] Therefore, the present invention provides an artificial intelligence-based enterprise middleware management method, system, and storage medium to solve the above-mentioned technical problems. Summary of the Invention

[0004] This application provides an AI-based enterprise middleware management method, system, and storage medium, which solves the technical problems of the silos of structured and unstructured data within existing enterprises, as well as the lack of intelligent processing and model self-evolution capabilities in traditional middleware.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, it provides an AI-based enterprise middleware management approach, including: S1: Obtain raw data from across departments within the enterprise, and process the raw data to obtain a preprocessed dataset; wherein, the raw data includes structured business data from the business database and unstructured document data from the file system; S2: Extract entity information and relationships between entities from the preprocessed dataset, and construct a dynamic enterprise knowledge graph that includes project, personnel, and financial dimensions; The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. By using a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library, thereby achieving semantic alignment between structured and unstructured data. S3: Receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow; among which, business requests include project progress query, intelligent recruitment screening, intelligent expense verification, and intelligent document retrieval; S4: Collect user interaction behavior data and final business decision results from S3 at the application layer; convert the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; use the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the self-evolution of the model; wherein, the interaction behavior data includes click location, page dwell time, download operation, and manual correction records, and the final business decision results include the hiring / elimination status and elimination reason in the recruitment process, and the approval / rejection status and rejection reason code in the expense reimbursement process.

[0006] In conjunction with the first aspect above, in one possible implementation, the original data is processed to obtain a preprocessed dataset, including: The structured business data is cleaned and formatted using an ETL engine to generate standardized structured data; the unstructured document data is denoised and extracted using optical character recognition technology and natural language preprocessing technology to generate standardized text data; a unified data dictionary is established to map the standardized structured data and the standardized text data to a unified data identification system to obtain a preprocessed dataset.

[0007] In conjunction with the first aspect above, one possible implementation involves extracting entity information and relationships between entities from the preprocessed dataset, and constructing a dynamic enterprise knowledge graph including project, personnel, and financial dimensions, including: Predefine an ontology containing four core entities: projects, personnel, finance, and documents, and their interrelationships, and establish the attribute fields and data type specifications for each entity; For structured data, entity nodes are instantiated directly according to field mapping rules and assigned a globally unique identifier; for unstructured text data, potential entities are extracted using a BERT-based named entity recognition model, and entity disambiguation and linking are performed by calculating the cosine similarity between the entity feature vector and the existing entity vector in the knowledge base to determine the entity identity. For structured data, relation edges between entities are directly generated based on foreign key associations and business logic rules; for unstructured text, a relation extraction model trained by remote supervised learning is used to identify semantic relationships between entities from the sentence context and generate corresponding relation edges. A timestamp attribute is introduced for each node attribute and relation edge. When a data change is detected, historical version records are retained and the effective time interval of the new version is marked, thereby forming a dynamic knowledge graph with time-series evolution capabilities. The nodes, relationships, and time-series attributes mentioned above are stored in a graph database, and a hash index based on entity ID and a full-text index based on text content are established to support subsequent millisecond-level graph traversal queries.

[0008] In conjunction with the first aspect above, in one possible implementation, a multimodal embedding model is used to encode features of the unstructured document data in the preprocessed dataset to generate high-dimensional semantic vectors, and the high-dimensional semantic vectors are stored in a multimodal vector library, including: The document image is divided into several image blocks and the coordinate information is extracted. At the same time, the document text is segmented. The spatial mapping relationship between text fragments and corresponding image regions is established to form a multimodal input sequence. Visual feature vectors of image patches are extracted using a visual Transformer model to capture the document’s layout, seals, and table structure information; semantic feature vectors of text are extracted using a pre-trained language model to capture the document’s semantic content. The visual feature vector is concatenated with the text feature vector, and modality type embedding and position embedding are added; the input is fed into a multimodal Transformer network, and the feature interaction and fusion of text semantics and visual layout are realized through the cross attention mechanism to generate a context feature sequence containing comprehensive text and image information; The attention pooling mechanism is used to weight and aggregate the context feature sequences to obtain a fixed-dimensional vector representing the semantics of the entire document. The fixed-dimensional vector is subjected to L2 normalization to generate the final high-dimensional semantic vector, which is then associated with the unique document identifier and stored in the multimodal vector library.

[0009] In conjunction with the first aspect mentioned above, one possible implementation involves using a unified data identification system to associate and map nodes in the dynamic enterprise knowledge graph with vectors in the multimodal vector library, including: Based on the preset globally unique identifier generation rules, a unique GUID is assigned to each structured business entity and unstructured document, and a mapping registry from the source system's original ID to the GUID is established. When constructing a dynamic enterprise knowledge graph, the GUID is written as a core attribute into the corresponding node in the graph database, and a unique index is established for the core attribute, so that the graph node can be uniquely addressed by the GUID. When storing the high-dimensional semantic vector of an unstructured document into a multimodal vector library, the corresponding document GUID is stored together with the vector as scalar metadata, and a filtering index is established for the scalar metadata field. The algorithm analyzes the reference edges between entities and documents in the knowledge graph, extracts the pairing relationship between the start node GUID and the end node GUID, and constructs a bidirectional mapping index of entity GUID-document GUID. Furthermore, for entity mentions extracted from the unstructured document, the algorithm generates a contextual semantic vector of the entity mention through the multimodal embedding model. After locating the corresponding entity node in the graph, the contextual semantic vector is used as the association vector of the entity node in a specific document context, stored in the multimodal vector library, and associated with the entity GUID, thereby achieving fine-grained semantic alignment between graph nodes and document fragments. During the data update process, a transaction mechanism is used to synchronously update the graph node attributes, vector library metadata, and bidirectional mapping index, ensuring that the logical relationships in the knowledge graph and the semantic data in the vector library remain consistent in real time.

[0010] In conjunction with the first aspect mentioned above, one possible implementation involves receiving a business request initiated by the user through the application layer and invoking the AI ​​engine layer to execute the processing flow, including: When the request is for a project progress query: the target project number in the request is parsed, and a breadth-first search algorithm is executed in the dynamic enterprise knowledge graph with the target project number as the starting node to traverse the associated financial payment nodes, project progress nodes, and human resource input nodes to obtain the status attribute values ​​of each node; the status attribute values ​​are aggregated using a graph neural network to calculate the project's comprehensive health score and risk warning index, and the comprehensive health score and risk warning index are returned to the application layer for display; When the request is for intelligent recruitment screening: The uploaded candidate resume file and preset job description text are retrieved; the candidate resume file is encoded using a multimodal embedding model to generate a resume feature vector; the job description text is encoded using a natural language processing model to generate a job requirement vector; the cosine similarity between the resume feature vector and the job requirement vector in the multimodal vector library is calculated to obtain a matching score; candidate information with a matching score higher than a preset threshold is used as the screening result and synchronized to the human resources module. When the request is for intelligent expense reimbursement verification: The uploaded invoice image file is retrieved; the layout and seal texture features of the invoice image file are extracted using a computer vision model, and the invoice text information is extracted using optical character recognition (OCR) technology; the invoice text information is compared with the authenticity verification result returned by the official tax interface, and the layout and seal texture features are input into a consistency verification model to determine if there is a logical conflict between the image features and the text information; if the authenticity verification passes and there is no logical conflict, a standardized expense reimbursement form is automatically generated and pushed to the financial approval process; otherwise, an anomaly marker is generated and the application layer is returned to prompt manual intervention. When the request is for intelligent document retrieval: obtain the natural language retrieval statement input by the user; use a semantic retrieval model to convert the natural language retrieval statement into a query vector; calculate the similarity between the query vector and the stored high-dimensional semantic vector in the multimodal vector library, and recall the top-ranked document IDs; read the corresponding original document from the file system according to the document ID, and return the original document to the application layer for display.

[0011] In conjunction with the first aspect above, in one possible implementation, the interactive behavior data is transformed into implicit reward signals, and the final business decision result is transformed into explicit labeled data, including: Collect user behavior logs during the search, recruitment, and expense reimbursement processes, including click locations, page dwell time, download operations, and manual correction records; calculate the above behavior logs into numerical implicit reward scores based on a preset behavior weight matrix, and perform time decay weighting and normalization processing to obtain implicit reward signals; Collect the decision results at the end of the business process, construct positive sample pairs between the candidates to be hired and the job descriptions, construct negative sample pairs between the candidates to be rejected and the job descriptions, and map the reasons for rejection to standard error type labels to form explicit labeled data.

[0012] In conjunction with the first aspect above, in one possible implementation, the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model are fine-tuned and updated using the implicit reward signal and explicit labeled data, including: The implicit reward signal is used to construct a ranking loss function to optimize the model's ability to rank documents or candidates based on their relevance; the explicit labeled data is used to construct a regression loss function to optimize the model's accuracy in identifying entity attributes and judging authenticity. Based on the ranking loss function and regression loss function, the gradient is calculated using the backpropagation algorithm. The parameters of the multimodal embedding model, semantic retrieval model and consistency verification model in the AI ​​engine layer are fine-tuned in small batches. A replay mechanism is used to prevent catastrophic forgetting and to achieve continuous iteration of model performance.

[0013] Secondly, it provides an AI-based enterprise middleware management system, including: a communication unit and a processing unit; The communication unit is used to acquire raw data from across departments within the enterprise and process the raw data to obtain a preprocessed dataset; wherein the raw data includes structured business data from the business database and unstructured document data from the file system; The processing unit is used to extract entity information and relationships between entities from the preprocessed dataset, and to construct a dynamic enterprise knowledge graph that includes project, personnel, and financial dimensions. The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. Through a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library; It receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow; among them, business requests include project progress query, intelligent recruitment screening, intelligent expense verification, and intelligent document retrieval; The system collects user interaction behavior data and final business decision results at the application layer; transforms the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; and uses the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the self-evolution of the model. The interaction behavior data includes click locations, page dwell time, download operations, and manual correction records; the final business decision results include the hiring / rejection status and rejection reasons in the recruitment process, and the approval / rejection status and rejection reason codes in the expense reimbursement process.

[0014] Thirdly, this application provides a storage medium storing instructions that, when executed on an AI-based enterprise middleware management system, cause the AI-based enterprise middleware management system to perform the methods described in the first aspect and any possible implementation thereof.

[0015] This application provides an enterprise middleware management method, system, and storage medium based on artificial intelligence, with the following beneficial effects: 1. This invention fundamentally solves the long-standing semantic disconnect between structured business data and unstructured document data in enterprise middleware by constructing a dual representation system of a dynamic enterprise knowledge graph and a multimodal vector library, and by utilizing a unified globally unique identifier to achieve deep association mapping between the two. Specifically, the dynamic knowledge graph accurately depicts the logical relationships and temporal evolution between core entities such as projects, personnel, and finance in a symbolic form, preserving the determinism and interpretability of business rules; while the multimodal vector library, through visual-text joint encoding, transforms unstructured documents such as invoices, resumes, and official documents into high-dimensional semantic vectors, capturing rich implicit semantics and layout information. The unified identifier system, acting as a "semantic anchor," not only enables bidirectional addressing between graph nodes and document vectors but also supports cross-modal information enhancement and verification during retrieval and reasoning. For example, in a project risk assessment scenario, the system can simultaneously retrieve the status of financial payment nodes in the graph and the semantic features of relevant contract documents in the vector library for multi-dimensional comprehensive analysis. This converged architecture transforms enterprise data assets from isolated "information fragments" into an interconnected "knowledge network," greatly improving data reuse rates and intelligent correlation analysis capabilities across business scenarios, providing a solid cognitive intelligence foundation for enterprise digital transformation.

[0016] 2. This invention innovatively designs a dual feedback mechanism based on implicit interactive behaviors and explicit business decisions, driving the AI ​​model to achieve continuous self-evolution and performance optimization in the actual enterprise operating environment, effectively breaking through the technical bottleneck of traditional AI platform models that are "fixed upon deployment and become less accurate with use." In specific implementation, the system transforms implicit user behaviors such as clicks, dwell times, downloads, and corrections into implicit reward signals with time decay, used to construct a ranking loss function, enabling the model to keenly capture real-time changes in user preferences for the relevance of search results. Simultaneously, explicit business final states such as recruitment / recruitment and expense reimbursement approval / rejection are transformed into standard labeled data, used to construct regression or classification loss functions, ensuring that the model output always aligns with the enterprise's rigid business rules. This dual feedback mechanism utilizes the statistical regularity of massive implicit behavioral data while ensuring the accuracy and compliance of key business decisions. Through small-batch online fine-tuning and experience playback strategies, the model can continuously absorb new knowledge and adapt to new rules without interrupting service, forming a positive flywheel effect of "business usage - feedback collection - model optimization - experience improvement." This not only significantly reduces the cost of manual annotation for model maintenance, but also enables AI capabilities to truly evolve dynamically along with the company's growth, significantly extending the life cycle and return on investment of AI systems.

[0017] 3. This invention establishes a consistency verification and unified identification management system covering the entire chain, providing reliable intelligent decision-making support for enterprise middleware in complex and ever-changing business scenarios, effectively avoiding business risks caused by data inconsistency or model illusions. At the data level, a unified data dictionary and GUID mapping registry ensure semantic consistency and reference integrity of heterogeneous data sources in all stages of access, processing, storage, and application. The transaction mechanism further ensures the synchronization of the graph and vector library under concurrent updates, eliminating decision-making biases caused by data conflicts at the root. At the model application level, especially in high-risk scenarios such as intelligent expense reimbursement verification, the system not only relies on the authenticity comparison between OCR text and tax interfaces, but also introduces a consistency verification model to cross-verify visual features such as the layout and seal texture of invoice images with text information, using a dual-tower neural network to determine whether the graphic and text logic is self-consistent. This multimodal cross-verification mechanism is equivalent to adding a "common sense verification" defense line to AI decision-making, effectively identifying fraudulent behaviors such as forged invoices and tampered information that are difficult to detect by traditional single-modal methods. Meanwhile, the time-series version control capability of the dynamic knowledge graph ensures that any business decision can be traced back to a specific point in time, meeting the stringent requirements of enterprise auditing and compliance supervision. In summary, this invention enhances the level of intelligence without sacrificing the determinism, security, and traceability necessary for enterprise-level applications, achieving an organic unity of intelligence and robustness.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A schematic diagram illustrating the steps of an AI-based enterprise middleware management method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of an AI-based enterprise middleware management system provided in an embodiment of this application. Detailed Implementation

[0020] In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. The "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. Furthermore, "at least one" means one or more, and "multiple" means two or more. The terms "first," "second," etc., do not limit the quantity or order of execution, and "first," "second," etc., do not necessarily imply differences.

[0021] It should be noted that, in this application, the terms "exemplary" or "for example" are used to indicate that something is being described as an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or design solutions. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a concrete manner.

[0022] like Figure 1 As shown in the embodiments of this application, the enterprise middleware management method based on artificial intelligence includes: S1: Obtain raw data from across departments within the enterprise, and process the raw data to obtain a preprocessed dataset; wherein, the raw data includes structured business data from the business database and unstructured document data from the file system; S2: Extract entity information and relationships between entities from the preprocessed dataset, and construct a dynamic enterprise knowledge graph that includes project, personnel, and financial dimensions; The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. Through a unified data identification system, nodes in the dynamic enterprise knowledge graph are associated and mapped with vectors in the multimodal vector library, achieving semantic alignment between structured and unstructured data. Specifically, this step constructs a dual-track representation system that combines symbolic knowledge and vectorized semantics. The dynamic enterprise knowledge graph focuses on expressing definite logical relationships and temporal evolution between entities, such as the attribution relationship between projects and personnel, and the causal relationship between financial payments and project progress, supporting accurate multi-hop reasoning and state tracing. The multimodal vector library, on the other hand, compresses the text content, layout, and image features of documents into high-dimensional dense vectors through an encoding model, capturing implicit semantic similarities that are difficult to describe with explicit rules, supporting fuzzy retrieval and semantic matching. More importantly, the unified data identification system bidirectionally anchors graph nodes and document vectors, enabling the system to seamlessly cross the two representation spaces when performing tasks. For example, when querying the risk of a project, it can either traverse along the graph edges to obtain associated financial anomaly nodes or directly recall the semantic vectors of relevant contract documents for content verification through identification mapping. This association mapping mechanism fundamentally solves the technical problem of the separation between structured logic and unstructured semantics, and realizes the holographic representation of enterprise knowledge.

[0023] S3: Receives business requests initiated by users through the application layer and invokes the AI ​​engine layer to execute processing flows. These business requests include project progress queries, intelligent recruitment screening, intelligent expense reimbursement verification, and intelligent document retrieval. Specifically, the application layer, as the human-computer interaction interface, translates users' natural language commands or form operations into standardized service call requests. The AI ​​engine layer, as the intelligent hub, dynamically orchestrates underlying capabilities based on the request type. For project progress queries, the AI ​​engine primarily relies on the traversal and aggregation capabilities of the knowledge graph to calculate health status. For intelligent recruitment and document retrieval, it focuses on utilizing the semantic similarity matching capabilities of the vector library. For intelligent expense reimbursement verification, it requires the coordinated invocation of multiple capabilities such as OCR recognition, knowledge graph rule verification, and vector consistency comparison. This step embodies the core value of the middle platform's "capability reuse," namely, shielding the complexity of underlying data and models through a unified AI engine layer, providing standardized and intelligent service interfaces to upper-layer applications, and supporting rapid business innovation and iteration.

[0024] S4: Collect user interaction behavior data and final business decision results from S3 at the application layer; convert the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; use the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the model's self-evolution; wherein, the interaction behavior data includes click location, page dwell time, download operation, and manual correction records, and the final business decision results include the hiring / elimination status and elimination reasons in the recruitment process, and the approval / rejection status and rejection reason code in the expense reimbursement process. Specifically, this step constructs a closed-loop self-evolution mechanism of "business use - feedback collection - model optimization," which is a key innovation that distinguishes it from the traditional static AI deployment model. Implicit reward signals originate from real-time interactions between users and the system. For example, user clicks, prolonged dwell times, or downloads on search results suggest the relevance and value of the results, while manual corrections directly indicate model biases. These signals are characterized by large data volumes, low acquisition costs, and immediate feedback, primarily used to optimize ranking strategies and user experience. Explicitly labeled data, on the other hand, originates from the final confirmation of business processes, such as HR hiring decisions or financial audit results. These represent definitive truth values ​​that comply with company rules and regulations, primarily used to calibrate the model's classification accuracy and compliance. By transforming these two complementary signals into a loss function and driving online or near-line fine-tuning of model parameters, the system can continuously absorb the knowledge and rule changes contained in the latest business data, effectively overcoming performance degradation caused by business environment drift after model deployment. For example, when a company adjusts its reimbursement standards or recruitment preferences, the model can automatically adapt to the new rules based on recent rejection records or screening behaviors without the need for manual relabeling of massive amounts of data, thus achieving synchronous evolution of AI capabilities and business development.

[0025] It should be noted that structured business data includes: financial data, engineering project data, and recruitment data; unstructured document data includes: resume documents, images of invoices and receipts, and official documents and archives.

[0026] In one possible implementation of this application embodiment, the above-mentioned S1 can be implemented by the following S101, which will be described in detail below: S101: Process the original data to obtain a preprocessed dataset, including: The structured business data is cleaned and formatted using an ETL engine to generate standardized structured data; the unstructured document data is denoised and extracted using optical character recognition technology and natural language preprocessing technology to generate standardized text data; a unified data dictionary is established to map the standardized structured data and the standardized text data to a unified data identification system to obtain a preprocessed dataset.

[0027] It should be noted that establishing a unified data dictionary specifically includes: Collect metadata information from various heterogeneous data sources, including database table structure, field definitions, and document type characteristics; Using pre-trained language models, the semantic similarity between fields from different source systems is calculated to identify field clusters with the same business meaning. For example, 'Employee Number', 'Employee ID', and 'User_Code' are identified as the same semantic entity. Construct an enterprise-level business ontology model, define standard entities and their attribute specifications, and form standard data; Establish a mapping rule library from source data fields to standard data. The mapping rule library includes data type conversion rules, value range enumeration mapping rules, and unit conversion rules. During data cleaning, the ETL engine uses the mapping rule base to convert heterogeneous data into standardized data that conforms to the standard, and generates a globally unique identifier for each data record based on the source system identifier and the original primary key, thereby obtaining a preprocessed dataset.

[0028] In one possible implementation of this application embodiment, the above-mentioned S2 can be implemented by the following S201, S202 and S203, which are described in detail below: S201: Extract entity information and relationships between entities from the preprocessed dataset, and construct a dynamic enterprise knowledge graph including project, personnel, and financial dimensions, including: Predefine an ontology containing four core entities: projects, personnel, finance, and documents, and their interrelationships, and establish the attribute fields and data type specifications for each entity; For structured data, entity nodes are directly instantiated according to field mapping rules and assigned globally unique identifiers. For unstructured text data, a BERT-based named entity recognition model is used to extract latent entities. Entity disambiguation and linking are performed by calculating the cosine similarity between the entity feature vector and existing entity vectors in the knowledge base to determine the entity's identity. Specifically, this step employs a hybrid extraction strategy to adapt to the characteristics of heterogeneous data sources. For structured data from business databases, which have clear table structures and field definitions, the system directly instantiates database records into entity nodes in the graph according to the field mapping rules in the unified data dictionary, inheriting the globally unique identifiers generated during the preprocessing stage. This approach ensures the accuracy and efficiency of inputting structured data into the graph. For unstructured text data such as contracts and meeting minutes, which lack explicit structure and contain numerous cases of ambiguous referencing or homonyms, the system uses a BERT-based named entity recognition model for latent entity extraction. The BERT model, with its deep bidirectional attention mechanism, effectively captures contextual information, significantly improving its ability to recognize complex linguistic phenomena such as omitted subjects and abbreviations. More importantly, to address the common entity ambiguity problem in unstructured text—for example, "Li Ming" in the text might refer to Li Ming A in the R&D department or Li Ming B in the sales department—this invention introduces an entity disambiguation mechanism based on vector similarity. The system converts entities extracted by the NER model into feature vectors and calculates their cosine similarity with existing entity vectors in the knowledge base. When the similarity exceeds a preset threshold, such as 0.85, it is determined to be the same entity and a link is established; if it is below the threshold but above the lower limit, such as 0.6, manual verification or secondary discrimination based on more contextual features is triggered; if it is below the lower limit, it is considered a new entity and a node is created. This disambiguation method based on semantic vectors rather than simple string matching significantly reduces graph noise caused by mislinks with the same name, ensuring the accuracy of entity identities in the knowledge graph.

[0029] For structured data, relation edges between entities are directly generated based on foreign key associations and business logic rules; for unstructured text, a relation extraction model trained by remote supervised learning is used to identify semantic relationships between entities from the sentence context and generate corresponding relation edges. A timestamp attribute is introduced for each node attribute and relation edge. When a data change is detected, historical version records are retained and the effective time interval of the new version is marked, thereby forming a dynamic knowledge graph with time-series evolution capabilities. The aforementioned nodes, relationships, and time-series attributes are stored in a graph database, and a hash index based on entity ID and a full-text index based on text content are established to support subsequent millisecond-level graph traversal queries. Specifically, relationship construction also distinguishes between structured and unstructured paths. The generation of relationship edges in structured data relies on database foreign keys and predefined business rules. For example, a "project-personnel" relationship edge can be directly established through the "person in charge ID" field in the "project table." This type of relationship has the characteristics of high determinism and strong confidence. Relationship extraction from unstructured text adopts a remote supervised learning paradigm, using existing knowledge base triples as weak labels to automatically annotate the corpus and train a relationship classification model. This allows for the mining of implicit semantic relationships from massive amounts of unlabeled text, solving the problem of traditional supervised learning's dependence on large-scale manually labeled data. Based on this, this invention introduces a dynamic time-series management mechanism, which is a key innovation that distinguishes it from the static snapshot graph in existing technologies. Specifically, the system adds a timestamp attribute, including "effective time" and "expiration time," to each node attribute value and each relationship edge in the graph. When source data changes, for example, the person in charge of a project changes from Zhang San to Li Si, the system does not directly overwrite the original relationship edges. Instead, it marks the "expiration time" of the original relationship edges with the current time, and simultaneously adds a new relationship edge pointing to the new person in charge, setting its "effective time" to the current time. This versioning management mechanism gives the knowledge graph complete historical backtracking capabilities, supporting time-sensitive queries such as "querying all persons in charge of the project in the previous quarter" or "tracing the project status at the time of a certain expense reimbursement approval." In scenarios such as enterprise compliance auditing, risk tracing, and business trend analysis, this dynamic evolution capability is crucial. It ensures that the graph not only reflects the current state of the business but also faithfully records the development path of the business, avoiding the loss of historical information due to data updates.

[0030] S202: Utilizing a multimodal embedding model to perform feature encoding on the unstructured document data in the preprocessed dataset, generating high-dimensional semantic vectors, and storing the high-dimensional semantic vectors in a multimodal vector library, including: The document image is divided into several image patches and their coordinate information is extracted. Simultaneously, the document text is segmented. A spatial mapping relationship between text fragments and corresponding image regions is established, forming a multimodal input sequence. Specifically, corporate documents, such as invoices, contracts, and resumes, have strong two-dimensional spatial attributes. The absolute position and relative layout of text on the page themselves carry key business semantics. For example, the "invoice code" of a VAT invoice is usually located in a specific area in the upper left corner, the "total price including tax" is located at the bottom of the table in the lower right corner, and the official seal must cover the name of the issuing unit. If the document is only treated as a one-dimensional text stream, this crucial layout information will be completely lost, causing subsequent models to be unable to distinguish between the main text content and header noise, and unable to verify the spatial logical constraints between the text and images. Therefore, this embodiment first divides the document image into fixed-size image patches, such as 16×16 pixels, and records the two-dimensional coordinates of each image patch in the original image. Simultaneously, the text recognized by OCR is segmented, and the bounding box information returned by the OCR engine is used to establish a spatial alignment mapping between the text token and its corresponding image patch. This explicit spatial mapping mechanism enables subsequent neural networks to perceive "which word appears in which position," laying a structured foundation for understanding complex document layouts.

[0031] This process utilizes a visual Transformer model to extract visual feature vectors from image patches, capturing document layout, seal, and table structure information. A pre-trained language model is used to extract semantic feature vectors from the text, capturing the document's semantic content. Specifically, this step employs a heterogeneous dual-encoder parallel processing strategy. The visual branch uses the Vision Transformer (ViT) architecture, whose self-attention mechanism can overcome the limitations of local receptive fields and globally model the dependencies between image patches, effectively extracting pure visual features such as table line direction, seal texture integrity, and paper background watermarks. These features are crucial for determining document authenticity and recognizing table row and column structures, features that traditional OCR text results cannot provide. The text branch uses a pre-trained language model (PLM) focused on understanding the deep semantic representations of words, phrases, and sentences, such as recognizing the meanings of business concepts like "Party A," "Party B," and "penalty for breach of contract." It should be understood that although this embodiment exemplarily uses a combination of ViT and PLM, in other embodiments, it can be replaced with visual backbone networks such as ResNet and CNN, or language model variants such as BERT and RoBERTa, depending on the computing power and accuracy requirements. As long as they respectively possess visual feature extraction and text semantic encoding capabilities, they all fall within the protection scope of this invention. This dual-stream design ensures that the physical morphological information and logical semantic information of a document are encoded independently and fully, avoiding the representation bottleneck of a single modal encoder when processing cross-modal tasks.

[0032] The visual feature vector is concatenated with the text feature vector, and modality type embedding and position embedding are added. This concatenation is then fed into a multimodal Transformer network. A cross-attention mechanism enables feature interaction and fusion between text semantics and visual layout, generating a contextual feature sequence containing comprehensive text and image information. Specifically, simple vector concatenation only achieves physical connection of features and does not achieve deep semantic fusion. Therefore, this invention introduces a multi-layer multimodal Transformer layer and embeds a crucial cross-attention module within it. Under this mechanism, a text token can be used as a query to retrieve a related image patch (key / value), and vice versa. This means the model can dynamically learn the correspondence between "text content" and "visual regions." For example, when processing the text fragment "Actual amount received: ¥10,000.00", cross-attention automatically focuses on the corresponding numerical pixel area in the image, using visual features to verify the accuracy of OCR recognition, and combining this with the presence of surrounding border lines to determine whether the number belongs to a table cell. Similarly, when detecting text related to "stamp," the model focuses on the red ink texture features of the corresponding area to determine whether the stamp is clear and whether it overlaps the text. This fine-grained interaction between text and images means that the generated contextual feature sequence is no longer an isolated text or image representation, but a comprehensive semantic representation that integrates "what the content is" and "where the content is and what it looks like," greatly improving the model's depth of understanding and robustness in handling complex business documents.

[0033] This embodiment employs attention pooling to weighted aggregate the contextual feature sequences, resulting in a fixed-dimensional vector representing the semantics of the entire document. Specifically, due to significant differences in document length, resolution, and content density, the directly output feature sequence length is variable, failing to meet the requirements of fixed-length indexes in vector databases. Traditional average pooling or max pooling often ignores the importance distribution within the document, easily introducing substantial background noise. The attention pooling mechanism used in this embodiment learns a trainable query token and weights the entire contextual feature sequence. This mechanism adaptively identifies and focuses on the most discriminative key areas in the document, such as titles, key fields, and signature areas, while suppressing the weight of low-value information like headers, footers, and blank areas. The final output fixed-dimensional vector, such as 768-dimensional or 1024-dimensional, not only has a uniform length but also highly condenses the core business intent of the document semantically, significantly improving the accuracy of subsequent retrieval and matching.

[0034] The fixed-dimensional vector is subjected to L2 normalization to generate the final high-dimensional semantic vector, which is then associated with the unique document identifier and stored in the multimodal vector library.

[0035] S203: Through a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library, including: Based on preset globally unique identifier generation rules, a unique GUID is assigned to each structured business entity and unstructured document, and a mapping registry from the source system's original ID to the GUID is established. Specifically, the GUID is the "semantic anchor" that runs through the entire enterprise middleware data link. Although Example 2 mentioned generating GUIDs during the data preprocessing stage, this example focuses on its registration and management mechanism in cross-database association. The mapping registry not only records the correspondence between GUIDs and the source system's original IDs, such as material codes in ERP and document numbers in OA, but also maintains the GUID's generation time, data source, and status flag. Regarding the generation rules, this invention is not limited to a specific algorithm. It can use standard UUID v4 to ensure global randomness, or distributed ID generation algorithms such as Snowflake to balance trend increment and sorting performance, as long as it satisfies global uniqueness and can be generated by machine. This independent mapping registry design allows the system to expand the identification system by updating the registry without modifying the underlying schema of the graph or vector library when connecting to new business systems or migrating historical data, greatly improving the system's compatibility and maintainability.

[0036] When constructing a dynamic enterprise knowledge graph, the GUID is written as a core attribute into the corresponding node in the graph database, and a unique index is established for the core attribute, so that the graph node can be uniquely addressed by the GUID. When storing high-dimensional semantic vectors of unstructured documents into a multimodal vector library, the corresponding document GUID is stored as scalar metadata along with the vector, and a filtering index is established for the scalar metadata field. Specifically, this dual-end anchoring strategy is a targeted optimization for the heterogeneous characteristics of graph databases and vector databases. On the graph database side, the GUID replaces the database's internal auto-incrementing ID as a business primary key. The unique index ensures that when performing precise queries such as "finding contracts associated with a certain project," nodes can be directly located, avoiding full table scans. On the vector database side, the GUID is not a dimension involved in similarity calculation but is stored as payload or metadata. The significance of establishing a filtering index is to support "constrained vector retrieval." For example, when retrieving "acceptance reports for project A," the system can first use the filtering index to filter out all vector subspaces where the GUIDs belong to the set of documents associated with "project A," and then perform semantic similarity searches within that subspace. This "filter before retrieval" or "filter during retrieval" mechanism, compared to searching the entire vector library and then filtering according to business rules, can reduce retrieval latency by an order of magnitude, while completely eliminating the risk of vector invalidation caused by changes in business IDs.

[0037] The algorithm analyzes the reference edges between entities and documents in the knowledge graph, extracts the pairing relationship between the start node GUID and the end node GUID, and constructs a bidirectional mapping index of entity GUID-document GUID. Furthermore, for entity mentions extracted from the unstructured document, the algorithm generates a contextual semantic vector of the entity mention through the multimodal embedding model. After locating the corresponding entity node in the graph, the contextual semantic vector is used as the association vector of the entity node in a specific document context, stored in the multimodal vector library, and associated with the entity GUID, thereby achieving fine-grained semantic alignment between graph nodes and document fragments. During the data update process, a transaction mechanism is used to synchronously update the graph node attributes, vector library metadata, and bidirectional mapping index, ensuring that the logical relationships in the knowledge graph and the semantic data in the vector library remain consistent in real time.

[0038] It should be noted that the multimodal vector library adopts a hierarchical storage structure: Document layer vector: Stores the global semantic vector of the entire document, used for coarse-grained tasks such as document retrieval; Fragment / Entity Layer Vectors: Store local semantic vectors of key paragraphs or identified entity references in a document, used for fine-grained knowledge graph association; Association Mapping: In the vector library, each vector record contains its own document ID and fragment ID; Retrieval Strategy: When performing document retrieval, document layer vectors are matched first; when performing knowledge graph completion or entity linking, fragment layer vectors are matched, and the original text is traced back through the own document ID.

[0039] In one possible implementation of this application embodiment, the above-mentioned S3 can be implemented by the following S301, which will be described in detail below: S301: Receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow, including: When the request is for a project progress query: the target project number in the request is parsed, and a breadth-first search algorithm is executed in the dynamic enterprise knowledge graph with the target project number as the starting node to traverse the associated financial payment nodes, project progress nodes, and human resource input nodes to obtain the status attribute values ​​of each node; the status attribute values ​​are aggregated using a graph neural network to calculate the project's comprehensive health score and risk warning index, and the comprehensive health score and risk warning index are returned to the application layer for display; When the request is for intelligent recruitment screening: The uploaded candidate resume file and preset job description text are retrieved; the candidate resume file is encoded using a multimodal embedding model to generate a resume feature vector; the job description text is encoded using a natural language processing model to generate a job requirement vector; the cosine similarity between the resume feature vector and the job requirement vector in the multimodal vector library is calculated to obtain a matching score; candidate information with a matching score higher than a preset threshold is used as the screening result and synchronized to the human resources module. When the request is for intelligent expense reimbursement verification: The uploaded invoice image file is retrieved; the layout and seal texture features of the invoice image file are extracted using a computer vision model, and the invoice text information is extracted using optical character recognition (OCR) technology; the invoice text information is compared with the authenticity verification result returned by the official tax interface, and the layout and seal texture features are input into a consistency verification model to determine if there is a logical conflict between the image features and the text information; if the authenticity verification passes and there is no logical conflict, a standardized expense reimbursement form is automatically generated and pushed to the financial approval process; otherwise, an anomaly marker is generated and the application layer is returned to prompt manual intervention. When the request is for intelligent document retrieval: obtain the natural language retrieval statement input by the user; use a semantic retrieval model to convert the natural language retrieval statement into a query vector; calculate the similarity between the query vector and the stored high-dimensional semantic vector in the multimodal vector library, and recall the top-ranked document IDs; read the corresponding original document from the file system according to the document ID, and return the original document to the application layer for display.

[0040] In one possible implementation of this application embodiment, the above-mentioned S4 can be implemented by the following S401 and S402, which are described in detail below: S401: The interactive behavior data is converted into implicit reward signals, and the final business decision result is converted into explicit labeled data, including: User behavior logs are collected during the search, recruitment, and expense reimbursement processes, including click locations, page dwell time, download operations, and manual correction records. Based on a pre-defined behavior weight matrix, these behavior logs are calculated into numerical implicit reward scores, which are then subjected to time-decay weighting and normalization to obtain implicit reward signals. Specifically, the generation of implicit reward signals is not a simple behavior count, but a quantification process that integrates business semantics and timeliness. The pre-defined behavior weight matrix defines the contribution and direction of different interaction actions to model optimization. For example, in a preferred configuration, a user's click on a search result is assigned a base score of 1.0, indicating a preliminary relevance to the query intent; if the user's dwell time on the result page exceeds 30 seconds, an additional 0.5 points are added, suggesting substantial value in the content; performing a download operation adds 0.8 points, representing a strong demand fulfillment; conversely, if the user manually corrects the information automatically filled in by the system, a deduction of -0.2 to -1.0 is applied based on the correction magnitude, directly indicating the model's prediction bias. This fine-grained weighting design enables the system to extract high-quality teaching signals from massive amounts of noisy user behavior. Furthermore, considering that business rules and user preferences evolve dynamically over time, this embodiment introduces a time-decay weighting mechanism. Actions more recent to the current moment are assigned higher weight coefficients, while the weight of older actions decreases exponentially. This ensures the model can accurately capture the latest business trends and avoids being misled by outdated historical data. Finally, the weighted and accumulated scores are normalized and mapped to the [0,1] interval, forming a standardized implicit reward signal, providing a continuous supervisory target for subsequent ranking optimization.

[0041] The system collects decision results at the end of business processes, constructs positive sample pairs between hired candidates and job descriptions, and negative sample pairs between rejected candidates and job descriptions. Rejection reasons are mapped to standard error type labels, forming explicit labeled data. Specifically, unlike the ambiguity of implicit signals, explicit labeled data represents the definitive truth value in the enterprise's business processes. In intelligent recruitment scenarios, HR's final hiring decision is considered the gold standard. The system uses "hired candidate resume - job description" pairings as positive samples to strengthen the model's memory of this matching pattern; simultaneously, "rejected candidate resume - job description" pairings are used as negative samples to help the model learn to distinguish the feature boundaries of mismatches. In intelligent expense reimbursement scenarios, the review results of finance personnel are also used in a structured manner: approved documents serve as positive examples of correct identification; while rejected documents, with their rejection reason codes, such as "blurred invoice stamp," "inconsistent amount in words and figures," and "over budget," are mapped to standard error type labels as fine-grained classification supervision signals. This mechanism, which transforms the final state of a business into machine learning labels, solves the problems of high manual labeling costs and the disconnect between labeling standards and actual business in traditional AI training, ensuring that the direction of model optimization always remains consistent with the rigid compliance requirements of enterprises.

[0042] It should be noted that the preset behavior weight matrix is ​​as follows: clicking on a location item earns a base score of 1.0; staying for a long time, such as more than 30 seconds, adds 0.5; performing a download operation adds 0.8; and manual correction is converted into a deduction of -0.2 to -1.0 according to the correction range. The final score is normalized to the [0,1] interval after time decay weighting to obtain the implicit reward score. When constructing the ranking loss function, this implicit reward score is used as the weighting factor of the sample in the ranking task. High reward samples are given higher weights to guide the model to prioritize learning high-value ranking patterns.

[0043] S402: Fine-tuning and updating the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model using the implicit reward signal and explicit labeled data, including: The implicit reward signal is used to construct a ranking loss function to optimize the model's ability to rank documents or candidates based on relevance. The explicit labeled data is used to construct a regression loss function to optimize the model's accuracy in identifying entity attributes and determining authenticity. Specifically, this invention employs a multi-task collaborative learning framework to unify the two types of heterogeneous signals into the model's optimization objective. For the implicit reward signal, the system constructs a ranking loss function (such as Listwise Loss or ApproxNDCG Loss), which focuses on the overall ranking quality of the candidate list rather than the absolute score of a single point. High-reward samples are given higher weights in the loss calculation, guiding the model to adjust parameters so that high-value results rank higher in the vector space, thereby directly improving the user's search experience and filtering efficiency. For the explicit labeled data, the system constructs a regression loss function or a cross-entropy classification loss function. For example, for a consistency verification model in expense reimbursement verification, the classification loss is calculated using the rejection reason label, and penalty weights are added for key risk categories such as "logical conflict" to solve the problem of imbalance between positive and negative samples, forcing the model to improve the recall rate of abnormal invoices. This composite loss design allows the model to learn "what users like" through massive implicit behaviors, while also safeguarding "what is business correct" through scarce explicit labels, thus achieving a dual guarantee of user experience and compliance bottom line.

[0044] Based on the ranking and regression loss functions, the gradient is calculated using the backpropagation algorithm. This allows for small-batch fine-tuning of the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model in the AI ​​engine layer. A replay mechanism is employed to prevent catastrophic forgetting, enabling continuous iteration of model performance. Specifically, to adapt to the characteristics of enterprise data streaming, this invention abandons the traditional full-scale retraining mode and instead adopts a small-batch online fine-tuning strategy. Whenever a certain number of new feedback samples are accumulated, the system triggers a gradient update, enabling the model to quickly absorb new knowledge. However, the biggest challenge of online fine-tuning is catastrophic forgetting, where the model may overwrite important old knowledge when learning new data distributions. Therefore, this embodiment introduces an experience replay mechanism. The system maintains a fixed-capacity experience replay buffer, persistently storing key samples from each modality during historical training and their corresponding explicit or implicit signals. Each time a gradient update is performed based on new data, the system randomly selects a certain proportion of historical samples from the buffer and mixes them with the current new samples to form a training batch. This hybrid sampling strategy is mathematically equivalent to applying a regularization constraint to the gradient update of new data, limiting the deviation of model parameters from historical optimal solutions, thus adapting to new changes while retaining the memory of old knowledge. For example, when a company recently processes a new type of invoice, causing the model to overfit to that feature, the historical regular invoice samples stored in the buffer will pull the model parameters back to the general representation space, preventing the ability to recognize older invoice types from deteriorating. Through this "new and old fusion" training paradigm, this invention successfully constructs an AI platform with long-term evolution capabilities, enabling it to maintain high performance and stability over an operational cycle of several years, truly achieving a resonance between artificial intelligence technology and enterprise business growth.

[0045] It should be noted that the replay mechanism includes: maintaining a fixed-capacity experience replay buffer to store key samples from each modality and their explicit / implicit signals during historical training; and randomly selecting a certain proportion of historical samples from the buffer to add to the current training batch each time a mini-batch gradient update is performed based on new data, in order to constrain the magnitude of model parameter shifts and prevent the forgetting of old knowledge.

[0046] It should be noted that the consistency verification model adopts a dual-tower neural network structure; the input end receives the ROI region feature vectors of the invoice image, including the amount, date, seal area, and structured text field vectors extracted by OCR; the output end outputs a confidence score, representing the logical consistency probability of the image and text information; the training labels are: the 'rejection reason codes' in the explicitly labeled data are mapped to binary labels (consistent / inconsistent) and multi-class labels (inconsistent amount, incorrect date, missing seal); the loss function is: the cross-entropy loss function is used, and a penalty weight is added for the 'logical conflict' category to solve the sample imbalance problem.

[0047] The above primarily describes the solutions of the embodiments of this application from the perspective of device implementation. It is understood that each device, such as an AI-based enterprise middleware management system, includes at least one of the hardware structures and software modules corresponding to the execution of each function in order to achieve the above-mentioned functions. Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0048] This application embodiment can divide the AI-based enterprise middleware management system into functional units according to the above method examples. For example, each function can be divided into different functional units, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division; other division methods may be used in actual implementation.

[0049] When using integrated units, Figure 2 The diagram illustrates a possible structure of the AI-based enterprise middleware management system (referred to as communication system 30) involved in the above embodiments. The communication system 30 includes a processing unit 301 and a communication unit 302, and may also include a storage medium (referred to as storage unit 303). Figure 2 The structural diagram shown can be used to illustrate the structure of the AI-based enterprise middleware management system involved in the above embodiments.

[0050] when Figure 2 The structural diagram shown is used to illustrate the structure of the AI-based enterprise middleware management system involved in the above embodiments. The processing unit 301 is used to control and manage the actions of the AI-based enterprise middleware management system, the communication unit 302 is used for the AI-based enterprise middleware management system to communicate with other devices, and the storage unit 303 is used to store the program code and data of the AI-based enterprise middleware management system.

[0051] For example, communication unit 302 is used to acquire raw data from across departments within an enterprise and process the raw data to obtain a preprocessed dataset; wherein the raw data includes structured business data from a business database and unstructured document data from a file system; Processing unit 301: used to extract entity information and relationships between entities from the preprocessed dataset, and to construct a dynamic enterprise knowledge graph containing project, personnel, and financial dimensions; The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. Through a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library; It receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow; among them, business requests include project progress query, intelligent recruitment screening, intelligent expense verification, and intelligent document retrieval; The system collects user interaction behavior data and final business decision results at the application layer; transforms the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; and uses the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the self-evolution of the model. The interaction behavior data includes click locations, page dwell time, download operations, and manual correction records; the final business decision results include the hiring / rejection status and rejection reasons in the recruitment process, and the approval / rejection status and rejection reason codes in the expense reimbursement process.

[0052] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.

[0053] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely illustrative descriptions of the application as defined by the appended claims, and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from the spirit and scope of this application. Thus, if such modifications and modifications fall within the scope of the claims of this application and their equivalents, this application is also intended to include such modifications and modifications.

[0054] Some of the data in the above calculation formula are obtained by removing dimensions and taking their numerical values. The calculation formula is a calculation formula that is closest to the real situation, obtained by software simulation of a large amount of collected data. The preset parameters and preset thresholds in the calculation formula are set by those skilled in the art according to the actual situation or obtained through simulation of a large amount of data.

Claims

1. An enterprise middle platform management method based on artificial intelligence, characterized in that: include: S1: Obtain raw data from across departments within the enterprise, and process the raw data to obtain a preprocessed dataset; wherein, the raw data includes structured business data from the business database and unstructured document data from the file system; S2: Extract entity information and relationships between entities from the preprocessed dataset, and construct a dynamic enterprise knowledge graph that includes project, personnel, and financial dimensions; The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. Through a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library; S3: Receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow; among which, business requests include project progress query, intelligent recruitment screening, intelligent expense verification, and intelligent document retrieval; S4: Collect user interaction behavior data and final business decision results from S3 at the application layer; convert the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; use the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the self-evolution of the model; wherein, the interaction behavior data includes click location, page dwell time, download operation, and manual correction records, and the final business decision results include the hiring / elimination status and elimination reason in the recruitment process, and the approval / rejection status and rejection reason code in the expense reimbursement process.

2. The enterprise middle platform management method based on artificial intelligence according to claim 1, characterized in that, The process of processing the original data to obtain the preprocessed dataset includes: The structured business data is cleaned and formatted using an ETL engine to generate standardized structured data; the unstructured document data is denoised and extracted using optical character recognition technology and natural language preprocessing technology to generate standardized text data; a unified data dictionary is established to map the standardized structured data and the standardized text data to a unified data identification system to obtain a preprocessed dataset.

3. The enterprise middle platform management method based on artificial intelligence according to claim 1, characterized in that, The step of extracting entity information and relationships between entities from the preprocessed dataset and constructing a dynamic enterprise knowledge graph including project, personnel, and financial dimensions includes: Predefine an ontology containing four core entities: projects, personnel, finance, and documents, and their interrelationships, and establish the attribute fields and data type specifications for each entity; For structured data, entity nodes are instantiated directly according to field mapping rules and assigned a globally unique identifier; for unstructured text data, potential entities are extracted using a BERT-based named entity recognition model, and entity disambiguation and linking are performed by calculating the cosine similarity between the entity feature vector and the existing entity vector in the knowledge base to determine the entity identity. For structured data, relation edges between entities are directly generated based on foreign key associations and business logic rules; for unstructured text, a relation extraction model trained by remote supervised learning is used to identify semantic relationships between entities from the sentence context and generate corresponding relation edges. A timestamp attribute is introduced for each node attribute and relation edge. When a data change is detected, historical version records are retained and the effective time interval of the new version is marked, thereby forming a dynamic knowledge graph. Store the aforementioned nodes, relationships, and time-series attributes into a graph database, and establish a hash index based on entity ID and a full-text index based on text content.

4. The enterprise middle platform management method based on artificial intelligence according to claim 1, characterized in that, The step of using a multimodal embedding model to encode features of unstructured document data in the preprocessed dataset, generating high-dimensional semantic vectors, and storing the high-dimensional semantic vectors in a multimodal vector library includes: The document image is divided into several image blocks and the coordinate information is extracted. At the same time, the document text is segmented. The spatial mapping relationship between text fragments and corresponding image regions is established to form a multimodal input sequence. Visual feature vectors of image patches are extracted using a visual Transformer model to capture the document’s layout, seals, and table structure information; semantic feature vectors of text are extracted using a pre-trained language model to capture the document’s semantic content. The visual feature vector is concatenated with the text feature vector, and modality type embedding and position embedding are added; the input is fed into a multimodal Transformer network, and the feature interaction and fusion of text semantics and visual layout are realized through the cross attention mechanism to generate a context feature sequence containing comprehensive text and image information; The attention pooling mechanism is used to weight and aggregate the context feature sequences to obtain a fixed-dimensional vector representing the semantics of the entire document. The fixed-dimensional vector is subjected to L2 normalization to generate the final high-dimensional semantic vector, which is then associated with the unique document identifier and stored in the multimodal vector library.

5. The enterprise middleware management method based on artificial intelligence according to claim 1, characterized in that, The process of associating and mapping nodes in the dynamic enterprise knowledge graph with vectors in the multimodal vector library through a unified data identification system includes: Based on the preset globally unique identifier generation rules, a unique GUID is assigned to each structured business entity and unstructured document, and a mapping registry from the source system's original ID to the GUID is established. When constructing a dynamic enterprise knowledge graph, the GUID is written as a core attribute into the corresponding node in the graph database, and a unique index is established for the core attribute, so that the graph node can be uniquely addressed by the GUID. When storing the high-dimensional semantic vector of an unstructured document into a multimodal vector library, the corresponding document GUID is stored together with the vector as scalar metadata, and a filtering index is established for the scalar metadata field. Analyze the reference edges between entities and documents in the knowledge graph, extract the pairing relationship between the start node GUID and the end node GUID, and construct a bidirectional mapping index of entity GUID-document GUID; During the data update process, a transaction mechanism is used to synchronously update the graph node attributes, vector library metadata, and bidirectional mapping index.

6. The enterprise middleware management method based on artificial intelligence according to claim 1, characterized in that, The process of receiving business requests initiated by users through the application layer and invoking the AI ​​engine layer to execute the processing flow includes: When the request is for a project progress query: the target project number in the request is parsed, and a breadth-first search algorithm is executed in the dynamic enterprise knowledge graph with the target project number as the starting node to traverse the associated financial payment nodes, project progress nodes, and human resource input nodes to obtain the status attribute values ​​of each node; the status attribute values ​​are aggregated using a graph neural network to calculate the project's comprehensive health score and risk warning index, and the comprehensive health score and risk warning index are returned to the application layer for display; When the request is for intelligent recruitment screening: The uploaded candidate resume file and preset job description text are retrieved; the candidate resume file is encoded using a multimodal embedding model to generate a resume feature vector; the job description text is encoded using a natural language processing model to generate a job requirement vector; the cosine similarity between the resume feature vector and the job requirement vector in the multimodal vector library is calculated to obtain a matching score; candidate information with a matching score higher than a preset threshold is used as the screening result and synchronized to the human resources module. When the request is for intelligent expense reimbursement verification: The uploaded invoice image file is retrieved; the layout and seal texture features of the invoice image file are extracted using a computer vision model, and the invoice text information is extracted using optical character recognition (OCR) technology; the invoice text information is compared with the authenticity verification result returned by the official tax interface, and the layout and seal texture features are input into a consistency verification model to determine if there is a logical conflict between the image features and the text information; if the authenticity verification passes and there is no logical conflict, a standardized expense reimbursement form is automatically generated and pushed to the financial approval process; otherwise, an anomaly marker is generated and the application layer is returned to prompt manual intervention. When the request is for intelligent document retrieval: obtain the natural language retrieval statement input by the user; use a semantic retrieval model to convert the natural language retrieval statement into a query vector; calculate the similarity between the query vector and the stored high-dimensional semantic vector in the multimodal vector library, and recall the top-ranked document IDs; read the corresponding original document from the file system according to the document ID, and return the original document to the application layer for display.

7. The enterprise middle platform management method based on artificial intelligence according to claim 1, characterized in that, The process of converting the interactive behavior data into implicit reward signals and converting the final business decision results into explicit labeled data includes: Collect user behavior logs during the search, recruitment, and expense reimbursement processes; calculate the above behavior logs into numerical implicit reward scores based on a preset behavior weight matrix, and perform time decay weighting and normalization processing to obtain implicit reward signals; wherein, the behavior logs include click locations, page dwell time, download operations, and manual correction records; Collect the decision results at the end of the business process, construct positive sample pairs between the candidates to be hired and the job descriptions, construct negative sample pairs between the candidates to be rejected and the job descriptions, and map the reasons for rejection to standard error type labels to form explicit labeled data.

8. The enterprise middle platform management method based on artificial intelligence according to claim 1, characterized in that, The step of fine-tuning and updating the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model using the implicit reward signal and explicit labeled data includes: The implicit reward signal is used to construct a ranking loss function to optimize the model's ability to rank documents or candidates based on their relevance; the explicit labeled data is used to construct a regression loss function to optimize the model's accuracy in identifying entity attributes and judging authenticity. Based on the ranking loss function and regression loss function, the gradient is calculated using the backpropagation algorithm, and the parameters of the multimodal embedding model, semantic retrieval model and consistency verification model in the AI ​​engine layer are fine-tuned in small batches to achieve continuous iteration of model performance.

9. An AI-based enterprise middleware management system, used to run the AI-based enterprise middleware management method according to any one of claims 1 to 8, characterized in that, include: Communication unit and processing unit; The communication unit is used to acquire raw data from across departments within the enterprise and process the raw data to obtain a preprocessed dataset; wherein the raw data includes structured business data from the business database and unstructured document data from the file system; The processing unit is used to extract entity information and relationships between entities from the preprocessed dataset, and to construct a dynamic enterprise knowledge graph that includes project, personnel, and financial dimensions. The unstructured document data in the preprocessed dataset is feature-encoded using a multimodal embedding model to generate high-dimensional semantic vectors, which are then stored in a multimodal vector library. Through a unified data identification system, the nodes in the dynamic enterprise knowledge graph are associated and mapped with the vectors in the multimodal vector library; It receives business requests initiated by users through the application layer and calls the AI ​​engine layer to execute the processing flow; among them, business requests include project progress query, intelligent recruitment screening, intelligent expense verification, and intelligent document retrieval; The system collects user interaction behavior data and final business decision results at the application layer; transforms the interaction behavior data into implicit reward signals and the final business decision results into explicit labeled data; and uses the implicit reward signals and explicit labeled data to fine-tune and update the parameters of the multimodal embedding model, semantic retrieval model, and consistency verification model to complete the self-evolution of the model. The interaction behavior data includes click locations, page dwell time, download operations, and manual correction records; the final business decision results include the hiring / rejection status and rejection reasons in the recruitment process, and the approval / rejection status and rejection reason codes in the expense reimbursement process.

10. A storage medium storing instructions that, when executed on an AI-based enterprise middleware management system, cause the AI-based enterprise middleware management system to perform the AI-based enterprise middleware management method according to any one of claims 1 to 8.

Images