A method, apparatus and program product for enterprise personnel affinity analysis

Abstract

The present application relates to the technical field of data processing, and discloses an enterprise personnel correlation analysis method, device and program product.The method comprises the following steps: obtaining an original binary document stream of a to-be-inspected enterprise's to-be-analyzed personnel document and a preset engine format adaptation degree matrix, and constructing a directed acyclic graph model; performing global optimal distribution and information extraction on the directed acyclic graph model respectively to obtain a table header global optimal distribution result and a hierarchical structured personnel information dataset; using an incremental collection strategy based on change event driving and a retry mechanism with exponential backoff to obtain an external information set of the to-be-inspected enterprise from an external public data source; constructing a four-element heterogeneous graph of the to-be-analyzed personnel document, and determining a comprehensive correlation strength value of each correlation relationship edge in the four-element heterogeneous graph; and generating a structured inspection report and an interactive correlation relationship graph of the to-be-inspected enterprise according to the hierarchical structured personnel information dataset, the external information set and the plurality of comprehensive correlation strength values.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method, apparatus, and program product for analyzing the correlation of personnel in an enterprise. Background Technology

[0002] With increasingly stringent compliance requirements, identifying potential conflicts of interest between internal and external personnel has become a crucial aspect of corporate risk management. Currently, companies face the following challenges when conducting personnel correlation analysis: The document formats are diverse: the company's personnel information forms are available in various formats such as PDF, Word, Excel, .doc, and .docx, making manual information extraction inefficient and prone to errors.

[0003] The table structure is complex: the information table contains various table formats such as nested tables, ordinary tables, and standard tables, which traditional parsing tools find difficult to fully identify.

[0004] Table header fields are often messed up: In actual documents, fields such as title, position, professional title, gender, and ethnicity are often filled with incorrect information, leading to errors in information extraction.

[0005] Information is scattered: internal personnel information is stored in the company's human resources system, while external personnel information is scattered through channels such as business registration and public disclosure, making it difficult to effectively integrate and compare.

[0006] Hidden connections: The relationships of interest between individuals are often hidden, making it difficult for traditional manual screening methods to fully identify potential risks.

[0007] Therefore, there is an urgent need for a technical solution that can automatically extract personnel information from multi-format documents, intelligently correct misaligned headers, and identify the relationships between internal and external personnel. Summary of the Invention

[0008] This invention provides a method, device, and program product for analyzing the relationships between enterprise personnel, in order to solve the problems that existing methods for analyzing the relationships between enterprise personnel cannot automatically extract personnel information from multi-format documents, correct misaligned table headers, or automatically identify relationships between external personnel.

[0009] In a first aspect, the present invention provides a method for analyzing the correlation between personnel in an enterprise, the method comprising: The process involves obtaining the original binary document stream and a preset engine format adaptation matrix of the personnel documents to be analyzed from the company under review. Multiple elements in the preset engine format adaptation matrix represent the historical average parsing quality scores of different parsing engines for documents of different formats. Based on the original binary document stream and the preset engine format adaptation matrix, a directed acyclic graph (DAG) model of the personnel documents to be analyzed is constructed using a multi-engine adaptive fusion unified representation transformation method. Global optimal allocation and information extraction are then performed on the DAG model to obtain the globally optimal allocation result of the table headers for the personnel documents to be analyzed. The system uses a hierarchical structured personnel information dataset; it employs an incremental data acquisition strategy driven by change events and a retry mechanism with exponential backoff to obtain external information sets of the companies under review from publicly available data sources; it constructs a quaternary heterogeneous graph of the personnel documents to be analyzed, and processes the quaternary heterogeneous graph using a method based on heterogeneous graph neural networks and spatiotemporal decay to obtain the comprehensive association strength value of each association edge in the quaternary heterogeneous graph; based on the hierarchical structured personnel information dataset, the external information set, and multiple comprehensive association strength values, it generates a structured review report and an interactive association graph of the companies under review.

[0010] The enterprise personnel correlation analysis method provided by this invention helps avoid the problems of inconsistent standards and low efficiency caused by manually configuring parsing rules and collecting external data separately, by acquiring a compatibility matrix between multi-source data and a preset engine format. Furthermore, through unified multi-format representation conversion and the construction of a directed acyclic graph of tables, it can complete the standardized structural conversion of heterogeneous documents and the hierarchical topology modeling of nested tables in one go, simultaneously solving the problems of poor compatibility of multi-format documents, loss of nested table hierarchy, and inability to quantify table spatial relationships. This provides a unified graph structure carrier with complete hierarchical, spatial, and layout characteristics for header error correction and personnel information extraction. Furthermore, through globally optimal header allocation and hierarchical personnel information extraction, it significantly reduces the probability of field mismatch and omission of nested personnel information, ultimately outputting standardized internal personnel data with aligned hierarchy and standardized fields. Furthermore, by utilizing an incremental collection strategy based on change events and a retry mechanism with exponential backoff to collect external information, it significantly reduces the number of interface requests, data storage, and computing power consumption while ensuring the timeliness of external data, solving the shortcomings of traditional full-data collection, such as long processing time and lack of fallback in data source rate limiting failures. Furthermore, by constructing a quaternary heterogeneous graph and processing it using a method for calculating implicit correlations based on heterogeneous graph neural networks and spatiotemporal decay, a unified correlation graph is built by integrating multiple types of entities. A dual decay method based on time and geography is introduced to quantify the true risk weight of relationships, quantifying the risk strength of each relationship and achieving a unified quantitative assessment of explicit and implicit correlations. Furthermore, by combining a hierarchical structured personnel information dataset, external information sets, and multiple comprehensive correlation strength values, a structured review report and interactive correlation graph for the company under review are generated, meeting the business needs of enterprise compliance review, audit traceability, and intuitive risk assessment. Therefore, by implementing this invention, the pain points of traditional manual review, such as inefficient multi-format document processing, disordered header fields, omission of nested personnel information, high cost of external data updates, difficulty in identifying hidden interest relationships, inability to quantify and classify risks, and lack of audit traceability, are solved. It also considers four core indicators: document parsing compatibility, information extraction accuracy, correlation mining depth, and compliance traceability, adapting to intelligent review of personnel correlation risks across all scenarios for enterprise personnel and supplier compliance.

[0011] In one optional implementation, based on the original binary document stream and a preset engine format adaptation matrix, a directed acyclic graph model of the personnel documents to be analyzed is constructed through a multi-engine adaptive fusion unified representation transformation method, including: Using the original binary document stream and a preset engine format adaptation matrix, the document format of the personnel document to be analyzed is adaptively identified, and multiple initial parsing engines matching the format of the personnel document to be analyzed are determined. The personnel document to be analyzed is then parsed using these multiple initial parsing engines, resulting in multiple engine parsing results. Each engine parsing result includes a text content matrix and a layout feature vector. Using evidence reasoning theory, the parsing results from multiple engines are fused to generate a document object model tree structure and a JSON text matrix set for the personnel document to be analyzed. A depth-first recursive traversal is performed on the document object model tree structure and the JSON text matrix set to construct a directed acyclic graph model.

[0012] The personnel correlation analysis method provided by this invention avoids problems such as corrupted files due to single magic number recognition, WPS / Word differentiation failure, and inconsistent accuracy of fixed engine parsing for various documents by adaptive document format recognition and matching engine filtering, thus improving the accuracy of format recognition and the basic quality of single document parsing. Furthermore, by using multi-engine parallel parsing and inputting the parsing results from multiple engines, it compensates for the deficiencies of single-engine table reconstruction. Furthermore, by fusing multi-engine parsing results through evidence reasoning, it overcomes the problem of conflicting and invalidating multi-engine results, quantifies the reliability confidence of each parsing result, reduces parsing omissions and distortions caused by document corruption and complex tables, and improves the completeness of standardized DOM and matrix data. Furthermore, by constructing a directed acyclic graph model through depth-first recursive traversal, it fully preserves the parent-child hierarchy, cell merging, and row and column spatial relationships of multi-level nested tables, enabling rapid location of any nested cell and solving the problem of losing the nested table topology in traditional flat parsing. Therefore, by implementing this invention, the preprocessing success rate of complex nested personnel tables is significantly improved.

[0013] In one optional implementation, the document format of the personnel document to be analyzed is adaptively identified using the original binary document stream and a preset engine format adaptation matrix, and multiple initial parsing engines that match the format of the personnel document to be analyzed are determined, including: Multiple byte windows of a preset length are extracted at multiple preset offset positions in the original binary document stream, and multiple frequency distribution histogram vectors corresponding to the multiple byte windows are calculated. The multiple frequency distribution histogram vectors and the byte information entropy vectors within the multiple byte windows are concatenated to construct a composite fingerprint vector of the personnel document to be analyzed. The byte information entropy vector is used to reflect the uniformity of byte distribution within the byte window. Multiple original weighted cosine similarities are calculated between the composite fingerprint vector and multiple pre-constructed standard fingerprint vectors. Based on the multiple original weighted cosine similarities, the format of the personnel document to be analyzed and the target weighted cosine similarity are determined, where the target weighted cosine similarity is the largest among the multiple original weighted cosine similarities. When the target weighted cosine similarity is greater than or equal to a preset first threshold, the format of the personnel document to be analyzed is determined to be an identified format, and the engine adaptation score of the format of the personnel document to be analyzed is determined based on a preset engine format adaptation matrix. The engine adaptation score is compared with a preset second threshold, and based on the comparison result, multiple initial parsing engines that match the format of the personnel document to be analyzed are determined.

[0014] The enterprise personnel correlation analysis method provided by this invention, by extracting histograms through multi-offset byte windows, overcomes the shortcomings of relying solely on file header pattern recognition for format identification. It can still stably extract distinguishing features from documents with damaged transmission or tampered headers, improving the robustness of damaged document identification. Furthermore, by constructing a composite fingerprint vector, it integrates both byte distribution histogram and byte information entropy features, addressing the limitation of traditional single-feature methods in distinguishing documents with the same container structure, thus refining the granularity of format differentiation. Furthermore, by using weighted cosine similarity matching, it helps strengthen the distinguishing role of header format features and entropy features, thereby accurately matching the standard sample fingerprint database and improving the accuracy of format determination. Furthermore, by determining document format through maximum similarity, it avoids format misjudgment that could lead to subsequent full-process parsing failure. Furthermore, by comparing and selecting initial engines based on engine adaptability scores, it avoids blindly calling all engines, reducing ineffective computing power consumption, while providing a fallback for documents without highly compatible engines, balancing computing efficiency and parsing coverage completeness. Therefore, by implementing this invention, high-precision automatic recognition of damaged, heterogeneous, and multi-format mixed personnel documents is achieved. It solves the problems of traditional single file header recognition failure, difficulty in distinguishing between WPS and Word, and waste of computing power due to indiscriminate engine calls. While ensuring the accuracy of format recognition, it intelligently reduces the scale of parsing engine calls and balances the robustness of document recognition with the system's computing power overhead.

[0015] In one optional implementation, evidence-based reasoning theory is used to fuse the parsing results from multiple engines, generating a document object model tree structure and a JSON text matrix set for the personnel documents to be analyzed, including: Based on multiple preset evaluation dimensions, a local quality evidence vector is generated for each target parsing engine, and a basic probability allocation function is constructed for each initial parsing engine. Based on each basic probability allocation function, the basic confidence probability value and basic unconfidence probability value of each engine's parsing result are calculated, and the basic confidence probability distribution of each engine's parsing result is determined. Using Dempster's combination rule, the basic confidence probability distributions of each engine's parsing result are recursively combined pairwise, and a global fusion confidence value is determined. When the global fusion confidence value is less than a preset fusion threshold, the highest basic confidence probability value is obtained, and a document object model tree structure and a JSON text matrix set are generated based on the engine parsing result corresponding to the highest basic confidence probability value.

[0016] The enterprise personnel correlation analysis method provided by this invention generates local quality evidence vectors for each target parsing engine according to multiple preset evaluation dimensions and constructs a basic probability allocation function for each initial parsing engine, quantifying the reliability of single-engine parsing. By replacing simple binary judgments of right and wrong, it refines the characterization of the degree of defects in each parsing result. Furthermore, based on each basic probability allocation function, it calculates the basic credibility probability value and basic unreliability probability value of each engine's parsing result and determines the basic credibility probability distribution of each engine's parsing result. Then, it uses Dempster's combination rule to perform pairwise recursive combination, solving the problem of weighted voting failure when multiple engine parsing results conflict. Finally, by integrating all engine evidence, it outputs a globally unified credibility, which can objectively reflect the overall parsing completeness of the document. Furthermore, when the global fusion confidence value is less than a preset fusion threshold, it avoids the amplification of errors in the fusion of conflicting results, ensuring the basic usability of the output data. Therefore, by implementing this invention, the limitations of traditional weighted voting in handling contradictory parsing results are overcome. It can accurately distinguish between high-quality, defective, and low-reliability document parsing outputs. Furthermore, by stably outputting low-distortion DOM trees and text matrices in multi-engine parallel parsing scenarios, it reduces the underlying data errors in subsequent header recognition and entity extraction.

[0017] In one alternative implementation, the method further includes: When the global fusion confidence value is greater than or equal to the preset fusion threshold, the probability difference between the highest and second-highest basic probability values ​​is calculated based on the basic confidence probability values ​​of each engine's parsing results. When the probability difference is greater than or equal to the preset third threshold, a document object model tree structure and a JSON text matrix set are generated based on the engine parsing results corresponding to the highest basic confidence probability value. When the probability difference is less than the preset third threshold, the document object model tree structure and a JSON text matrix set are generated after fine-grained completion of the engine parsing results corresponding to the highest basic confidence probability value, using the basic confidence probability values ​​of each engine's parsing results as weights.

[0018] The enterprise personnel correlation analysis method provided by this invention quantifies the quality gap between the optimal and second-best analytical results by calculating the probability difference between the highest and second-highest basic probability values. This helps distinguish between scenarios where a single engine is significantly optimal and multi-engine results are of similar quality, providing a basis for differentiated data completion strategies. Furthermore, when the probability difference is greater than or equal to a preset third threshold, the analytical result corresponding to the highest basic credibility probability value is directly selected, generating a Document Object Model (DOM) tree structure and a JSON text matrix set. This avoids irrelevant errors introduced by multi-engine fusion, improving processing efficiency and data purity. Further, when the probability difference is less than the preset third threshold, the analytical result corresponding to the highest basic credibility probability value is fine-grained and completed using the basic credibility probability value of each engine's analytical result as a weight. This generates a DOM tree structure and a JSON text matrix set, integrating effective information from multiple engines to compensate for missing cells in a single engine, further improving the completeness of table data. Therefore, by implementing this invention, by selecting a single optimal result or using multi-engine weighted fusion for completion based on different scenarios, the missing fields in tables are maximized while ensuring data accuracy, further reducing the probability of missing information in complex personnel tables.

[0019] In one optional implementation, the directed acyclic graph model is subjected to global optimal allocation and information extraction respectively to obtain the global optimal allocation result of the header of the personnel document to be analyzed and the hierarchical structured personnel information dataset, including: Based on a directed acyclic graph (DAG) model, a header recognition and error correction method based on bidirectional attention flow and global optimal allocation is used to obtain the global optimal allocation result of the header of the personnel document to be analyzed. A multidimensional position encoding vector is constructed for each node in the DAG model, and the topological context vector of each node is calculated. Based on the multidimensional position encoding vector and topological context vector of each node, an initial input representation vector for each node is generated. Based on an edge type awareness mechanism, the initial input representation vector of each node is processed through a graph attention network to obtain the target hidden layer representation vector of each node, which incorporates neighborhood information. Based on the global optimal allocation result of the header and preset nested topological constraints, the target hidden layer representation vector of each node is used to perform multi-task decoding of the personnel document to be analyzed, generating a hierarchical structured personnel information dataset.

[0020] The enterprise personnel correlation analysis method provided by this invention utilizes a bidirectional attention global header error correction and edge-aware graph attention nested extraction architecture to simultaneously solve three core problems: header field disorder, incomplete extraction of personnel entities from nested tables, and entity hierarchy confusion. This enables the output of a structured internal personnel dataset with standardized fields, clear hierarchy, and complete family member information. Compared with traditional flat NER, it significantly improves the completeness of nested table information extraction and the F1 score for entity recognition.

[0021] In one optional implementation, based on a directed acyclic graph model, and processed by a header recognition and error correction method based on bidirectional attention flow and global optimal allocation, the globally optimal header allocation result of the personnel document to be analyzed is obtained, including: The method extracts multiple header texts from the directed acyclic graph model to obtain multiple header cells in the personnel document to be analyzed. These header texts and multiple preset standard categories are then mapped using a pre-trained semantic encoder to obtain multiple first semantic vectors and multiple second semantic vectors. Based on these first and second semantic vectors, a forward attention flow is used to calculate the first attention weight value of each header text for each preset standard category. The matching score of each column of data in each header text with the data type of each preset standard category is calculated, and the aggregated content feature vector of each column is determined. Based on these aggregated content feature vectors and multiple second semantic vectors, a backward attention flow is used to calculate the second attention weight value of each column of data for each header text. The multiple first and second attention weight values ​​are weighted and fused, and an optimization function is constructed, where the optimization function represents the minimum total allocation cost. Within a preset time complexity, the Hungarian algorithm is used to solve the optimization function, and the globally optimal header allocation result for the personnel document to be analyzed is determined.

[0022] The enterprise personnel correlation analysis method provided by this invention completely solves the defects of traditional one-way keyword matching in that it cannot identify header string errors and multiple header category conflicts by establishing an intelligent error correction system with two-way verification of header text and column data and global mutual exclusion constraints. It achieves fully automatic and accurate correction of a large number of non-standard and misaligned headers in personnel files without the need for manual calibration, which greatly improves the standardization of personnel document fields and review efficiency.

[0023] In one optional implementation, the quaternary heterogeneous graph is processed using a method for calculating implicit association based on heterogeneous graph neural networks and spatiotemporal decay to obtain the comprehensive association strength value of each association edge in the quaternary heterogeneous graph, including: Based on preset time decay functions and preset spatial decay functions, the quaternary heterogeneous graph is processed by heterogeneous graph convolution to obtain the target embedding representation vector of each node in the quaternary heterogeneous graph that integrates multi-hop heterogeneous neighborhood information. Link prediction is performed using the bilinear mapping of the target embedding representation vector of each node to obtain the predicted probability value of each edge in the quaternary heterogeneous graph. Based on the predicted probability value of each edge, multiple association edges of the quaternary heterogeneous graph are determined. The scores of each association edge on preset multiple association dimensions are calculated and the comprehensive association strength value of each association edge is determined.

[0024] The enterprise personnel correlation analysis method provided by this invention realizes unified quantitative assessment of explicit and implicit personnel and enterprise correlations. It restores the true risk value of correlations through spatiotemporal attenuation. At the same time, it uses heterogeneous graph multi-hop reasoning to uncover hidden interest relationships and combines a multi-dimensional collaborative risk model to accurately quantify the intensity of correlation risks. It solves the pain points of traditional review that can only identify direct correlations, has a single risk assessment dimension, and cannot distinguish the timeliness of relationships. It provides a quantifiable numerical basis for compliance risk classification and early warning.

[0025] Secondly, the present invention provides an enterprise personnel correlation analysis device, the device comprising: The first acquisition module acquires the original binary document stream and a preset engine format adaptation matrix of the personnel documents to be analyzed from the company under review. Multiple elements in the preset engine format adaptation matrix represent the historical average parsing quality scores of different parsing engines for documents of different formats. The first processing module constructs a directed acyclic graph (DAG) model of the personnel documents to be analyzed, based on the original binary document stream and the preset engine format adaptation matrix, using a multi-engine adaptive fusion unified representation transformation method. The second processing module performs global optimal allocation and information extraction on the DAG model to obtain the globally optimal allocation result of the header of the personnel documents to be analyzed. The system comprises the following modules: a first module and a second module, which acquires external information about the company under review from publicly available data sources using an incremental acquisition strategy driven by change events and a retry mechanism with exponential backoff; a third module, which constructs a quaternary heterogeneous graph of the personnel documents to be analyzed and processes it using a method based on a heterogeneous graph neural network and spatiotemporal decay to calculate the implicit correlation strength of each relation edge in the quaternary heterogeneous graph; and a fourth module, which generates a structured review report and an interactive relational graph of the company under review based on the hierarchical structured personnel information dataset, the external information set, and multiple comprehensive correlation strength values.

[0026] Thirdly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the enterprise personnel correlation analysis method described in the first aspect or any corresponding embodiment. Attached Figure Description

[0027] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the enterprise personnel correlation analysis method according to an embodiment of the present invention; Figure 3 This is a structural block diagram of an enterprise personnel correlation analysis device according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0031] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0032] As an optional application scenario of this invention, considering the specific application environment architecture or specific hardware architecture upon which the enterprise personnel correlation analysis method depends, the specific application environment architecture or specific hardware architecture is described here. For example... Figure 1 As shown, the architecture system may include at least one terminal device and at least one server. Figure 1The system is illustrated in the example, which includes a computer 101, a mobile terminal 102, and a server 103, and the terminal devices such as the computer 101 and the mobile terminal 102 are connected to the server 103 through a network 110.

[0033] Specifically, the terminal device can be a smartphone, tablet, laptop, PDA, desktop computer, game console, smart TV, smart wearable device, in-vehicle terminal, VR (Virtual Reality) device, AR (Augmented Reality) device, etc. Server 103 can be a standalone physical server, a server cluster, a distributed system, or a cloud server providing cloud services. Network 110 can be a wired or wireless network, examples of which include, but are not limited to, the Internet, corporate intranet, local area network, wide area network, mobile communication network, and combinations thereof.

[0034] According to an embodiment of the present invention, an embodiment of a method for analyzing the correlation of personnel in an enterprise is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a method for analyzing the correlation between enterprise personnel, which can be used on the aforementioned mobile terminals, such as mobile phones and tablets. Figure 2 This is a flowchart of an enterprise personnel correlation analysis method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain the original binary document stream and preset engine format adaptation matrix of the personnel documents to be analyzed from the enterprise to be reviewed.

[0036] In one optional embodiment, the personnel documents to be analyzed represent various personnel files storing personnel information in a corporate compliance review scenario. These may include cadre appointment and dismissal approval forms, personal information report forms, etc., and the files may cover multiple formats such as .doc, .docx, .wps, PDF, scanned images, and OCR text. Furthermore, the documents may contain ordinary tables and multi-level nested tables, recording core fields required for review, such as basic information of internal personnel, family members, and social relationships.

[0037] Furthermore, the raw binary document stream represents the underlying binary byte sequence of the personnel document to be analyzed before format transcoding and text extraction, denoted as... The total length in bytes is .

[0038] Furthermore, the preset engine format adaptation matrix Each element Indicates the first The parsing engine for the first... The historical average parsing quality score for documents in this format.

[0039] Step S202: Based on the original binary document stream and the preset engine format adaptation matrix, the document is processed by a multi-format document unified representation conversion method based on multi-engine adaptive fusion to construct a directed acyclic graph model of the personnel documents to be analyzed.

[0040] In one optional embodiment, the multi-engine adaptive fusion-based multi-format document unified representation conversion method represents a method that can automatically select the optimal parsing strategy and uniformly convert heterogeneous documents (such as .docx, .doc, .wps, PDF, and scanned text preprocessed by OCR) into standardized structured data representations when faced with personnel files of heterogeneous origin and diverse formats (such as .docx, .doc, .wps, PDF, and scanned text preprocessed by OCR).

[0041] Furthermore, the directed acyclic graph model This represents an acyclic topology graph constructed with all cells within the document as independent nodes and cell nesting / row / column / merging relationships as directed edges. The node set... Each node in Associating a 7-tuple property is shown in the following relation (1):

[0042] in, It is a globally unique identifier. The parent table identifier. For nesting level depth, For spatial location and span, For cell text, For cell type, For additional metadata.

[0043] Furthermore, edge set For a directed edge, edge Represents child nodes Nested in the parent node In the cell it represents. This represents the maximum nesting depth.

[0044] In one optional embodiment, the underlying storage structure, encoding, and table rendering logic of various personnel documents differ greatly. A single parsing engine cannot handle all formats and complex nested tables. Furthermore, relying solely on the file header pattern is prone to format recognition failures. In this embodiment, using the original binary document stream and a preset engine format adaptation matrix as input, a multi-engine adaptive fusion-based unified representation conversion method for multi-format documents can transform the original personnel documents, which are characterized by chaotic formats and heterogeneous underlying storage, into a standardized DOM and text matrix. Then, by recursively traversing the table elements using a depth-first approach, a complete directed acyclic graph model that preserves multi-level nested relationships can be constructed.

[0045] Step S203 involves performing global optimal allocation and information extraction on the directed acyclic graph model to obtain the global optimal allocation result of the header of the personnel document to be analyzed and the hierarchical structured personnel information dataset.

[0046] In one optional embodiment, the global optimal allocation result of the table header represents the one-to-one mapping relationship between all table headers and standard personnel fields. It may include the original table header text, the corrected standard category, the allocation cost, and a complete set of audit records marked by manual review. It automatically corrects misalignment and string filling problems in the document table header fields, and achieves standardized classification of table headers without duplication or conflict.

[0047] Furthermore, the hierarchical structured personnel information dataset is represented by a table nesting hierarchy as the basis for division. It uses a graph attention network with topological constraints to extract personnel entity information from the table graph, and includes a clear, standardized, and traceable personnel data set with entity identification confidence and data legality verification labels.

[0048] In one optional embodiment, personnel forms commonly suffer from problems such as mismatch between header text and column data, multiple headers competing for the same standard field, and easy loss of family member information in multi-layered nested structures. Relying solely on unidirectional header text matching cannot correct these errors, and flattened extraction loses the context of nested hierarchies. In this embodiment, using a directed acyclic graph model as input, and sequentially performing bidirectional attention-based global optimal header matching and nested topological constraint graph attention-based entity extraction, it can output corrected header standard mapping results and a hierarchically structured personnel information dataset, respectively.

[0049] Step S204: Using an incremental data collection strategy driven by change events and a retry mechanism with exponential backoff, obtain the external information set of the company to be reviewed from external public data sources.

[0050] In one optional embodiment, the incremental collection strategy based on change event driving represents a lightweight data collection solution for external enterprises. This strategy does not periodically pull all of the enterprise's business registration information, but only compares the enterprise's historical synchronized records and only collects field data that has changed since the last synchronization. This reduces the amount of interface requests, storage and computing power consumption, while ensuring that information on changes in enterprise shareholders, executives, equity and other information is valid in real time.

[0051] Furthermore, the retry mechanism with exponential backoff represents a fault-tolerant retry strategy for external data source interface failures.

[0052] Furthermore, the external information set refers to a standardized structured data set collected from publicly available external data sources such as business registration information and corporate credit information. It can include tuple data such as basic corporate registration information, shareholders, directors, supervisors and senior executives, equity changes, external investments, and changes in business scope. Each data entry is bound to a unique corporate identifier, a data version hash, and a last synchronization timestamp, which is used for cross-comparison with internal personnel data to uncover conflicts of interest.

[0053] In one optional embodiment, an incremental acquisition strategy based on change events is used to collect external information, which can reduce the retrieval of invalid data. At the same time, by combining an exponential backoff retry mechanism to deal with the instability of external interfaces, standardized external business registration, equity, and personnel information of all companies on the review list can be obtained in batches, and finally a corresponding unified external information set can be generated.

[0054] For example, the system reads the complete list of companies to be reviewed and retrieves a historical synchronization snapshot for each company from the system database. Furthermore, this snapshot stores two markers: the last synchronization timestamp. Current data version hash , as an incremental comparison benchmark.

[0055] Furthermore, for each company on the review list During data collection, the system first queries the change announcement interface of the enterprise information disclosure system to obtain the data. List of changes since then .

[0056] Furthermore, if If empty, it indicates that the company information has not changed, and this data collection will be skipped; if If not empty, only the fields involved in the change will be captured (e.g., if the shareholder changes, only the shareholder information will be collected; if the business scope changes, only the business scope will be updated), instead of re-capturing all fields.

[0057] Furthermore, the collected data undergoes pattern alignment, is uniformly modeled into enterprise information tuples, and then updated. and .

[0058] Furthermore, considering the possibility of temporary unavailability of external data sources (such as during system maintenance or API rate limiting), this embodiment implements a retry mechanism with exponential backoff: for collection requests from the same enterprise, the number of consecutive failed attempts is limited. The corresponding retry waiting time is ,in Base waiting time (60 seconds). The maximum waiting time is set to 86,400 seconds, or 24 hours. Furthermore, when... If the threshold (5 times) is exceeded, the enterprise will be marked as "collection abnormal" and the operation and maintenance personnel will be notified.

[0059] Furthermore, all successfully collected and standardized enterprise tuple data are aggregated and bound with the corresponding enterprise identifier, change time, version hash, and collection timestamp, and then uniformly packaged to form a complete external information set.

[0060] Step S205: Construct a quaternary heterogeneous graph of the personnel documents to be analyzed, and process the quaternary heterogeneous graph using a method for calculating implicit correlation based on heterogeneous graph neural networks and spatiotemporal decay to obtain the comprehensive correlation strength value of each correlation edge in the quaternary heterogeneous graph.

[0061] In an alternative embodiment, a quaternary heterogeneous diagram This represents a heterogeneous topological network composed of four types of differentiated nodes and directed edges of multiple semantic types.

[0062] in, There are four types of nodes; This is a set of relation (edge) types, which can include nine directed relations: an individual works for a company, an individual invests in a company, an individual is related to family members or relatives, family members work for a company, family members invest in a company, a company is related to the legal representative of a key person, a company is related to the senior management of a key person, a company is related to the shareholder of a key person, and a key person is related to the supplier of a company.

[0063] Furthermore, different types of nodes have different attribute feature spaces, and different types of edges have different semantic meanings and decay characteristics.

[0064] Furthermore, the implicit correlation calculation method based on heterogeneous graph neural networks and spatiotemporal decay represents a method for multi-hop implicit relation reasoning through the message passing mechanism of heterogeneous graph neural networks, and introduces time decay and spatial decay factors in the correlation calculation to achieve dynamic quantitative evaluation of correlation strength.

[0065] Furthermore, the relationship edges are directed edges in the quaternary heterogeneous graph, used to represent the real or potential interest connections between nodes. They are divided into explicit relationship edges (documents, business data directly contain records of employment, shareholding, relatives, etc.) and implicit relationship edges (discovered by multi-hop reasoning of the model, without direct original records).

[0066] Furthermore, the comprehensive correlation strength value represents a standardized value obtained by integrating the six risk dimensions and incorporating the synergistic amplification effect of the dimensions, ranging from 0 to 1. In this embodiment, the risk is further divided into high, medium, and low levels based on the value; a higher value indicates a stronger connection of interests between the individual and the enterprise, and a greater risk of corruption.

[0067] In one alternative embodiment, internal personnel, family members, and external enterprises belong to different data sources, with significant differences in entity types and semantic relationships. Ordinary isomorphic graphs cannot distinguish the timeliness and geographical risks of diverse entities and differentiated relationships. Therefore, by constructing a quaternary heterogeneous network to isolate four types of nodes and nine types of relationships, and by using a heterogeneous graph neural network with spatiotemporal decay constraints to complete multi-hop information propagation and implicit link prediction, a unified quantitative comprehensive association strength value can be output for all explicit and implicit association edges within the graph.

[0068] Step S206: Based on the hierarchical structured personnel information dataset, external information set, and multiple comprehensive correlation strength values, generate a structured review report and interactive correlation graph for the company to be reviewed.

[0069] In one optional embodiment, the structured review report is output in Excel or CSV format using a dynamic column generation mechanism: fixed columns may include the internal personnel's name, position, department, source document path, and extraction confidence level; the dynamic columns are based on the number of family members of each internal personnel. Dynamically expand to Columns include (title, name, date of birth, political affiliation, work unit and position); related columns may include the name of the related external enterprise, unified social credit code, type of association, and strength of association. Risk level, reasoning path (complete multi-hop path of implicit association), evidence tracing link (original document path and specific page / line number, external data collection timestamp and data version hash).

[0070] Furthermore, the risk level is determined according to... The values ​​are divided into three levels: For high risk, For medium risk, Low risk. For example, ; .

[0071] In one optional embodiment, the relationship graph is presented as an interactive HTML page, with four types of nodes distinguished by different colors and shapes (internal personnel are blue circles, external enterprises are gray squares, key personnel are orange rhombuses, and family members are green triangles); the thickness of the edges is proportional to the relationship strength R; the color of the edges corresponds to the risk level (red / yellow / green); and the size of the nodes is proportional to the degree centrality of the nodes, which facilitates the rapid identification of key hub nodes in the graph.

[0072] Furthermore, the graph supports the following interactive operations: clicking on a node brings up a details panel (displaying all attributes and a table of relationships for that node), hovering over an edge displays the reasoning path and source of evidence, filtering by risk level, relationship type, and node type, searching and locating by company and personnel name, and exporting the current view as a PNG / SVG image file.

[0073] Furthermore, all displayed data can be traced back to the corresponding location in the original document with a single click, meeting the audit trail requirements for compliance review.

[0074] The enterprise personnel correlation analysis method provided in this embodiment solves the pain points of traditional manual review, such as inefficient multi-format document processing, disordered header fields, omission of nested personnel information, high cost of external data update, difficulty in identifying hidden interest relationships, inability to quantify and classify risks, and lack of audit traceability. It takes into account four core indicators: document parsing compatibility, information extraction accuracy, correlation mining depth, and compliance traceability. It is suitable for intelligent review of personnel correlation risks in all scenarios of enterprise cadres and personnel and supplier compliance.

[0075] In some optional implementations, step S202 above includes: Step S2021: Using the original binary document stream and the preset engine format adaptation matrix, the document format of the personnel document to be analyzed is adaptively identified, and multiple initial parsing engines that match the format of the personnel document to be analyzed are determined.

[0076] Traditional methods rely solely on the magic number at a fixed offset in the file header for format determination, which has two inherent drawbacks: first, they cannot correctly identify files when the file header is corrupted during transmission or intentionally tampered with; second, they cannot distinguish between .doc and .wps files that share the same OLE2 container structure, as both use 0xD0CF11E0 as the file header identifier, and deeper internal structural features are required for effective differentiation.

[0077] Specifically, step S2021 above includes: Step a1: Extract multiple byte windows of a preset length at multiple preset offset positions in the original binary document stream, and calculate multiple frequency distribution histogram vectors corresponding to the multiple byte windows.

[0078] In one optional embodiment, the frequency distribution histogram vector represents the 256-dimensional numerical vector obtained by counting the frequency of all byte values ​​from 0 to 255 within a single 256-byte sampling window and normalizing it. This vector is used to quantify the proportion of various byte types within the window and can intuitively reflect the local underlying byte distribution characteristics of the document.

[0079] For example, for the input document binary stream (Total length is) byte), in Extract the length at each preset offset position. byte window ( Take 5. (Take 256 bytes) to obtain the sample window set. .

[0080] in, Located at the beginning of the file (offset) ), lie in At the byte level, lie in At the byte level, Located in the middle of the document ( ), Located at the end of the document ( ).

[0081] Furthermore, the above segmented sampling strategy covers different structural levels of the file: the header contains the format magic number, the middle section contains document metadata and style definitions, the middle section contains the byte distribution characteristics of the text content, and the tail contains cross-reference tables or file directory structures.

[0082] Furthermore, for each sample window Calculate its 256-dimensional byte value frequency distribution histogram vector. The following relation (2) is shown:

[0083] In the formula: Indicates the first Frequency distribution histogram vector of a byte window; Indicates the first A byte sequence of a byte window; This indicates an indicator function; it returns 1 if the condition is true, and 0 otherwise. This represents the vector dimension index, corresponding to a value of 0 to 255 bytes.

[0084] Step a2 involves concatenating multiple frequency distribution histogram vectors and byte information entropy vectors within multiple byte windows to construct a composite fingerprint vector of the personnel document to be analyzed.

[0085] In an alternative embodiment, the byte information entropy vector is used to reflect the uniformity of byte distribution within the byte window.

[0086] In an alternative embodiment, The histogram vector of each window is concatenated with the byte information entropy feature of each window to form a composite fingerprint vector of the personnel document to be analyzed. The following relation (3) is shown:

[0087] In the formula: This represents a vector concatenation operation; This represents the byte information entropy vector for each window, where each information entropy... Defined as the following relation (4):

[0088] In the formula: Indicates avoidance The smoothing constant for numerical anomalies.

[0089] Furthermore, information entropy This reflects the uniformity of the byte distribution within the window: for .doc files (OLE2 composite documents with a multi-level storage structure of directory streams and data streams), the entropy values ​​of each window show an alternating pattern of high and low values; for .wps files (which contain a mixture of XML markup text and binary data streams), the entropy value of the header window is relatively low (mainly XML text), while the entropy values ​​of the middle and tail windows are moderately high.

[0090] Furthermore, by simultaneously applying histogram similarity constraints to the entropy distribution patterns of multiple windows, effective differentiation of heterogeneous formats sharing the same file header signature is achieved.

[0091] Step a3: Calculate multiple original weighted cosine similarities between the composite fingerprint vector and multiple pre-constructed standard fingerprint vectors.

[0092] In an alternative embodiment, for each known document format Pre-construct standard fingerprint vectors (The average value was obtained by extracting fingerprints from a large number of sample files of this format using the method described above).

[0093] Furthermore, for the document to be identified The weighted cosine similarity between the fingerprint and the reference fingerprint of each format is calculated as shown in the following equation (5):

[0094] In the formula: Indicates the document to be identified With format Weighted cosine similarity; This represents element-wise multiplication; The feature weight vector is designed according to the principle of "high weight in the first window, low weight in the last window, and weights enhanced by entropy components": [corresponding to...] The weights of the 256 histogram components (in the file header) are: The weights of the entropy components corresponding to each window are: The remaining components have a weight of 1.0.

[0095] Step a4: Determine the format and target weighted cosine similarity of the personnel document to be analyzed based on multiple original weighted cosine similarities.

[0096] In one alternative embodiment, the target weighted cosine similarity is the largest among multiple original weighted cosine similarities.

[0097] In one optional embodiment, all similarity values ​​are traversed, and the document type corresponding to the maximum value is selected as the judgment format. At the same time, the maximum similarity is recorded as the format recognition confidence level to determine whether the recognition result is reliable.

[0098] The determination formulas are shown in the following relations (6) and (7):

[0099] In the formula: Indicates the document to be identified The determination format number; This represents the target weighted cosine similarity (maximum similarity).

[0100] Step a5: When the target weighted cosine similarity is greater than or equal to the preset first threshold, the format of the personnel document to be analyzed is determined to be an identified format, and the engine adaptation score of the format of the personnel document to be analyzed is determined based on the preset engine format adaptation matrix.

[0101] In an alternative embodiment, the target weighted cosine similarity is... and preset first threshold Perform a comparison, if The document to be identified will be... Mark it as "unrecognized format" and forward it for manual confirmation.

[0102] Furthermore, if If the format of the personnel document to be analyzed is determined to be an identified format, the engine adaptation score of the format of the personnel document to be analyzed is determined using the preset engine format adaptation matrix.

[0103] Specifically, the preset engine format adaptation matrix is ​​dynamically maintained through a Bayesian online learning mechanism: Let the engine... In format The performance on follows a Beta distribution. The initial prior is (Uniform prior), the posterior parameters are updated based on manual review feedback after each parsing, as shown in the following relationships (8) and (9):

[0104] Then the engine compatibility score .

[0105] Step a6: Compare the engine compatibility score with the preset second threshold, and determine multiple initial parsing engines that match the format of the personnel document to be analyzed based on the comparison results.

[0106] In an optional embodiment, for the format determined as The document in question is the personnel document to be analyzed, which has already been identified in its format. Priority will be given to scheduling the document with the highest fit score that is not lower than the preset second threshold. Top-N engines ( (instead of indiscriminately calling all engines),

[0107] Furthermore, when hour( Taking 0.5 indicates that this format lacks a highly adaptable engine, so all scheduling is performed. Multiple engines are used to ensure coverage.

[0108] Step S2022: Use multiple initial parsing engines to parse the personnel documents to be analyzed, and obtain the parsing results of multiple engines.

[0109] In an alternative embodiment, the parsing result of each engine includes a text content matrix and a layout feature vector.

[0110] In an alternative embodiment, the scheduled One engine ( Parallel processing of documents Perform analysis.

[0111] Furthermore, each engine Output its parsing results It can include a text content matrix ( To identify the row number, To identify the number of columns and the layout feature vector consisting of three indicators. .

[0112] Among them, text density Defined as the proportion of non-empty cells in the matrix; row height uniformity. Defined as 1 minus the normalized value of the standard deviation of the text length of each line; column width alignment. The average Pearson correlation coefficient, defined as the text length vectors between adjacent columns, is used to measure the regularity of table column division. These three metrics together reflect the quality of the engine's table structure recovery.

[0113] Step S2023: Using the theory of evidence reasoning, the parsing results of multiple engines are fused and processed to generate a document object model tree structure and a JSON text matrix set of the personnel documents to be analyzed.

[0114] In one alternative embodiment, the Evidential Reasoning (DS Reasoning) theory represents an imprecise reasoning method that allows for probability allocation and synthesis of a set of hypotheses when the evidence is incomplete, imprecise, or conflicting. Furthermore, this theory defines the credibility interval of propositions through a trust function and a likelihood function, and utilizes orthogonal sums and rules to fuse multi-source evidence.

[0115] Furthermore, the Document Object Model (DOM) tree structure represents a tree-like structured carrier that organizes paragraphs, tables, cells, nested sub-tables, table headers, and text blocks within a document according to hierarchical parent-child relationships, and can completely restore the original document layout, table nesting, and cell merging relationships.

[0116] Furthermore, the JSON text matrix collection represents a two-dimensional table matrix dataset stored in JSON serialization format. Each matrix corresponds to a table parsed by the engine, and the matrix rows and columns correspond one-to-one with document cells, used to store cell text, position, and merge attributes.

[0117] Furthermore, this embodiment employs evidence-based reasoning theory to achieve confidence fusion of multi-engine parsing results, overcoming the failure problem of traditional weighted voting methods when faced with highly conflicting engine results.

[0118] Specifically, step S2023 above includes: Step b1: Generate local quality evidence vectors for each target parsing engine according to multiple preset evaluation dimensions, and construct the basic probability allocation function for each initial parsing engine.

[0119] In one optional embodiment, multiple preset evaluation dimensions represent four standardized evaluation indicators for uniformly measuring the quality of a single-engine parsing output, which may include text integrity. (Percentage of non-empty text), table structure retention (Whether the number of rows and columns is within a reasonable range), character encoding compliance rate (Percentage of valid Unicode characters), Detection rate of key fields (Preset keyword hit rate).

[0120] Furthermore, the Basic Probability Assignment (BPA) function represents a mapping defined in evidence reasoning theory on the power set of the recognition framework, used to directly assign trust to any subset rather than just a single-element set.

[0121] In an alternative embodiment, for the engine Combining the four evaluation dimensions mentioned above, we can define its role in the analysis results. Local quality evidence vector on .

[0122] Furthermore, the basic probability assignment function is constructed by combining the local quality evidence vector, as shown in the following equations (10) to (12):

[0123] In the formula: Represents the Sigmoid function; Represents the recognition framework; This represents the quality of uncertainty, i.e., the probability of uncertainty in the analytical result; Indicates the weight of each evaluation dimension; Indicates engine No. Scores for each evaluation dimension; Indicates the weight of engine compatibility; Indicates engine The score for adaptability to the current document format; Indicates the weight of logical consistency loss; This represents the logical consistency loss term, used to measure the internal consistency of the parsing results. For example, the loss term increases when the identified ID number field fails the checksum verification. , , The determination can be optimized on the validation set using grid search; This represents the basic probability of reliable single-engine parsing. This represents the base probability that the engine's parsing is unreliable.

[0124] Step b2: Based on each basic probability allocation function, calculate the basic confidence probability value and basic unconfidence probability value of each engine's parsing result, and determine the basic confidence probability distribution of each engine's parsing result.

[0125] In an optional embodiment, the basic confidence probability value of each engine's parsing result can be calculated using the above relation (10). Meanwhile, the basic unreliability probability value of each engine's parsing result can be calculated using the above relationship (11). .

[0126] Furthermore, this is combined with the calculated basic credibility probability value. and the basic unbelievability probability value The basic confidence probability distribution of each engine's parsing result can be determined through the above relation (12). .

[0127] Step b3: Using Dempster's combination rules, the basic confidence probability distribution of each engine's parsing results is recursively combined pairwise, and the global fusion confidence value is determined.

[0128] In an optional embodiment, the Dempster combination rule represents the mathematical formula in evidence reasoning theory used to fuse two independent basic probability assignments (BPAs). After orthogonal summation and elimination of conflicting parts, it is normalized to generate a new BPA, as shown in the following relation (13):

[0129] In the formula: , This represents the BPA probability distribution of two independent engines; , This represents the subset of the identification framework corresponding to each of the two pieces of evidence: Taken from the entire subset of the recognition framework of the first engine , Taken from the entire subset of the second engine's recognition framework ; This represents the term where two pieces of evidence are in complete conflict, and the denominator is used to normalize the probability of canceling out the conflict. This represents the subset after the two pieces of evidence are merged. The new probability assignment value.

[0130] Furthermore, using the above relation (13), for all The BPA of each engine is recursively combined pairwise to obtain the final confidence score. .

[0131] Step b4: When the global fusion confidence value is less than the preset fusion threshold, obtain the highest basic confidence probability value, and generate a document object model tree structure and a JSON text matrix set based on the engine parsing result corresponding to the highest basic confidence probability value.

[0132] In an optional embodiment, the global fusion confidence value is... With preset fusion threshold Perform a comparison, if If so, all engine parsing results of the personnel documents to be analyzed will be marked with a "low confidence" tag.

[0133] Furthermore, the engine parsing result corresponding to the highest "reliable" confidence level (i.e., the highest basic confidence probability value) in BPA is selected as the main result. Based on the text content matrix, layout, and cell position information contained in the engine parsing result, a standardized document DOM tree is constructed in layers, and the table, nesting, and cell hierarchy relationships are fully preserved.

[0134] Furthermore, all the two-dimensional matrices and cell attributes of the tables are serialized into a standard JSON dataset, forming a collection of JSON text matrices.

[0135] Step S2024: Perform a depth-first recursive traversal of the document object model tree structure and the JSON text matrix set, and construct a directed acyclic graph model.

[0136] In one alternative embodiment, depth-first recursive traversal represents a tree structure traversal algorithm that starts from the root node of the DOM tree, recursively visits sub-tables and nested cells layer by layer downwards, until the plain text leaf node is reached, and then backtracks to the upper layer.

[0137] In one alternative embodiment, starting from the root node of the document DOM tree, a table container node is created for each table node, and all its rows and cells are traversed. If a cell contains nested sub-tables, then that cell is marked as... It recursively calls itself to process sub-tables, increments the recursion level by 1, and sets the parent identifier to the current table identifier.

[0138] Furthermore, the recursion terminates at plain text paragraph nodes (leaf nodes). The algorithm automatically identifies merged cells ( or ), mark it as The merge span is recorded. Furthermore, all text content is standardized (removing extra whitespace, converting full-width and half-width characters, and converting traditional characters to simplified characters).

[0139] For example, load the DOM tree structure, the entire JSON text matrix collection, and create an empty node collection. empty directed edge set This is used to store all nodes and nested hierarchical relationships in the table graph. Furthermore, all cell text is preprocessed uniformly: extra blank spaces are removed, full-width / half-width characters are converted, and traditional Chinese characters are standardized to simplified Chinese.

[0140] Furthermore, it reads all rows and cells of the current table and generates an independent global table for the current table. As the parent container number Then, parse each cell one by one and extract the rows. Cross-column And determine the cell type: if or Then mark If a cell contains a sub-table DOM branch, mark it as such. If a cell contains only text and no embedded table, distinguish between header cells and data cells.

[0141] Furthermore, assemble the current cell node according to the seven-tuple formula. And store it in the node set. Furthermore, if the cell is a nested container... Then a directed edge is created from the parent table to the child table. And store it in the directed edge set. Furthermore, nesting depth .

[0142] Furthermore, if a cell contains only plain text, it is determined to be a leaf node, the current branch recursion is terminated, and the process backtracks to the upper-level table to continue traversing the remaining cells.

[0143] Furthermore, after the recursive traversal is completed, each node in the traversal node set... and along Tracing back to the top-level master table, and simultaneously concatenating the complete hierarchical ID link according to the path encoding formula, the result is stored in the node metadata field. This enables quick hierarchical positioning in any cell.

[0144] Among them, to support downstream modules Time complexity of node-level location, for each node Hierarchical path encoding is assigned, and the path encoding formula is shown in the following relation (14):

[0145] In the formula: Represents the complete source path of a node, supports bidirectional fast location of nested level cells, and has a low query complexity. ; Indicates the ID of the top-level main table; This indicates the table ID of the current node's level.

[0146] Furthermore, the encoding records the identifier sequence of each level table along the complete path from the root node to node v, supporting bidirectional tracing along the path.

[0147] Furthermore, traversing all nodes The maximum value is used as the global nesting depth of the entire table chart. Then combine the node set Set of directed edges Generate a standard nested table graph model, i.e., a directed acyclic graph model. .

[0148] In some optional implementations, step S2023 above further includes: Step b5: When the global fusion confidence value is greater than or equal to the preset fusion threshold, calculate the probability difference between the highest and second-highest basic probability values ​​based on the basic confidence probability values ​​of the parsing results of each engine.

[0149] Step b6: When the probability difference is greater than or equal to the preset third threshold, generate a document object model tree structure and a JSON text matrix set based on the engine parsing result corresponding to the highest basic confidence probability value.

[0150] Step b7: When the probability difference is less than the preset third threshold, the document object model tree structure and JSON text matrix set are generated after fine-grained completion of the engine parsing result corresponding to the highest basic confidence probability value, using the basic confidence probability value of each engine parsing result as the weight.

[0151] In an alternative embodiment, if Based on the calculated base confidence probability value of each engine's parsing result. Determine the highest base probability and the second highest base probability, and then calculate the probability difference between the highest base probability and the second highest base probability.

[0152] Furthermore, this probability difference is compared with a preset third threshold ( The comparison is performed, and if the probability difference is greater than or equal to the preset third threshold... If the result is zero, it means that there are no multiple engines with comparable results. In this case, the parsing result of the engine corresponding to the highest basic confidence probability value is directly used to generate the document object model tree structure and the JSON text matrix set. The specific generation process can be found in the relevant description in step b4 above, and will not be repeated here.

[0153] Furthermore, if the probability difference is less than a preset third threshold This indicates that multiple engines have similar results. In this case, multi-engine fine-grained completion is triggered: for each missing cell in the main result (i.e., the engine parsing result corresponding to the highest basic confidence probability value), candidate values ​​are extracted from the corresponding positions in other engines, and a weighted majority vote is used (the weight is the percentage of each engine's score). Determine the final fill value.

[0154] Furthermore, the engine parsing result corresponding to the highest base confidence probability value after completion is used to generate a document object model tree structure and a JSON text matrix set. The specific generation process can be found in the relevant description in step b4 above, and will not be repeated here.

[0155] In some optional implementations, step S203 above includes: Step S2031: Based on the directed acyclic graph model, the header recognition and error correction method based on bidirectional attention flow and global optimal allocation is used to obtain the global optimal allocation result of the personnel document to be analyzed.

[0156] Existing header recognition methods generally adopt a one-way matching strategy of "header text → standard category", that is, calculating the similarity between header keywords and standard category keywords and taking the maximum value. This approach has three fundamental drawbacks: (a) it ignores the "reverse verification" information of the data content of the header column on the header semantics - when the actual data in the header "gender" column is a title such as "father" or "mother", relying solely on one-way matching of header text cannot detect the error; (b) the category assignment of each header is independent of each other and lacks global consistency constraints - it is possible for two different headers to be assigned to the same standard category; (c) it does not make sufficient use of the header context (the semantic mutual exclusion relationship between adjacent headers in the same row).

[0157] This embodiment proposes a header recognition and error correction method based on bidirectional attention flow and global optimal assignment, which can fundamentally solve the above three defects.

[0158] Specifically, step S2031 includes: Step c1 involves extracting the directed acyclic graph model to obtain multiple header texts from multiple header cells of the personnel document to be analyzed.

[0159] Step c2 involves mapping multiple header texts and multiple preset standard categories through a pre-trained semantic encoder to obtain multiple first semantic vectors and multiple second semantic vectors.

[0160] Step c3: Based on multiple first semantic vectors and multiple second semantic vectors, use forward attention flow to calculate the first attention weight value of each header text for each preset standard category.

[0161] Step c4: Calculate the matching score of each column of data in each header text with each preset standard category of data type, and determine the aggregated content feature vector of each column of data.

[0162] Step c5: Based on multiple aggregated content feature vectors and multiple second semantic vectors, the back attention flow is used to calculate the second attention weight value of each column of data for each header text.

[0163] Step c6 involves weighted fusion of multiple first attention weight values ​​and multiple second attention weight values, and constructing an optimization function.

[0164] Step c7: Within the preset time complexity, use the Hungarian algorithm to solve the optimization function and determine the globally optimal header allocation result of the personnel document to be analyzed.

[0165] In one alternative embodiment, a pre-trained semantic encoder (such as RoBERTa-zh) represents a deep neural network model based on a Transformer encoder architecture that learns general bidirectional contextual representations on large-scale unlabeled corpora through self-supervised tasks (such as masked language modeling) and can be transferred and fine-tuned to downstream NLP tasks.

[0166] Furthermore, the optimization function represents the minimum total cost of allocation.

[0167] Furthermore, the Hungarian algorithm (Kuhn-Munkres algorithm) represents a combinatorial optimization algorithm for solving the task allocation problem in polynomial time.

[0168] In an optional embodiment, the header row of the document table contains Header text The standard header category set is (common The categories include name, gender, ethnicity, date of birth, place of origin, political affiliation, education and degree, position, rank, professional and technical title, ID number, contact number, home address, family member titles, family member names, family member ages, family member political affiliation, family member work unit and position, etc.

[0169] Furthermore, each standard category Associate a keyword extension set The comprehensive semantic vector, i.e., the second semantic vector, is obtained by mapping through a pre-trained semantic encoder (such as RoBERTa-zh). At the same time, the text of each header will be... is mapped to a first semantic vector through a pre-trained semantic encoder (such as RoBERTa-zh) . Wherein, represents the dimension of the semantic vector.

[0170] Further, based on the plurality of first semantic vectors and the plurality of second semantic vectors, forward attention flow is utilized to calculate the header text for each standard category the attention weight, i.e., the first attention weight value, whose physical meaning is "the header semantically has a certain probability of belonging to the category ", as shown in the following relational expression (15):

[0171] In the formula: represents learnable query and key projection matrices; represents a scaling factor; represents the forward attention weight of the header for the standard category .

[0172] Further, the distribution of the forward attention weights reflects the semantic attribution tendency of the header text for each standard category.

[0173] Further, assume that the column where the header is located contains rows of data . First, the entire column of data is aggregated into a column content representation vector. Specifically, for each row of data , the data type matching score between it and each standard category is calculated, and a category tendency vector of the data item is constructed , as shown in the following relational expression (16):

[0174] In the formula: represents a data format verification function based on regular expressions and a rule base. For example, for the category "birth date", it is verified that conforms to the date format (such as "YYYYMM", "YYYY-MM-DD", "YYYY.MM", etc.); for the category "gender", it is verified that it exactly matches "male" or "female"; for the category "ID number", it is verified that it conforms to the 18-digit format and passes the ISO 7064:1983 MOD 11-2 check digit verification; for the category "ethnic group", it is verified that it matches the dictionary of 56 ethnic group names.

[0175] Further, the category tendency vectors of rows are averaged to obtain the aggregated content feature vector of this column .

[0176] Furthermore, based on multiple aggregated content feature vectors and multiple second semantic vectors Using backward attention flow, the attention weight of column data content to table header hi is calculated, which is the second attention weight value. Its physical meaning is that the "actual data content of the column" corrects the "degree of semantic belonging of the table header" in reverse, as shown in the following relation (17):

[0177] In the formula: This represents a projection matrix independent of the forward flow (allowing the model to learn differential matching patterns between header text and column data). Indicates column data against standard categories Backward attention weights.

[0178] Furthermore, the forward and backward attention flows are weighted and fused to obtain the table header. With category Overall matching score The following relation (18) is shown:

[0179] Where: parameters , The contribution ratios of forward attention, backward attention, and context constraints are determined through cross-validation on a validation set (in this embodiment, we take...). , ); The context consistency reward term is represented by the following relation (19):

[0180] In the formula: Indicates table header The set of header indices of the peer neighbor list.

[0181] Furthermore, the physical meaning of this reward is: if the adjacent head... Category Forward attention weights Very high, indicating If it is already occupied by the head of the adjacent list, then decrease. Assigned to The score introduces a mutual exclusion constraint for category assignment.

[0182] Furthermore, based on the comprehensive matching score Build dimensional allocation cost matrix The following relation (20) is shown:

[0183] In the formula: Indicates the conflict penalty item (in this embodiment, it is taken as...) ); Returns true in the following cases: The original label (such as the original header text that can be determined not to belong to this category by preliminary rule judgment) and There are known mutual exclusion relationships, such as when the original header is clearly "date of birth" but the candidate category is "ethnicity", a conflict penalty is triggered.

[0184] Furthermore, the problem of assigning table headers to standard categories is formalized as a minimum weight matching problem in a bipartite graph: in a bipartite graph In the middle, find a double shot (or partial match, when) When the total cost of allocation is minimized, the optimization function is shown in the following relation (21):

[0185] Furthermore, the Kuhn-Munkres algorithm (Hungarian algorithm) is used in... The above optimized function can be solved exactly within the time complexity.

[0186] Furthermore, for In this case, introduce Each dummy header has a cost of 0, ensuring that the output of the Hungarian algorithm contains only [number of dummy headers]. One header is assigned to the actual category, and the rest... One category was not assigned.

[0187] Furthermore, when the globally optimal allocation Once determined, for each table header If its allocation cost (This embodiment takes) If the misalignment is detected, a misalignment correction alarm will be triggered: Check Is it greater than the threshold? (Take 0.2), if so, then... Replace it with the category with the second-highest score and mark the entry as "requires manual review". Furthermore, record the mapping relationship between the original header text and the final assigned category, along with the operation log, to support audit traceability and obtain the globally optimal header assignment result for the personnel documents to be analyzed.

[0188] Furthermore, by combining the aforementioned bidirectional attention flow with global optimal allocation, the following are achieved: (a) mutual verification between the semantics of the header text and the content of the column data; (b) mutual exclusion constraints between adjacent headers; and (c) globally consistent optimal category allocation for all headers. Furthermore, this joint optimization at three levels significantly improves the robustness of header recognition and error correction.

[0189] Step S2032: Construct a multi-dimensional position encoding vector for each node in the directed acyclic graph model, and calculate the topological context vector for each node in the directed acyclic graph model.

[0190] In one alternative embodiment, in the constructed directed acyclic graph model Based on this, for each node Construct its multidimensional positional encoding vector It can include encoded information in four dimensions: 1. Absolute position encoding The sine-cosine position encoding function is used to encode the row and column coordinates of the nodes. Mapped to 3D continuous vector; 2. Hierarchical depth coding : Nesting depth Through learnable embedding layers mapped to dimensional vector; 3. Cell type encoding :Will (header / data / nested_table / merged) after one-hot encoding and linear projection is dimensional vector; 4. Merge span codes Merge row and column spans Through learnable embedding layers mapped to Dimensional vector.

[0191] Furthermore, the final position encoding The concatenation of the encoded information for the above four dimensions is shown in the following relation (22):

[0192] Furthermore, to capture the topological constraints of nested tables, a topological context vector is computed for each node. , is defined as the chain from the node to its ancestor node (along... The aggregation of type labels at each level, tracing back to the root, is shown in the following relation (23):

[0193] In the formula: This indicates a learnable embedding layer of a hierarchical type, used to convert tabular data into model-computable numerical features; Indicates from the root node to The first path The type of layer node.

[0194] Furthermore, the topological context vector provides the model with global semantic signals about the nested environment of the cell. For example, the topological context vectors of cells located in the "Family Members" nested table are significantly different from those of cells located in the "Basic Information" area of ​​the main table, which can guide the model to adopt different information extraction strategies.

[0195] Step S2033: Generate the initial input representation vector for each node based on the multidimensional location encoding vector and topological context vector of each node.

[0196] In an alternative embodiment, the initial input representation vector of a node is represented as a fusion of text embedding and positional encoding, i.e., combining the multidimensional positional encoding vector of each node. and topological context vector It can generate the initial input representation vector for each node. The following relation (24) is shown:

[0197] In the formula: Represents a node The text content; express Contextual semantic vectors encoded by a pre-trained BERT model; This indicates a splicing operation.

[0198] Step S2034: Based on the edge type awareness mechanism, the initial input representation vector of each node is processed by the graph attention network to obtain the target hidden layer representation vector of each node that integrates neighborhood information.

[0199] In one optional embodiment, an edge type awareness mechanism is introduced during information propagation between nodes to distinguish the following four edge types: spatial adjacency edges (adjacent cells vertically, horizontally, and vertically, with type labeled as...). ), Alignment margins in the same column (cells in the same column, type marked as) ), and the same row alignment (cells in the same row, with the type marked as ). ) and nesting hierarchy edges (between nested table containers and the content of the nested table, type marked as ).

[0200] Furthermore, different edge types use independent attention parameters, enabling the model to aggregate information differently based on relation type.

[0201] Furthermore, for the first Layered graph attention network, nodes The update rule is shown in the following relation (25):

[0202] In the formula: Indicates the number of heads of multi-head attention ( Take 8); Represents a set of edge types; Represents a node In edge type The set of neighbors below; Indicates the first Size in the side type The learnable transformation matrix below; The attention coefficient is represented by the following relationship (26):

[0203] In the formula: Representing edge type Learnable Relation Embedding; This represents the attention vector parameters.

[0204] Furthermore, by explicitly embedding edge types into the attention computation, the model can learn the prior knowledge that the information transmission method on nested hierarchical edges should be different from that of spatially adjacent edges.

[0205] Furthermore, after Layered graph convolution ( After taking step 3), each node The final hidden layer representation with fused neighborhood information was obtained. .

[0206] Step S2035: Based on the global optimal allocation result of the table header and the preset nested topological constraints, the target hidden layer representation vector of each node is used to perform multi-task decoding of the personnel documents to be analyzed, and generate a hierarchical structured personnel information dataset.

[0207] In an optional embodiment, the preset nested topology constraint represents a table-level path encoding. Node nesting depth The extraction constraints are constructed to distinguish between basic information cells in the main table and nested sub-table cells containing family members.

[0208] In an optional embodiment, the target hidden layer representation vector of each node is obtained. Then, multi-task decoding is performed on it, and the entity label of each node is jointly predicted (using the BIO annotation scheme, entity types include PERSON, TITLE, ORG, DATE, GENDER, ETHNIC, POLITICAL, ID_CARD, PHONE, ADDRESS, RELATION, etc.).

[0209] Specifically, for nodes The conditional probability distribution of its entity labels The following relation (27) is shown:

[0210] in, To model the transition constraints between adjacent node labels for node-level conditional probabilities, a Conditional Random Field (CRF) layer is used. This represents a BIO entity tag.

[0211] Furthermore, the CRF emission score is obtained by linear projection of the node hidden layer representation, as shown in the following equation (28):

[0212] in, Represents a node Predicting entity labels The launch score, in different side types Using different transition matrices This is because the label transition pattern on nested hierarchical edges (such as transitioning from the "RELATION" label to the "PERSON" label) is fundamentally different from the transition pattern on spatial adjacency edges (such as transitioning from the "PERSON" label to the "PERSON" label, indicating that the name may span multiple columns). Indicates label The dedicated linear projection matrix; This indicates the label bias.

[0213] Finally, the globally optimal label sequence is obtained through Viterbi decoding, and the extracted entity information is assembled in a nested hierarchical structure according to the hierarchical path encoding shown in the above relation (14), forming a corresponding hierarchical structured personnel information dataset. Further, after extraction, data quality verification is performed, including ID card number verification, date validity verification, and temporal logical consistency verification between the year of birth and the year of employment.

[0214] Furthermore, unlike traditional methods that process each row of a table independently into a flat sequence before performing named entity recognition (NER), this embodiment models the table as a graph structure—each cell of the table is treated as a node, and the spatial adjacency (adjacent vertically, horizontally, and vertically), hierarchical nesting (between the nested table container and the nested table), and row and column alignment (same row, same column) between cells are treated as different types of edges. Then, a graph attention network (GAT) is used to extract node-level information on this heterogeneous graph, and nested topological constraints are used to ensure the consistency of the information extracted from the nested table with its host context, thus solving the problem of accurately extracting enterprise personnel information and family member information in nested table structures.

[0215] In some optional implementations, step S205 above processes the quaternary heterogeneous graph using a method for calculating implicit correlation based on heterogeneous graph neural networks and spatiotemporal decay to obtain the comprehensive correlation strength value of each correlation edge in the quaternary heterogeneous graph, including: Step d1: Based on the preset time decay function and the preset spatial decay function, the quaternary heterogeneous graph is processed by heterogeneous graph convolution to obtain the target embedding representation vector of each node in the quaternary heterogeneous graph that integrates multi-hop heterogeneous neighborhood information.

[0216] In an alternative embodiment, the time decay function It describes the domain knowledge of the decay of the time-related relationship: the longer the relationship has been established, the lower its reference value for current risk assessment.

[0217] Specifically, set the edge The corresponding related event occurred at the following time points: The current review benchmark time is The time span is (in months). Furthermore, different types of relationships use different decay rates, as shown in the following relationship (29):

[0218] In the formula: Indicates association type The corresponding time decay coefficient takes a value of (0, 1], and the larger the value, the stronger the correlation with timeliness. Nine types of business relationships are indicated: personal employment, personal investment, kinship, family member employment / investment, legal entity relationship, senior management relationship, shareholder relationship, and supplier relationship; Representing relation type The time decay half-life parameter (in months) can be set according to different business needs, such as the half-life of the employment relationship. The investment relationship may have a lasting impact even after the employee leaves the company, lasting for several months (5 years); the investment relationship half-life. Months (3 years), because equity changes are usually more frequent; supplier partnership half-life Two months (2 years) is sufficient because supplier partnerships are time-sensitive. Furthermore, through differentiated half-life design, the model can naturally depict the domain pattern that "different types of relationships have different durations of time sensitivity."

[0219] Furthermore, the reset effect of time decay on change events is introduced. Specifically, for relationship types with historical change records (such as equity changes and executive changes), if in If any changes occurred during this period, the time span starts from the time of the most recent change. The calculation begins from the initial association time point, as shown in the following relationship (30):

[0220] Furthermore, this correction mechanism ensures that even if an investment relationship initially occurs early on, as long as there is a recent change in shareholding (increase, decrease, or transfer), the time decay of the relationship will be reset, maintaining a high timeliness weight in the model.

[0221] In an optional embodiment, a preset spatial decay function is used. The weakening effect of geographical distance on the strength of associations is characterized: the closer the geographical locations, the greater the possibility of informal interest associations between individuals and enterprises (such as private exchanges within the same city or the transfer of benefits based on geographical proximity), while the opportunity cost of informal interactions is higher and the strength of associations is relatively lower for geographically dispersed relationships. Furthermore, the spatial decay function is presupposed to be defined in Gaussian kernel form, as shown in the following relation (31):

[0222] In the formula: Represents a node and The spherical distance between associated geographic locations (based on great circle distance in latitude and longitude). Representing relation type Spatial attenuation bandwidth parameters, The smaller the distance, the more drastic the decrease in correlation strength due to spatial distance.

[0223] Furthermore, regarding family member relationships, Choose a smaller value (e.g., 50 kilometers) because the connection between family members is relatively less affected by geographical distance (even if they live in different places, the kinship still exists); for supplier partnerships... Choose a larger value (e.g., 500 km) because supplier partnerships are highly dependent on geographical accessibility.

[0224] In an alternative embodiment, in the heterogeneous diagram When passing messages, different edge types use independent transformation matrices and attention mechanisms.

[0225] Specifically, for edge types Source node To the target node The message being transmitted is defined by the following relation (32):

[0226] In the formula: Representing relation type Down Transmit to The feature message vector; Representing relation type Learnable embeddings; This represents a message transformation multilayer perceptron specifically for edge types.

[0227] Furthermore, the target node In the The layer is represented by aggregating the message updates of all incoming edges, as shown in the following relation (33):

[0228] In the formula: Represents the self-loop characteristic transformation matrix of a node; Indicates the relationship point to The set of source nodes; Represents the set of all 9 types of relation edges; Representing an edge In relation The attention weights are introduced in this embodiment by introducing a spatiotemporal decay factor, which is defined as the product of the basic attention weights and the time and space decay factors, as shown in the following relationship (34):

[0229] In the formula: The standard scaled dot product attention weights are calculated from the similarity of the hidden layer representations of the source and target nodes.

[0230] Furthermore, based on preset time decay functions and preset spatial decay functions, the quaternary heterogeneous graph is processed... Layer heterogeneous graph convolution processing ( (Choose 3), each node can obtain the final embedding representation, i.e., the target embedding representation vector, which incorporates information from multi-hop heterogeneous neighborhoods. .

[0231] Step d2 involves using the bilinear mapping of the target embedding representation vector of each node to perform link prediction, thereby obtaining the predicted probability value of each edge in the quaternary heterogeneous graph.

[0232] In an alternative embodiment, link prediction is performed using a bilinear mapping of node embedding vectors, which can discover latent associations in the graph that have not yet been connected, as shown in the following relation (35):

[0233] In the formula: Represents a node , There is a hidden relationship The predicted probability; Representing relation type A dedicated bilinear scoring matrix.

[0234] Step d3: Determine the multiple association edges of the quaternary heterogeneous graph based on the predicted probability value of each edge.

[0235] In an optional embodiment, based on the calculated predicted probability value This can distinguish between explicit and implicit related edges.

[0236] For example, when predicting probabilities (This embodiment takes) When ), the edge Mark as latent association discovery and record the reasoning path. This reasoning path provides reviewers with a complete and interpretable chain of evidence.

[0237] Furthermore, by merging explicit related edges and implicit related edges that meet the threshold, we can obtain the set of all valid related edges in the graph, i.e., multiple related edges.

[0238] Step d4: Calculate the score of each association edge on multiple preset association dimensions and determine the comprehensive association strength value of each association edge.

[0239] In one optional embodiment, for each discovered (explicit or implicit) association relationship, i.e., the association edge, its association score across six dimensions is calculated and weighted, and the overall association strength is obtained. The following relation (36) is shown:

[0240] Where: coefficient and It can be obtained by training on historical labeled data through L1 regularized logistic regression. The regularization strength is selected through cross-validation to control model complexity and prevent overfitting.

[0241] Furthermore, the six dimensions of association are as follows: Name similarity dimension (including exact and fuzzy matching of the names of the individual and their family members with key personnel of external companies). Unique identifier matching dimension (exact match of ID card number or unified social credit code); Job-related dimension (the degree of overlap between the scope of authority of internal personnel and the business scope of external enterprises). Time overlap dimension (the ratio of the overlap between the on-the-job time of internal staff and the operating time of external enterprises); Investment-related dimension (shareholding ratio of the individual and family members in external companies, truncated after min normalization). Address overlap dimension (geographical distance between the registered or permanent address of the individual / family member and the registered address of an external company, normalized by Gaussian decay).

[0242] Furthermore, second-order interaction terms It captures the synergistic amplification effect between dimensions. For example, highly similar names (high... And the investment and shareholding ratio is high (high) The combined risk of the two dimensions is far greater than the simple linear sum of the risks of the two dimensions.

[0243] The enterprise personnel correlation analysis method provided in this embodiment has the following beneficial effects: 1. Strong compatibility with multiple formats: Supports multiple document formats such as .doc, .docx, PDF, and Excel, and automatically completes format conversion; 2. Comprehensive table parsing: Employing a recursive traversal algorithm, it can identify top-level tables and nested tables, avoiding information omissions; 3. Intelligent header correction: When fields such as title, position, professional title, gender, and ethnicity are filled in, the system automatically recognizes the semantic meaning of the content and matches the correct header, greatly improving the accuracy of information extraction. 4. Robust fault tolerance mechanism: Supports both WPS and Word document conversion methods, and attempts multiple COM names to ensure successful document conversion; 5. Strict data validation: Name length validation, Chinese character validation, and kinship keyword validation ensure the quality of extracted data; 6. Dynamic column generation: Dynamically generate Excel columns based on the number of the largest family members, adapting to different document structures; 7. Resource Management Standards: Ensure document objects are closed correctly, and automatically clean up temporary files to avoid consuming system resources; 8. Accurate identification of relationships: Intelligent algorithms are used for personnel matching and relationship analysis to improve the accuracy of relationship identification; 9. Strong real-time early warning capability: Supports dynamic monitoring and real-time early warning, enabling timely detection of potential risks; 10. Good traceability: Provides a function to trace the source of relationships, records the header to correct mapping relationships, and facilitates compliance review and audit.

[0244] This embodiment also provides an enterprise personnel correlation analysis device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated for details already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0245] This embodiment provides a device for analyzing the correlation between enterprise personnel, such as... Figure 3 As shown, the device includes: The first acquisition module 301 is used to acquire the original binary document stream of the personnel documents to be analyzed from the enterprise to be reviewed and the preset engine format adaptation matrix. The preset engine format adaptation matrix contains multiple elements that represent the historical average parsing quality scores of different parsing engines for documents of different formats.

[0246] The first processing module 302 is used to construct a directed acyclic graph model of the personnel documents to be analyzed based on the original binary document stream and the preset engine format adaptation matrix, through a multi-format document unified representation conversion method based on multi-engine adaptive fusion.

[0247] The second processing module 303 is used to perform global optimal allocation and information extraction on the directed acyclic graph model to obtain the global optimal allocation result of the table header of the personnel document to be analyzed and the hierarchical structured personnel information dataset.

[0248] The second acquisition module 304 is used to acquire the external information set of the enterprise to be reviewed from external public data sources by using an incremental acquisition strategy based on change event-driven and a retry mechanism with exponential backoff.

[0249] The third processing module 305 is used to construct a quaternary heterogeneous graph of the personnel documents to be analyzed, and to process the quaternary heterogeneous graph through a method for calculating the implicit correlation degree based on heterogeneous graph neural network and spatiotemporal decay, so as to obtain the comprehensive correlation strength value of each correlation edge in the quaternary heterogeneous graph.

[0250] The generation module 306 is used to generate a structured review report and an interactive relational graph of the enterprise to be reviewed based on the hierarchical structured personnel information dataset, external information set and multiple comprehensive correlation strength values.

[0251] The enterprise personnel correlation analysis device provided in this embodiment of the invention can execute the enterprise personnel correlation analysis method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the above modules are the same as in the corresponding embodiments described above, and will not be repeated here.

[0252] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0253] The following is a detailed reference. Figure 4 This diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 401, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 402 or a program loaded from memory 408 into random access memory (RAM) 403. The RAM 403 also stores various programs and data required for the operation of the electronic device. The processor 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0254] Typically, the following devices can be connected to I / O interface 405: input devices 406 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 407 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 408 including, for example, magnetic tapes, hard disks, etc.; and communication devices 409. Communication device 409 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 4 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0255] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 409, or installed from a memory 408, or installed from a ROM 402. When the computer program is executed by the processor 401, it performs the functions defined in the enterprise personnel correlation analysis method of the embodiments of the present invention.

[0256] Figure 4The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0257] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium after being downloaded via a network. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the enterprise personnel correlation analysis method shown in the above embodiments is implemented.

[0258] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0259] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for analyzing the correlation of personnel within an enterprise, characterized in that, The method includes: Obtain the raw binary document stream of the personnel documents to be analyzed from the company to be reviewed and the preset engine format adaptation matrix. In the preset engine format adaptation matrix, multiple elements represent the historical average parsing quality scores of different parsing engines for documents of different formats. Based on the original binary document stream and the preset engine format adaptation matrix, a directed acyclic graph model of the personnel document to be analyzed is constructed through a multi-format document unified representation conversion method based on multi-engine adaptive fusion. Global optimal allocation and information extraction are performed on the directed acyclic graph model respectively to obtain the global optimal allocation result of the table header of the personnel document to be analyzed and the hierarchical structured personnel information dataset; By utilizing an incremental data collection strategy driven by change events and a retry mechanism with exponential backoff, the external information set of the company under review is obtained from publicly available external data sources. Construct a quaternary heterogeneous graph of the personnel documents to be analyzed, and process the quaternary heterogeneous graph through a method for calculating the implicit correlation degree based on heterogeneous graph neural network and spatiotemporal decay to obtain the comprehensive correlation strength value of each correlation edge in the quaternary heterogeneous graph. Based on the hierarchical structured personnel information dataset, the external information set, and multiple comprehensive correlation strength values, a structured review report and an interactive correlation graph of the enterprise to be reviewed are generated.

2. The method according to claim 1, characterized in that, Based on the original binary document stream and the preset engine format adaptation matrix, and processed by a multi-engine adaptive fusion-based unified representation transformation method for multi-format documents, a directed acyclic graph model of the personnel document to be analyzed is constructed, including: Using the original binary document stream and the multiple pre-built standard fingerprint vectors, the document format of the personnel document to be analyzed is adaptively identified, and multiple initial parsing engines that match the format of the personnel document to be analyzed are determined. Multiple initial parsing engines are used to parse the personnel document to be analyzed, and multiple engine parsing results are obtained. Each engine parsing result includes a text content matrix and a layout feature vector. Using the theory of evidence reasoning, the parsing results of the multiple engines are fused and processed to generate a document object model tree structure and a JSON text matrix set of the personnel document to be analyzed; A depth-first recursive traversal is performed on the document object model tree structure and the JSON text matrix set to construct the directed acyclic graph model.

3. The method according to claim 2, characterized in that, Using the original binary document stream and the preset engine format adaptation matrix, the document format of the personnel document to be analyzed is adaptively identified, and multiple initial parsing engines matching the format of the personnel document to be analyzed are determined, including: Extract multiple byte windows of a preset length at multiple preset offset positions in the original binary document stream, and calculate multiple frequency distribution histogram vectors corresponding to the multiple byte windows; The multiple frequency distribution histogram vectors and the multiple byte information entropy vectors within the multiple byte windows are concatenated to construct the composite fingerprint vector of the personnel document to be analyzed, wherein the byte information entropy vector is used to reflect the uniformity of byte distribution within the byte window; Calculate multiple raw weighted cosine similarities between the composite fingerprint vector and multiple pre-constructed standard fingerprint vectors; Based on the multiple original weighted cosine similarities, the format and target weighted cosine similarity of the personnel document to be analyzed are determined, wherein the target weighted cosine similarity is the largest among the multiple original weighted cosine similarities; When the target weighted cosine similarity is greater than or equal to a preset first threshold, the format of the personnel document to be analyzed is determined to be an identified format, and the engine adaptation score of the format of the personnel document to be analyzed is determined based on the preset engine format adaptation matrix. The engine compatibility score is compared with a preset second threshold, and based on the comparison result, the plurality of initial parsing engines that match the format of the personnel document to be analyzed are determined.

4. The method according to claim 2, characterized in that, Using evidence-based reasoning theory, the parsing results from the multiple engines are fused and processed to generate a document object model tree structure and a JSON text matrix set for the personnel documents to be analyzed, including: Based on multiple preset evaluation dimensions, a local quality evidence vector for each target parsing engine is generated, and a basic probability allocation function for each initial parsing engine is constructed. Based on each basic probability assignment function, calculate the basic confidence probability value and basic unconfidence probability value of each engine's parsing result, and determine the basic confidence probability distribution of each engine's parsing result; Using the Dempster combination rule, the basic confidence probability distribution of the parsing results of each engine is recursively combined pairwise, and the global fusion confidence value is determined. When the global fusion confidence value is less than the preset fusion threshold, the highest basic confidence probability value is obtained, and the document object model tree structure and the JSON text matrix set are generated based on the engine parsing result corresponding to the highest basic confidence probability value.

5. The method according to claim 4, characterized in that, The method further includes: When the global fusion confidence value is greater than or equal to the preset fusion threshold, the probability difference between the highest basic probability value and the second highest basic probability value is calculated based on the basic confidence probability value of the parsing result of each engine. When the probability difference is greater than or equal to a preset third threshold, the document object model tree structure and the JSON text matrix set are generated based on the engine parsing result corresponding to the highest basic confidence probability value. When the probability difference is less than the preset third threshold, the document object model tree structure and the JSON text matrix set are generated by using the basic credibility probability value of each engine parsing result as the weight to perform fine-grained completion on the engine parsing result corresponding to the highest basic credibility probability value.

6. The method according to claim 1, characterized in that, Global optimal allocation and information extraction are performed on the directed acyclic graph model to obtain the global optimal allocation result of the table headers of the personnel documents to be analyzed and the hierarchical structured personnel information dataset, including: Based on the aforementioned directed acyclic graph model, and after processing by a header recognition and error correction method based on bidirectional attention flow and global optimal allocation, the global optimal allocation result of the header of the personnel document to be analyzed is obtained. Construct a multidimensional position encoding vector for each node in the directed acyclic graph model, and calculate the topological context vector for each node in the directed acyclic graph model; Based on the multidimensional location encoding vector and the topological context vector of each node, an initial input representation vector for each node is generated; Based on the edge type awareness mechanism, the initial input representation vector of each node is processed by a graph attention network to obtain the target hidden layer representation vector of each node that integrates neighborhood information. Based on the global optimal allocation result of the table header and the preset nested topological constraints, the target hidden layer representation vector of each node is used to perform multi-task decoding on the personnel document to be analyzed, and generate the hierarchical structured personnel information dataset.

7. The method according to claim 6, characterized in that, Based on the aforementioned directed acyclic graph model, and after processing using a header recognition and error correction method based on bidirectional attention flow and global optimal allocation, the global optimal allocation result of the header of the personnel document to be analyzed is obtained, including: The directed acyclic graph model is extracted to obtain multiple header texts for multiple header cells in the personnel document to be analyzed; The multiple header texts and multiple preset standard categories are respectively mapped by a pre-trained semantic encoder to obtain multiple first semantic vectors and multiple second semantic vectors; Based on the plurality of first semantic vectors and the plurality of second semantic vectors, the first attention weight value of each header text for each preset standard category is calculated using forward attention flow; Calculate the matching score between each column of data in each header text and each preset standard category of data type, and determine the aggregated content feature vector of each column of data; Based on multiple aggregated content feature vectors and the multiple second semantic vectors, the second attention weight value of each column of data to each header text is calculated using backward attention flow; Multiple first attention weights and multiple second attention weights are weighted and fused, and an optimization function is constructed, wherein the optimization function represents the minimum total cost of allocation; Within a preset time complexity, the optimization function is solved using the Hungarian algorithm, and the globally optimal header allocation result of the personnel document to be analyzed is determined.

8. The method according to claim 1, characterized in that, The quaternary heterogeneous graph is processed using a method for calculating implicit association strength based on heterogeneous graph neural networks and spatiotemporal decay to obtain the comprehensive association strength value of each association edge in the quaternary heterogeneous graph, including: Based on a preset time decay function and a preset spatial decay function, the quaternary heterogeneous graph is processed by heterogeneous graph convolution to obtain the target embedding representation vector of each node in the quaternary heterogeneous graph that fuses multi-hop heterogeneous neighborhood information. Link prediction is performed using the bilinear mapping of the target embedding representation vector of each node to obtain the predicted probability value of each edge in the quaternary heterogeneous graph. Based on the predicted probability value of each edge, determine the multiple association edges of the quaternary heterogeneous graph; Calculate the score of each association edge on multiple preset association dimensions and determine the comprehensive association strength value of each association edge.

9. A device for analyzing the correlation of personnel within an enterprise, characterized in that, The device includes: The first acquisition module is used to acquire the original binary document stream of the personnel documents to be analyzed from the enterprise to be reviewed and the preset engine format adaptation matrix. In the preset engine format adaptation matrix, multiple elements represent the historical average parsing quality scores of different parsing engines for documents of different formats. The first processing module is used to construct a directed acyclic graph model of the personnel document to be analyzed based on the original binary document stream and the preset engine format adaptation matrix, through a multi-format document unified representation conversion method based on multi-engine adaptive fusion. The second processing module is used to perform global optimal allocation and information extraction on the directed acyclic graph model respectively, to obtain the global optimal allocation result of the table header of the personnel document to be analyzed and the hierarchical structured personnel information dataset. The second acquisition module is used to acquire the external information set of the enterprise under review from external public data sources by using an incremental acquisition strategy based on change event-driven and a retry mechanism with exponential backoff. The third processing module is used to construct a quaternary heterogeneous graph of the personnel documents to be analyzed, and process the quaternary heterogeneous graph through a method for calculating the implicit correlation degree based on heterogeneous graph neural network and spatiotemporal decay to obtain the comprehensive correlation strength value of each correlation edge in the quaternary heterogeneous graph. The generation module is used to generate a structured review report and an interactive relational graph of the enterprise to be reviewed based on the hierarchical structured personnel information dataset, the external information set, and multiple comprehensive correlation strength values.

10. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute the enterprise personnel correlation analysis method according to any one of claims 1 to 8.

Images