Enterprise data asset entry system based on artificial intelligence

Abstract

The invention discloses an enterprise data asset entry system based on artificial intelligence, and the system comprises an original document receiving module which is used for generating an original document set; the multi-modal data construction module is used for constructing a multi-modal input data set; the artificial intelligence identification module is used for identifying a field set with data asset attributes; the structured asset modeling module is used for converting the identification field set into a structured data asset object; the rule matching and template selection module is used for generating a field mapping rule and table entry trigger configuration; the field standardization mapping module is used for generating a structured data entry record set; the table entry metadata supplementing module is used for adding a template number, a mapping timestamp and an operator identifier for each data table entry record; the ledger writing and auditing module is used for writing the data entry record into the enterprise asset ledger system; and the table entering state management module is used for generating a state mapping table. According to the invention, automatic data asset entry is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of enterprise data management, and in particular to an enterprise data asset entry system based on artificial intelligence. Background Art

[0002] In the current digital transformation of enterprises, data assets, as a crucial component of core production factors, are gradually evolving from decentralized storage and extensive management to structured, auditable, and operationalized systems. Enterprises generate a large number of business documents in their daily operations. These documents come from a wide range of sources, including contract texts, financial statements, business approval materials, project records, and system-derived data snapshots, typically in PDF, Word, Excel, image, or scanned formats. Traditional data asset management methods rely heavily on manual organization and archiving, requiring managers to review document content one by one, identify key fields, and enter them into asset ledgers or management systems. This approach is inefficient, error-prone, and difficult to adapt to the demands of processing surging data volumes.

[0003] With the development of natural language processing and document intelligent parsing technology, some companies have begun to introduce OCR recognition tools, structured extraction components or rule-based parsing systems to extract keywords and field information from documents. Although these tools can reduce the intensity of manual intervention to a certain extent, their accuracy and adaptability are still limited by multiple factors. On the one hand, traditional OCR technology has difficulty processing document types that contain complex layout structures, nested tables or seal interference; on the other hand, structured extraction solutions based on fixed rules or templates are highly dependent on document formats, and when faced with non-standard formats or files generated across systems, accurate matching is often not possible. In addition, in the process of field recognition, existing solutions often find it difficult to comprehensively consider the contextual semantics, spatial position relationships and contextual associations of fields, resulting in missed detections or misjudgments in the recognition results of data asset attributes.

[0004] In addition, in the structured modeling stage after field extraction, existing methods generally lack intelligent aggregation and semantic association capabilities, and are unable to effectively identify the logical attribution relationship between multiple scattered fields, thus affecting the accurate construction of data asset objects. The table entry process also mostly uses manual selection of templates and field mapping methods, which cannot automatically select the optimal table entry template based on field content, format and context, and complete standardized field conversion. This process that relies on manual judgment is not only inefficient, but also difficult to ensure the integrity, consistency and audit traceability of asset management. In the data record writing link, although some systems support logging functions, they lack a unified metadata completion mechanism and status management system, resulting in a lack of effective support for subsequent record auditing, version traceability and exception handling.

[0005] Therefore, how to provide an enterprise data asset entry system based on artificial intelligence is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0006] One purpose of the present invention is to propose an enterprise data asset entry system based on artificial intelligence. The present invention combines multimodal document parsing technology with the LayoutLMv3 visual document pre-training model, and describes in detail the entire process from automatic recognition of original business documents, structured modeling to rule-driven entry. It has the advantages of high recognition accuracy, adaptability to multiple document types, high data entry efficiency and strong audit traceability.

[0007] An enterprise data asset entry system based on artificial intelligence according to an embodiment of the present invention includes:

[0008] The original document receiving module is used to receive data source files from multiple systems within the enterprise and generate an original document collection;

[0009] A multimodal data construction module is used to perform text extraction, layout coordinate analysis and image data preprocessing on the original document collection to construct a multimodal input data collection;

[0010] The AI ​​recognition module is used to input a multimodal input data set into the LayoutLMv3 model fine-tuned with the enterprise document corpus, perform image and text semantic modeling, and identify the set of fields with data asset attributes;

[0011] A structured asset modeling module, used to convert a set of identification fields into a structured data asset object;

[0012] The rule matching and template selection module is used to perform feature matching calculations on structured data asset objects, select the optimal table entry template, and generate field mapping rules and table entry trigger configurations;

[0013] The field standardization mapping module is used to convert values, unify formats, and complete missing fields according to field mapping rules to generate a structured data entry record set;

[0014] The table entry metadata supplement module is used to add a template number, mapping timestamp, and operator identification to each data entry record;

[0015] The ledger writing and auditing module is used to write data entry records into the enterprise asset ledger system and simultaneously generate entry logs and field mapping audit information;

[0016] The entry status management module is used to confirm the writing status of each data record and generate a status mapping table.

[0017] Optionally, modules can be connected using the following methods:

[0018] S1. Receive original business documents from multiple data sources within the enterprise and generate an original document collection;

[0019] S2. Perform preprocessing operations on the original document set to extract text content, layout coordinate information, and image pixel information to construct a multimodal input data set;

[0020] S3. Input the multimodal input data set into the pre-trained model LayoutLMv3 based on visual document understanding. The LayoutLMv3 model is fine-tuned and trained on the enterprise data document corpus to perform joint image and text modeling and generate a data asset field set.

[0021] S4. Construct a structured data asset object set based on the data asset field set;

[0022] S5. Input the structured data asset object set into a preset rule matching engine, perform a matching operation based on the attribute information of each structured data asset object, and determine an entry template set;

[0023] S6. Based on the table entry template set, convert the field contents in the structured data asset object set into a standardized form field format to generate a data entry record set;

[0024] S7. Write the data entry record set into the asset ledger system, recording the execution time of the entry operation, the operator ID, and the template ID used.

[0025] Optionally, the original document collection includes contract documents in PDF format, business reports in Word format, financial ledgers in Excel format, scanned reports in image format, archive scans in TIFF format, and structured document snapshots exported through an office system.

[0026] Optionally, the S2 specifically includes:

[0027] S21. Performing structured parsing processing on the PDF, Word, and Excel documents in the original document collection, extracting paragraph text, table fields, and form structures, and numbering them sequentially according to the document pages to construct a preliminary text content set and a structure tag set;

[0028] S22, performing OCR recognition processing on the image format documents and TIFF format documents in the original document set, using an optical character recognition engine to extract valid text information in the image, recording the position range of each text block in combination with the image position information, and generating an OCR text set and a position coordinate set;

[0029] S23, establishing a binding relationship between each text unit in the preliminary text content set and the OCR text set and a corresponding two-dimensional spatial position, wherein the two-dimensional spatial position is represented as a rectangular box coordinate;

[0030] S24. Perform standardized preprocessing operations on the image format document, including unifying the image resolution, image size, color channel arrangement and compression format, generating an image pixel matrix set, and pairing the image pixel matrix set with the text content and spatial coordinate information to construct a multimodal input data set as input data for a subsequent recognition model.

[0031] Optionally, the S3 specifically includes:

[0032] S31, uniformly encode the text content, the corresponding two-dimensional space coordinate position and the image pixel matrix set in the multimodal input data set, where the input of each text block is represented as a triple (T i ,B i ,I), where T i Represents the Token sequence of the i-th text block, B i represents the corresponding rectangular coordinate frame, and I represents the image pixel matrix of the corresponding page;

[0033] S32. Input each component element of the triple into the embedding layer of the LayoutLMv3 model respectively, and perform an embedding extraction operation on the LayoutLMv3 model after fine-tuning and training the model with the enterprise data document corpus to obtain a text embedding vector, a layout position embedding vector, and a visual feature embedding vector;

[0034] S33. Perform weighted fusion on the above three types of embedding vectors and generate the fused multimodal feature vector through the Gaussian error linear unit activation function:

[0035]

[0036] in, is the fused multimodal input feature vector corresponding to the i-th text block, is the text embedding vector of the i-th text block, is the layout position embedding vector of the i-th text block, is the visual feature embedding vector of the i-th text block, W t W is the weight matrix that projects the text embedding vector into the joint space. l is the weight matrix that projects the layout embedding vector into the joint space, W v is the weight matrix that projects the visual embedding vector into the joint space, b is the bias term in the multimodal fusion linear combination process, and GELU(·) is the Gaussian error linear unit activation function:

[0037]

[0038] Among them, x is the intermediate vector after the fusion linear transformation, that is, tanh(·) is the hyperbolic tangent function;

[0039] S34. Input the fused multimodal feature vector sequence into the Transformer encoding structure of the LayoutLMv3 model, perform multi-head self-attention mechanism and feedforward network calculation, and generate a context embedding vector set;

[0040] S35. Input the context embedding vector set into the classifier head structure of the LayoutLMv3 model, perform sequence labeling based on the Softmax function, and generate a data asset field set. Each field in the data asset field set includes field content, field type label, context window fragment, and original coordinate frame position for subsequent structured modeling and template matching.

[0041] Optionally, the S32 specifically includes:

[0042] S321, the text Token sequence T in the triple i Input to the text embedding sublayer to generate the corresponding text embedding vector The embedding process is based on a word vector mapping function fine-tuned from the enterprise corpus. The enterprise data document corpus includes historical contracts, financial ledgers, business reports, meeting minutes, audit records, and archive scans, covering the semantics and format structures commonly used in enterprise data asset management.

[0043] S322, the rectangular coordinate frame B in the triplet i Input to the layout embedding sublayer to generate the corresponding layout position embedding vector The position embedding function is mapped into a set of high-dimensional space coordinate embeddings;

[0044] S323: Input the image pixel matrix I into the visual embedding sublayer, first divide the image pixel matrix I into multiple fixed-size image patch sets, and then input it into the visual convolutional network to extract the visual feature embedding vector

[0045] S324, embed the text obtained above into a vector Layout position embedding vector and visual feature embedding vectors Keep the dimensions aligned and output to the multimodal fusion sublayer to generate the fused multimodal input feature vector.

[0046] Optionally, the S4 specifically includes:

[0047] S41. Parse each field element in the data asset field set, extract the field content, field type label, context embedding vector, and original coordinate frame information, and construct a field structure set. Each field structure is represented as a four-tuple.

[0048] S42. Perform a field aggregation operation on the field structure set, using an aggregation function based on semantic similarity and spatial proximity to aggregate multiple field structures belonging to the same asset entity into an asset object unit:

[0049] O j =Agg({F i |Sim(H i ,H k )>θ1∧Dist(B i ,B k )<θ2});

[0050] Among them, O j is the jth asset object unit, Agg(·) is the field aggregation function, F i For the i-th field structure, Sim(H i ,H k ) is the semantic similarity function, H i is the context embedding vector of the i-th field, H k is the context embedding vector of the k-th candidate field, θ1 is the semantic similarity threshold, Dist(B i ,B k ) is the spatial distance function, B i is the rectangular position box of the i-th field in the original document, B k is the rectangular location box of the k-th candidate field, θ2 is the spatial distance threshold, i is the index number of the current candidate field structure, j is the number of the asset object unit generated after aggregation, and k is the number of the center field structure currently being used as a reference;

[0051] S43, for each asset object unit O j Perform field standardization and map the subordinate field contents into a standardized field set {N1, N2, ..., N k ,…,N p}, where each standardized field N k Corresponding to one of the asset name, department, person in charge, source system, or field value. If the original field is missing, it is set to "null value" and a missing tag is recorded;

[0052] S44. Output all standardized field sets as a structured data asset object set, with each object accompanied by a corresponding original coordinate frame set for subsequent table entry template matching and audit tracking.

[0053] Optionally, the S5 specifically includes:

[0054] S51. Receive a standardized field set in a structured data asset object set as a matching input set, wherein the standardized field content includes asset name, department, responsible person, source system, field value, and field type label;

[0055] S52: Input the matching input set into a rule matching engine, which includes a field attribute parsing module, a template matching scoring module, and a trigger condition judgment module;

[0056] S53, the field attribute parsing module performs a k Constructing matching vectors The matching vector is composed of the field semantic label, historical matching frequency and location coordinate encoding to generate a matching vector sequence;

[0057] S54. The template matching scoring module performs scoring matching on the matching vector sequence and the field mapping rules in the preset table entry template set in sequence:

[0058]

[0059] Among them, S j,m is the matching score, k is the field number of the matching input field in the current object, p is the number of standardized fields extracted from the current structured data asset object, α is the weighted coefficient of the field semantic label similarity, Sim(·,·) is the similarity scoring function, and l k Encode the semantic label of the kth field, For table entry template T m The standard semantic label encoding of the kth field in , β is the weighted coefficient of the field’s historical frequency similarity, f k is the matching frequency statistics of the kth field in the historical table entry template, For table entry template T m The reference frequency value of the kth field in , γ is the weighted coefficient of the field position coordinate similarity, p k Encode the two-dimensional position coordinates of the k-th field in the original business document, For table entry template T m The standard position vector of the kth field in ;

[0060] S55, the trigger condition judgment module will match the score S j,m With threshold θ T Compare, when the score is greater than or equal to the threshold, select the corresponding template T mAs a template candidate for the current asset object, the field mapping rules, form format configuration, and table entry trigger conditions are output to form a table entry template set.

[0061] Optionally, the S6 specifically includes:

[0062] S61, read the field mapping rules in the table template set, for each asset object unit O j The standardized field set is determined in the table template T m The corresponding standard form field set, record field mapping relationship pair (N k ,F k );

[0063] S62. Standardize the original field values ​​according to the field mapping rules, including unified encoding format, unit conversion, and null value filling strategy, execute the conversion function on each field, and generate form field values;

[0064] S63, combine all mapped field pairs and converted form field values ​​to form a data entry record object R j ;

[0065] S64, record object R in the table j The data entry table metadata is appended to the template, including the template number, field mapping timestamp, and mapping operator ID, to generate a data entry table record set.

[0066] Optionally, the S7 specifically includes:

[0067] S71. Each structured entry record object R in the data entry record set j Write to the asset ledger system. The writing process includes field docking, form positioning and data submission to complete the asset record registration operation;

[0068] S72. During the writing process, a table entry operation log is synchronously generated. The log content includes the template number, table entry record number, operation timestamp, operator ID, and target ledger form ID, forming a table entry operation log set.

[0069] S73. Generate an audit trail data set based on the field mapping information, template number, and structured field source during the writing process, to describe the mapping path between the field source and the target location;

[0070] S74. The entry log set and audit trail data set are transmitted back to the data governance platform or model training module as data support for subsequent model fine-tuning, rule engine modification, and abnormal audit analysis.

[0071] S75, confirm the status of each data entry record and generate a status indicator S j∈ success, failure}, and is bound to the entry record number to form a write status mapping table for feedback control and subsequent fault-tolerant processing.

[0072] The beneficial effects of the present invention are:

[0073] This paper constructs an AI-based enterprise data asset entry system, achieving a complete automated process from receiving original business documents, constructing multimodal data, intelligently identifying and modeling structured data, to rule matching, standardized mapping, and account entry. The system utilizes the LayoutLMv3 visual document understanding model, fine-tuned and trained on a corpus of enterprise data documents, enabling it to more accurately identify field information related to data assets. It also leverages the combined features of text semantics, layout, and image content to enhance recognition stability and adaptability in complex document scenarios.

[0074] During the structured modeling process after field identification, this invention enhances data organization and relevance by constructing a standardized set of objects containing asset name, ownership, responsible person, source system, and field location information, facilitating subsequent unified management. For table entry rule matching, this invention introduces a scoring mechanism based on field content, spatial location, and semantic similarity. This automatically selects the most appropriate table entry template and mapping configuration, replacing traditional manual template selection, reducing human intervention, and improving table entry accuracy and efficiency.

[0075] In addition, the present invention integrates format conversion, unit unification, and missing completion in the field standardization process to ensure the consistency of written fields. During the data record writing phase, the system supports the automatic supplementation of operation information, template identifiers, and timestamps, realizing a closed-loop audit of the data asset entry process, facilitating future tracing and responsibility division. The overall system design improves the automation level of data asset identification, modeling, and entry in large-scale documents, alleviating the problems of large manpower investment, strong reliance on field identification, and insufficient audit coverage in enterprise data management. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a flow chart of a method for entering enterprise data assets into a table system based on artificial intelligence proposed by the present invention;

[0078] Figure 2 This is a system diagram of an enterprise data asset entry system based on artificial intelligence proposed by the present invention;

[0079] Figure 3This is a structural diagram of the graphic and text semantic joint modeling based on the LayoutLMv3 model in the artificial intelligence-based enterprise data asset entry system proposed by the present invention. DETAILED DESCRIPTION

[0080] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0081] refer to Figure 1-3 , an enterprise data asset entry system based on artificial intelligence, including:

[0082] The original document receiving module is used to receive data source files from multiple systems within the enterprise and generate an original document collection;

[0083] A multimodal data construction module is used to perform text extraction, layout coordinate analysis and image data preprocessing on the original document collection to construct a multimodal input data collection;

[0084] The AI ​​recognition module is used to input a multimodal input data set into the LayoutLMv3 model fine-tuned with the enterprise document corpus, perform image and text semantic modeling, and identify the set of fields with data asset attributes;

[0085] A structured asset modeling module, used to convert a set of identification fields into a structured data asset object;

[0086] The rule matching and template selection module is used to perform feature matching calculations on structured data asset objects, select the optimal table entry template, and generate field mapping rules and table entry trigger configurations;

[0087] The field standardization mapping module is used to convert values, unify formats, and complete missing fields according to field mapping rules to generate a structured data entry record set;

[0088] The table entry metadata supplement module is used to add a template number, mapping timestamp, and operator identification to each data entry record;

[0089] The ledger writing and auditing module is used to write data entry records into the enterprise asset ledger system and simultaneously generate entry logs and field mapping audit information;

[0090] The entry status management module is used to confirm the writing status of each data record and generate a status mapping table.

[0091] The present invention realizes the automatic entry of enterprise multi-source business documents into structured data assets by constructing an enterprise data asset entry system based on artificial intelligence. In the document parsing stage, the accuracy and stability of field recognition are significantly improved by combining text content, layout information and image features. In the field modeling and rule matching link, an artificial intelligence-based graphic and text semantic joint modeling method is introduced to enhance the contextual understanding ability of asset fields and optimize the efficiency of template selection and field mapping. Through field standardization processing and metadata supplementation operations, the format consistency and audit traceability of the entry records are ensured. At the same time, the system supports status tracking and write confirmation, which improves the controllability and management transparency of the data entry process. After testing, the system has significantly improved in processing efficiency, field recognition rate, template matching accuracy and audit integrity compared with manual methods, and has good engineering practicality and promotion value.

[0092] In this embodiment, the modules are connected through the following methods:

[0093] S1. Receive original business documents from multiple data sources within the enterprise and generate an original document collection;

[0094] S2. Perform preprocessing operations on the original document set to extract text content, layout coordinate information, and image pixel information to construct a multimodal input data set;

[0095] S3. Input the multimodal input data set into the pre-trained model LayoutLMv3 based on visual document understanding. The LayoutLMv3 model is fine-tuned and trained on the enterprise data document corpus to perform joint image and text modeling and generate a data asset field set.

[0096] S4. Construct a structured data asset object set based on the data asset field set;

[0097] S5. Input the structured data asset object set into a preset rule matching engine, perform a matching operation based on the attribute information of each structured data asset object, and determine an entry template set;

[0098] S6. Based on the table entry template set, convert the field contents in the structured data asset object set into a standardized form field format to generate a data entry record set;

[0099] S7. Write the data entry record set into the asset ledger system, recording the execution time of the entry operation, the operator ID, and the template ID used.

[0100] The present invention proposes an artificial intelligence-based enterprise data asset entry method, which realizes the automation of the entire process from original document parsing to structured entry into the table through a multi-step process. By jointly modeling text, layout and image information, the recognition accuracy of asset fields in complex business documents is improved. Combined with the LayoutLMv3 model that has been fine-tuned and trained with the enterprise document corpus, the adaptability to enterprise terminology and document structure is enhanced. In the process of rule matching and template selection, accurate mapping between field content and template configuration is achieved, reducing the frequency of manual intervention. Ultimately, through data standardization processing and write confirmation mechanism, the integrity, timeliness and traceability of asset ledger data are ensured. Experiments have shown that this method is superior to traditional manual methods in terms of field extraction accuracy, template matching efficiency and data entry success rate, and has strong engineering feasibility and promotion prospects.

[0101] In this embodiment, the original document collection includes contract documents in PDF format, business reports in Word format, financial ledgers in Excel format, scanned reports in image format, scanned archives in TIFF format, and structured document snapshots exported through the office system.

[0102] The present invention supports unified access and processing of original business documents in multiple formats, including contracts, reports, financial ledgers, scanned reports and structured snapshots, improving the system's compatibility and parsing capabilities for heterogeneous enterprise data sources, and providing a reliable foundation for subsequent data asset identification and entry into tables.

[0103] In this embodiment, S2 specifically includes:

[0104] S21. Performing structured parsing processing on the PDF, Word, and Excel documents in the original document collection, extracting paragraph text, table fields, and form structures, and numbering them sequentially according to the document pages to construct a preliminary text content set and a structure tag set;

[0105] S22, performing OCR recognition processing on the image format documents and TIFF format documents in the original document set, using an optical character recognition engine to extract valid text information in the image, recording the position range of each text block in combination with the image position information, and generating an OCR text set and a position coordinate set;

[0106] S23, establishing a binding relationship between each text unit in the preliminary text content set and the OCR text set and a corresponding two-dimensional spatial position, wherein the two-dimensional spatial position is represented as a rectangular box coordinate;

[0107] S24. Perform standardized preprocessing operations on the image format document, including unifying the image resolution, image size, color channel arrangement and compression format, generating an image pixel matrix set, and pairing the image pixel matrix set with the text content and spatial coordinate information to construct a multimodal input data set as input data for a subsequent recognition model.

[0108] This method performs structured parsing and OCR on documents in various formats, extracting paragraph text, table fields, and spatial location information. Combined with image preprocessing techniques, it constructs a multimodal input data set, effectively integrating text, structure, and image features. This approach improves the recognition coverage and spatial understanding of asset fields in unstructured and semi-structured documents, providing a high-quality, uniformly formatted data foundation for subsequent model input.

[0109] In this embodiment, S3 specifically includes:

[0110] S31, uniformly encode the text content, the corresponding two-dimensional space coordinate position and the image pixel matrix set in the multimodal input data set, where the input of each text block is represented as a triple (T i ,B i ,I), where T i Represents the Token sequence of the i-th text block, B i represents the corresponding rectangular coordinate frame, and I represents the image pixel matrix of the corresponding page;

[0111] S32. Input each component element of the triple into the embedding layer of the LayoutLMv3 model respectively, and perform an embedding extraction operation on the LayoutLMv3 model after fine-tuning and training the model with the enterprise data document corpus to obtain a text embedding vector, a layout position embedding vector, and a visual feature embedding vector;

[0112] S33. Perform weighted fusion on the above three types of embedding vectors and generate the fused multimodal feature vector through the Gaussian error linear unit activation function:

[0113]

[0114] in, is the fused multimodal input feature vector corresponding to the i-th text block, is the text embedding vector of the i-th text block, is the layout position embedding vector of the i-th text block, is the visual feature embedding vector of the i-th text block, W t W is the weight matrix that projects the text embedding vector into the joint space. l is the weight matrix that projects the layout embedding vector into the joint space, Wv is the weight matrix that projects the visual embedding vector into the joint space, b is the bias term in the multimodal fusion linear combination process, and GELU(·) is the Gaussian error linear unit activation function:

[0115]

[0116] Among them, x is the intermediate vector after the fusion linear transformation, that is, tanh(·) is the hyperbolic tangent function;

[0117] S34. Input the fused multimodal feature vector sequence into the Transformer encoding structure of the LayoutLMv3 model, perform multi-head self-attention mechanism and feedforward network calculation, and generate a context embedding vector set;

[0118] S35. Input the context embedding vector set into the classifier head structure of the LayoutLMv3 model, perform sequence labeling based on the Softmax function, and generate a data asset field set. Each field in the data asset field set includes field content, field type label, context window fragment, and original coordinate frame position for subsequent structured modeling and template matching.

[0119] The present invention realizes the joint expression of text content, layout position and image features by constructing a multimodal embedding structure. The system encodes each text block into a text vector, a layout vector and an image vector respectively, and performs weighted fusion on the three vectors as a unified input representation of the model. The fused representation is input into a Gaussian error linear unit activation function, which has continuity and smoothness in numerical processing and can effectively enhance the model's ability to express boundary semantics and local changes. This mechanism not only retains the integrity of text semantics, but also strengthens the modeling ability of spatial structure and visual features, thereby improving the system's perception and extraction accuracy of asset fields in documents. In business documents with dense field arrangement, complex layout structure or mixed text and graphics, the present invention has stronger robustness and adaptability, providing a reliable information basis for subsequent data asset modeling and standardized table entry.

[0120] In this embodiment, the S32 specifically includes:

[0121] S321, the text Token sequence T in the triple i Input to the text embedding sublayer to generate the corresponding text embedding vector The embedding process is based on a word vector mapping function fine-tuned from the enterprise corpus. The enterprise data document corpus includes historical contracts, financial ledgers, business reports, meeting minutes, audit records, and archive scans, covering the semantics and format structures commonly used in enterprise data asset management.

[0122] S322, the rectangular coordinate frame B in the triplet i Input to the layout embedding sublayer to generate the corresponding layout position embedding vector The position embedding function is mapped into a set of high-dimensional space coordinate embeddings;

[0123] S323: Input the image pixel matrix I into the visual embedding sublayer, first divide the image pixel matrix I into multiple fixed-size image patch sets, and then input it into the visual convolutional network to extract the visual feature embedding vector

[0124] S324, embed the text obtained above into a vector Layout position embedding vector and visual feature embedding vectors Keep the dimensions aligned and output to the multimodal fusion sublayer to generate the fused multimodal input feature vector.

[0125] This method constructs embedding vectors for text content, layout coordinates, and image features, and aligns them in a multimodal fusion sublayer, achieving a coordinated representation of semantic information, spatial structure, and visual content in enterprise documents. Text embedding, trained on enterprise-specific corpus, enhances the recognition of domain terminology. Layout embedding strengthens the modeling of relative field positions. Image embedding improves the parsing of mixed text and image documents. This embedding mechanism improves the accuracy and stability of the model's recognition of asset fields in complex documents.

[0126] In this embodiment, the S4 specifically includes:

[0127] S41. Parse each field element in the data asset field set, extract the field content, field type label, context embedding vector, and original coordinate frame information, and construct a field structure set. Each field structure is represented as a four-tuple.

[0128] S42. Perform a field aggregation operation on the field structure set, using an aggregation function based on semantic similarity and spatial proximity to aggregate multiple field structures belonging to the same asset entity into an asset object unit:

[0129] O j =Agg({F i |Sim(H i ,H k )>θ1∧Dist(B i ,B k )<θ2});

[0130] Among them, O jis the jth asset object unit, Agg(·) is the field aggregation function, F i For the i-th field structure, Sim(H i ,H k ) is the semantic similarity function, H i is the context embedding vector of the i-th field, H k is the context embedding vector of the k-th candidate field, θ1 is the semantic similarity threshold, Dist(B i ,B k ) is the spatial distance function, B i is the rectangular position box of the i-th field in the original document, B k is the rectangular location box of the k-th candidate field, θ2 is the spatial distance threshold, i is the index number of the current candidate field structure, j is the number of the asset object unit generated after aggregation, and k is the number of the center field structure currently being used as a reference;

[0131] S43, for each asset object unit O j Perform field standardization and map the subordinate field contents into a standardized field set {N1, N2, ..., N k ,…,N p}, where each standardized field N k Corresponding to one of the asset name, department, person in charge, source system, or field value. If the original field is missing, it is set to "null value" and a missing tag is recorded;

[0132] S44. Output all standardized field sets as a structured data asset object set, with each object accompanied by a corresponding original coordinate frame set for subsequent table entry template matching and audit tracking.

[0133] This invention improves the accuracy of determining the attribution relationship between fields through a field aggregation mechanism driven by both semantic similarity and spatial distance. The system uses a weighted function to aggregate field structures. The aggregation result simultaneously considers the semantic similarity of the fields and the proximity of their positions in the original document, enabling the aggregation of multiple fields into a unified data asset object. This mechanism effectively solves the problem of classification errors that can easily occur when fields are dispersed and semantic associations are weak, improving the construction quality of structured data asset objects and the contextual consistency of table mappings.

[0134] In this embodiment, the S5 specifically includes:

[0135] S51. Receive a standardized field set in a structured data asset object set as a matching input set, wherein the standardized field content includes asset name, department, responsible person, source system, field value, and field type label;

[0136] S52: Input the matching input set into a rule matching engine, which includes a field attribute parsing module, a template matching scoring module, and a trigger condition judgment module;

[0137] S53, the field attribute parsing module performs a k Constructing matching vectors The matching vector is composed of the field semantic label, historical matching frequency and location coordinate encoding to generate a matching vector sequence;

[0138] S54. The template matching scoring module performs scoring matching on the matching vector sequence and the field mapping rules in the preset table entry template set in sequence:

[0139]

[0140] Among them, S j,m is the matching score, k is the field number of the matching input field in the current object, p is the number of standardized fields extracted from the current structured data asset object, α is the weighted coefficient of the field semantic label similarity, Sim(·,·) is the similarity scoring function, and l k Encode the semantic label of the kth field, For table entry template T m The standard semantic label encoding of the kth field in , β is the weighted coefficient of the field’s historical frequency similarity, f k is the matching frequency statistics of the kth field in the historical table entry template, For table entry template T m The reference frequency value of the kth field in , γ is the weighted coefficient of the field position coordinate similarity, p k Encode the two-dimensional position coordinates of the k-th field in the original business document, For table entry template T m The standard position vector of the kth field in ;

[0141] S55, the trigger condition judgment module will match the score S j,m With threshold θ T Compare, when the score is greater than or equal to the threshold, select the corresponding template T m As a template candidate for the current asset object, the field mapping rules, form format configuration, and table entry trigger conditions are output to form a table entry template set.

[0142] The present invention achieves precise matching between standardized fields in structured data asset objects and preset table entry templates by constructing a field attribute parsing and template matching scoring mechanism. The system introduces a weighted scoring function that integrates semantic tag similarity, historical matching frequency, and two-dimensional position similarity to comprehensively evaluate candidate fields and template fields, generate ranked matching results, and automatically complete field mapping. This mechanism effectively improves the accuracy of template selection and the rationality of field configuration. It is particularly suitable for business form matching scenarios with complex structures and similar fields, helping to reduce the risk of field mismatching and improve table entry efficiency.

[0143] In this embodiment, S6 specifically includes:

[0144] S61, read the field mapping rules in the table template set, for each asset object unit O j The standardized field set is determined in the table template T m The corresponding standard form field set, record field mapping relationship pair (N k ,F k );

[0145] S62. Standardize the original field values ​​according to the field mapping rules, including unified encoding format, unit conversion, and null value filling strategy, execute the conversion function on each field, and generate form field values;

[0146] S63, combine all mapped field pairs and converted form field values ​​to form a data entry record object R j ;

[0147] S64, record object R in the table j The data entry table metadata is appended to the template, including the template number, field mapping timestamp, and mapping operator ID, to generate a data entry table record set.

[0148] This invention uses a standardized processing mechanism driven by field mapping rules to achieve unified formatting, standardized units, and null value completion for raw field values, improving the consistency and integrity of table data. Furthermore, by attaching template numbers and mapping metadata, the traceability and auditability of table records are enhanced.

[0149] In this embodiment, the S7 specifically includes:

[0150] S71. Each structured entry record object R in the data entry record set j Write to the asset ledger system. The writing process includes field docking, form positioning and data submission to complete the asset record registration operation;

[0151] S72. During the writing process, a table entry operation log is synchronously generated. The log content includes the template number, table entry record number, operation timestamp, operator ID, and target ledger form ID, forming a table entry operation log set.

[0152] S73. Generate an audit trail data set based on the field mapping information, template number, and structured field source during the writing process, to describe the mapping path between the field source and the target location;

[0153] S74. The entry log set and audit trail data set are transmitted back to the data governance platform or model training module as data support for subsequent model fine-tuning, rule engine modification, and abnormal audit analysis.

[0154] S75, confirm the status of each data entry record and generate a status indicator S j ∈ success, failure}, and is bound to the entry record number to form a write status mapping table for feedback control and subsequent fault-tolerant processing.

[0155] This invention achieves automatic entry of structured data records, process log generation, and field source tracking by building a complete closed-loop mechanism for ledger writing and auditing. During the writing process, the system synchronously records the template number, operation time, and field mapping path, improving the registration accuracy and auditability of data assets. Furthermore, through a status confirmation and feedback mechanism, it supports feedback optimization on the data governance platform, providing continuous support for model fine-tuning and rule updates, and enhancing the system's stability and fault tolerance in large-scale table entry scenarios.

[0156] Example 1:

[0157] To verify the feasibility of this invention, we applied it to the data governance process of a large manufacturing enterprise, focusing on the data assetization requirements of its three core business modules: contract management, financial accounting, and indicator reporting. This enterprise faced issues such as decentralized management across multiple departments, inconsistent data recording formats, and delayed data asset registration. In particular, during the contract approval process, manual extraction of key fields and matching of ledger table templates often resulted in data omissions, format inconsistencies, and unclear accountability, impacting the enterprise's unified inventory and management of data assets.

[0158] Before deploying the system of the present invention, the company used traditional manual methods to enter contract information, financial records, and related indicator data, processing approximately 450 documents per week. The average entry time for a single document was 5.3 minutes, and the error rate reached 7.6%. After the system was launched, through the preset data asset rule templates, the fine-tuned LayoutLMv3 recognition model, and the standardized field mapping mechanism, the entire process from original document reception, structured field recognition, rule matching, form conversion to writing into the asset ledger system was automated. The document processing capability was significantly improved, the processing time was shortened to 1.4 minutes per document, and the accuracy rate was increased to 98.3%, of which the field extraction accuracy increased by approximately 14%, and the template matching accuracy increased by more than 25%.

[0159] The system has been successfully integrated into the company's OA contract approval system, financial settlement platform, and data governance platform. It automatically parses and structures various document formats (such as PDF contracts, Excel financial ledgers, and scanned business documents) into tables, generating ledger records with field source, responsible department, generation time, and audit identification. The system supports rule version iteration and regular model retraining, offering excellent adaptability and maintainability, meeting the company's needs for full lifecycle management of data assets.

[0160] By comparing and analyzing data statistics over four consecutive weeks, we compiled the following "Performance Comparison Table of Automatic Data Asset Entry Systems." This table illustrates the changes in various metrics before and after the system's introduction, including entry efficiency, field extraction accuracy, template recognition accuracy, and manual intervention frequency. This further validates the practicality and promotional value of this invention in real-world applications.

[0161] Table 1 Performance comparison of data asset automatic entry system

[0162]

[0163] As can be seen from the above table, the enterprise data asset entry system based on artificial intelligence proposed in the present invention has significant improvements in multiple key performance indicators compared to traditional manual entry methods. First, in terms of processing efficiency, the average processing time of a single document was reduced from 5.3 minutes to 1.4 minutes after the system was deployed, and the overall processing efficiency was improved by more than 70%, significantly reducing the manpower cost and time expenditure required for data asset management. Secondly, in terms of field extraction accuracy, the system uses the fine-tuned LayoutLMv3 model to achieve accurate recognition of multimodal document content, and the accuracy rate has increased from 84.1% of manual extraction to 98.3%, effectively avoiding the risks of missed recognition and misrecognition.

[0164] In terms of template matching accuracy, the system relies on a rule-matching engine to automatically adapt structured data asset objects, increasing recognition accuracy from 71.6% to 96.5%, significantly reducing the probability of misconfiguration during form field matching. At the same time, the frequency of manual intervention dropped from 17.2 times per 100 documents to 2.1 times, demonstrating the system's high degree of automatic processing stability and its ability to guarantee data quality without the need for frequent manual proofreading. Furthermore, the completeness rate of entry record fields increased from 81.5% to 97.4%, demonstrating the system's strong expressive capabilities in integrating multiple document fields and retaining contextual information, effectively supporting the complete registration of data assets.

[0165] Finally, in terms of asset ledger audit consistency, the score increased from 74 points compared to the traditional method to 93.6 points, demonstrating the system's enhanced audit capabilities in terms of ledger traceability, operation tracing, and field attribution determination. Overall, this invention achieves intelligent process reconstruction from raw data document parsing to structured table entry and then to a closed-loop audit loop, addressing the low efficiency, error-proneness, and unclear accountability issues inherent in the traditional process. It demonstrates strong practicality and potential for widespread adoption.

[0166] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An enterprise data asset entry system based on artificial intelligence, characterized by: include: The original document receiving module is used to receive data source files from multiple systems within the enterprise and generate an original document collection; A multimodal data construction module is used to perform text extraction, layout coordinate analysis and image data preprocessing on the original document collection to construct a multimodal input data collection; The AI ​​recognition module is used to input a multimodal input data set into the LayoutLMv3 model fine-tuned with the enterprise document corpus, perform image and text semantic modeling, and identify the set of fields with data asset attributes; A structured asset modeling module, used to convert a set of identification fields into a structured data asset object; The rule matching and template selection module is used to perform feature matching calculations on structured data asset objects, select the optimal table entry template, and generate field mapping rules and table entry trigger configurations; The field standardization mapping module is used to convert values, unify formats, and complete missing fields according to field mapping rules to generate a structured data entry record set; The table entry metadata supplement module is used to add a template number, mapping timestamp, and operator identification to each data entry record; The ledger writing and auditing module is used to write data entry records into the enterprise asset ledger system and simultaneously generate entry logs and field mapping audit information; The entry status management module is used to confirm the writing status of each data record and generate a status mapping table.

2. The enterprise data asset entry system based on artificial intelligence according to claim 1 is characterized in that: The modules are implemented as follows: S1. Receive original business documents from multiple data sources within the enterprise and generate an original document collection; S2. Perform preprocessing operations on the original document set to extract text content, layout coordinate information, and image pixel information to construct a multimodal input data set; S3. Input the multimodal input data set into the pre-trained model LayoutLMv3 based on visual document understanding. The LayoutLMv3 model is fine-tuned and trained on the enterprise data document corpus to perform joint image and text modeling and generate a data asset field set. S4. Construct a structured data asset object set based on the data asset field set; S5. Input the structured data asset object set into a preset rule matching engine, perform a matching operation based on the attribute information of each structured data asset object, and determine an entry template set; S6. Based on the table entry template set, convert the field contents in the structured data asset object set into a standardized form field format to generate a data entry record set; S7. Write the data entry record set into the asset ledger system, recording the execution time of the entry operation, the operator ID, and the template ID used.

3. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The original document collection includes contract documents in PDF format, business reports in Word format, financial ledgers in Excel format, scanned reports in image format, scanned archive copies in TIFF format, and structured document snapshots exported through the office system.

4. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S2 specifically includes: S21. Performing structured parsing processing on the PDF, Word, and Excel documents in the original document collection, extracting paragraph text, table fields, and form structures, and numbering them sequentially according to the document pages to construct a preliminary text content set and a structure tag set; S22, performing OCR recognition processing on the image format documents and TIFF format documents in the original document set, using an optical character recognition engine to extract valid text information in the image, recording the position range of each text block in combination with the image position information, and generating an OCR text set and a position coordinate set; S23, establishing a binding relationship between each text unit in the preliminary text content set and the OCR text set and a corresponding two-dimensional spatial position, wherein the two-dimensional spatial position is represented as a rectangular box coordinate; S24. Perform standardized preprocessing operations on the image format document, including unifying the image resolution, image size, color channel arrangement and compression format, generating an image pixel matrix set, and pairing the image pixel matrix set with the text content and spatial coordinate information to construct a multimodal input data set.

5. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S3 specifically includes: S31, uniformly encode the text content, the corresponding two-dimensional space coordinate position and the image pixel matrix set in the multimodal input data set, where the input of each text block is represented as a triple (T i ,B i ,I), where T i Represents the Token sequence of the i-th text block, B i represents the corresponding rectangular coordinate frame, and I represents the image pixel matrix of the corresponding page; S32. Input each component element of the triple into the embedding layer of the LayoutLMv3 model respectively, and perform an embedding extraction operation on the LayoutLMv3 model after fine-tuning and training the model with the enterprise data document corpus to obtain a text embedding vector, a layout position embedding vector, and a visual feature embedding vector; S33. Perform weighted fusion on the above three types of embedding vectors and generate the fused multimodal feature vector through the Gaussian error linear unit activation function: in, is the fused multimodal input feature vector corresponding to the i-th text block, is the text embedding vector of the i-th text block, is the layout position embedding vector of the i-th text block, is the visual feature embedding vector of the i-th text block, W t W is the weight matrix that projects the text embedding vector into the joint space. l is the weight matrix that projects the layout embedding vector into the joint space, W v is the weight matrix that projects the visual embedding vector into the joint space, b is the bias term in the multimodal fusion linear combination process, and GELU(·) is the Gaussian error linear unit activation function; S34. Input the fused multimodal feature vector sequence into the Transformer encoding structure of the LayoutLMv3 model, perform multi-head self-attention mechanism and feedforward network calculation, and generate a context embedding vector set; S35. Input the context embedding vector set into the classifier head structure of the LayoutLMv3 model, perform sequence labeling based on the Softmax function, and generate a data asset field set. Each field in the data asset field set includes field content, field type label, context window fragment, and original coordinate box position.

6. The enterprise data asset entry system based on artificial intelligence according to claim 5 is characterized in that: The S32 specifically includes: S321, the text Token sequence T in the triple i Input to the text embedding sublayer to generate the corresponding text embedding vector The embedding process is based on the word vector mapping function fine-tuned by the enterprise corpus; S322, the rectangular coordinate frame B in the triplet i Input to the layout embedding sublayer to generate the corresponding layout position embedding vector The position embedding function is mapped into a set of high-dimensional space coordinate embeddings; S323: Input the image pixel matrix I into the visual embedding sublayer, first divide the image pixel matrix I into multiple fixed-size image patch sets, and then input it into the visual convolutional network to extract the visual feature embedding vector S324, embed the text obtained above into a vector Layout position embedding vector and visual feature embedding vector Keep the dimensions aligned and output to the multimodal fusion sublayer.

7. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S4 specifically includes: S41. Parse each field element in the data asset field set, extract the field content, field type label, context embedding vector, and original coordinate frame information, and construct a field structure set. Each field structure is represented as a four-tuple. S42. Perform a field aggregation operation on the field structure set, using an aggregation function based on semantic similarity and spatial proximity to aggregate multiple field structures belonging to the same asset entity into an asset object unit: O j =Agg({F i ∣Sim(H i ,H k )>θ1∧Dist(B i ,B k )<θ2}); Among them, O j is the jth asset object unit, Agg(·) is the field aggregation function, F i For the i-th field structure, Sim(H i ,H k ) is the semantic similarity function, H i is the context embedding vector of the i-th field, H k is the context embedding vector of the k-th candidate field, θ1 is the semantic similarity threshold, Dist(B i ,B k ) is the spatial distance function, B i is the rectangular position box of the i-th field in the original document, B k is the rectangular location box of the k-th candidate field, θ2 is the spatial distance threshold, i is the index number of the current candidate field structure, j is the number of the asset object unit generated after aggregation, and k is the number of the center field structure currently being used as a reference; S43, for each asset object unit O j Perform field standardization and map the subordinate field contents into a standardized field set {N1, N2, ..., N k ,…,N p }, where N k is a standardized field; S44. Output all standardized field sets as a structured data asset object set, with each object being accompanied by a corresponding original coordinate frame set.

8. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S5 specifically includes: S51. Receive a standardized field set in a structured data asset object set as a matching input set, wherein the standardized field content includes asset name, department, responsible person, source system, field value, and field type label; S52: Input the matching input set into a rule matching engine, which includes a field attribute parsing module, a template matching scoring module, and a trigger condition judgment module; S53, the field attribute parsing module performs a k Constructing matching vectors The matching vector is composed of the field semantic label, historical matching frequency and location coordinate encoding to generate a matching vector sequence; S54. The template matching scoring module performs scoring matching on the matching vector sequence and the field mapping rules in the preset table entry template set in sequence: Among them, S j,m is the matching score, k is the field number of the matching input field in the current object, p is the number of standardized fields extracted from the current structured data asset object, α is the weighted coefficient of the field semantic label similarity, Sim(·,·) is the similarity scoring function, and l k Encode the semantic label of the kth field, For table entry template T m The standard semantic label encoding of the kth field in , β is the weighted coefficient of the field’s historical frequency similarity, f k is the matching frequency statistics of the kth field in the historical table entry template, For table entry template T m The reference frequency value of the kth field in , γ is the weighted coefficient of the field position coordinate similarity, p k Encode the two-dimensional position coordinates of the k-th field in the original business document, For table entry template T m The standard position vector of the kth field in ; S55, the trigger condition judgment module will match the score S j,m With threshold θ T Compare, when the score is greater than or equal to the threshold, select the corresponding template T m As a template candidate for the current asset object, the field mapping rules, form format configuration, and table entry trigger conditions are output to form a table entry template set.

9. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S6 specifically includes: S61, read the field mapping rules in the table template set, for each asset object unit O j The standardized field set is determined in the table template T m The corresponding standard form field set, record field mapping relationship pair (N k ,F k ); S62. Standardize the original field values ​​according to the field mapping rules, including unified encoding format, unit conversion, and null value filling strategy, execute the conversion function on each field, and generate form field values; S63, combine all mapped field pairs and converted form field values ​​to form a data entry record object R j ; S64, record object R in the table j The data entry table metadata is appended to the template, including the template number, field mapping timestamp, and mapping operator ID, to generate a data entry table record set.

10. The enterprise data asset entry system based on artificial intelligence according to claim 2 is characterized in that: The S7 specifically includes: S71, each structured entry record object R in the data entry record set j Write to the asset ledger system. The writing process includes field docking, form positioning and data submission to complete the asset record registration operation; S72. Synchronously generate an entry operation log during the writing process. The log content includes the template number, entry record number, operation timestamp, operator ID, and target ledger form ID, forming an entry operation log set. S73. Generate an audit trail data set based on the field mapping information, template number, and structured field source during the writing process; S74. Send the entry log set and the audit trail data set back to the data governance platform or model training module; S75, confirm the status of each data entry record and generate a status indicator S j ∈ success, failure}, and bind it with the entry record number to form a write status mapping table.

Images