File review method, electronic device, and program product

Abstract

The present disclosure provides a file auditing method, an electronic device and a program product. The method of the present disclosure comprises: in response to an auditing request for a target file, extracting auditing information and a key field from the target file, taking the auditing information as an auditing node, and taking the key field as a field node; determining a first directed edge based on a constraint relationship between a plurality of auditing nodes; determining a second directed edge based on a dependency relationship and a constraint relationship between a plurality of field nodes; determining a third directed edge based on a judgment relationship between the auditing nodes and the field nodes; constructing a graph structure using the auditing nodes, the field nodes, the first directed edge, the second directed edge and the third directed edge; inputting the graph structure into a neural network model for auditing to generate an auditing result corresponding to each of the field nodes and the auditing nodes; and generating an auditing conclusion based on a plurality of auditing results.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more particularly to document review methods, electronic devices, and program products. Background Technology

[0002] With the development of computer technology, computers can be used for document review.

[0003] In existing technologies, document review can be automated based on preset rules. However, rule-based automation suffers from limitations such as scattered rules, high maintenance costs, and difficulty in expressing complex dependencies. Furthermore, review rules are relatively fixed, presenting significant limitations. Moreover, increasing business complexity and data volume make the review process more complex, potentially requiring the rewriting of new review rules, which imposes a significant workload on staff. Additionally, the effectiveness of such fixed review rules is unsatisfactory. Therefore, a solution is needed to improve the efficiency and accuracy of document review. Summary of the Invention

[0004] This disclosure provides document review methods, electronic devices, and program products.

[0005] According to a first aspect of this disclosure, a document review method is provided. The method specifically includes: in response to a review request for a target document, extracting review information and key fields from the target document, and using the review information as review nodes and the key fields as field nodes; determining a first directed edge based on the constraint relationships between multiple review nodes; determining a second directed edge based on the dependency and constraint relationships between multiple field nodes; determining a third directed edge based on the decision relationships between review nodes and field nodes; constructing a graph structure using the review nodes, field nodes, the first directed edge, the second directed edge, and the third directed edge; inputting the graph structure into a neural network model for review, generating review results corresponding to the field nodes and review nodes respectively; and generating a review conclusion based on the multiple review results.

[0006] Determining a first directed edge based on the constraint relationship between multiple review nodes according to at least one embodiment of this disclosure includes: determining the review order of each review node in the review process; and determining a first directed edge representing the constraint relationship between each review node according to the review order.

[0007] Determining a second directed edge based on dependencies and constraints between multiple field nodes according to at least one embodiment of this disclosure includes: determining the matching frequency between each field node; determining a second directed edge representing the dependencies between field nodes based on the matching frequency; and / or determining the business rules for each field node; and determining a second directed edge representing the constraints between field nodes based on the business rules.

[0008] Determining a third directed edge based on the determination relationship between an audit node and a field node according to at least one embodiment of this disclosure includes: determining a determination record between audit information and field information based on a determination log in historical audit records; and determining a third directed edge representing the determination relationship between the audit node and the field node based on the determination record.

[0009] According to at least one embodiment of this disclosure, in response to a request to review a target file, review information and key fields are extracted from the target file, and the review information is used as a review node, and the key fields are used as field nodes, including: extracting review information and key fields from the target file; performing feature encoding on the extracted review information and key fields, including converting text into vectors, standardizing numerical values, and performing feature encoding on categories; and creating review nodes for the review information and creating field nodes for the key fields based on the encoding results.

[0010] According to at least one embodiment of this disclosure, a graph structure is input into a neural network model, and the review nodes, field nodes, and directed edges contained in the graph structure are reviewed by the neural network model to generate review results corresponding to each node, including: inputting the graph structure into a pre-trained neural network model for aggregation processing; determining a first review result for the field nodes; wherein the first review result includes: a valid verification result and a risk assessment result; determining a second review result for the review nodes; wherein the second review result includes: a judgment conclusion.

[0011] According to at least one embodiment of this disclosure, an audit conclusion is generated based on multiple audit results, including: weighting and summing the valid verification results and risk assessment results according to preset weights to generate an overall field evaluation score; weighting and fusing the overall field evaluation score with the judgment conclusion according to business priority to generate a final comprehensive audit score; and determining the audit conclusion based on the final comprehensive audit score; wherein the audit conclusion includes: pass, reject, and manual review.

[0012] According to at least one embodiment of this disclosure, the neural network model training method includes: collecting historical audit form data, including audit node information, key fields, and historical audit conclusions; constructing a graph structure sample from the historical audit form data, including extracting field sample nodes, audit sample nodes, and audit conclusion sample nodes, as well as directed edges representing the logical relationships between nodes; and inputting the graph structure sample into the model to be trained for training to obtain a trained neural network model.

[0013] According to a second aspect of this disclosure, an electronic device is provided, comprising: a memory storing execution instructions; and a processor executing the execution instructions stored in the memory, such that the processor performs a first aspect of any embodiment of this disclosure.

[0014] According to a third aspect of this disclosure, a readable storage medium is provided, wherein executable instructions are stored therein, which, when executed by a processor, are used to implement a first aspect of any embodiment of this disclosure.

[0015] According to a fourth aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements a first aspect of any embodiment of this disclosure. Attached Figure Description

[0016] The accompanying drawings illustrate exemplary embodiments of the present disclosure and, together with the description thereof, serve to explain the principles of the present disclosure. These drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification.

[0017] Figure 1 This is a flowchart illustrating the document review method provided in this embodiment of the disclosure.

[0018] Figure 2 This is a flowchart illustrating the graph structure construction method provided in an embodiment of the present disclosure.

[0019] Figure 3 This is a flowchart illustrating the first directed edge generation method provided in an embodiment of the present disclosure.

[0020] Figure 4 This is a flowchart illustrating the second directed edge generation method provided in an embodiment of this disclosure.

[0021] Figure 5 This is a flowchart illustrating the third directed edge generation method provided in an embodiment of this disclosure.

[0022] Figure 6 This is a flowchart illustrating the key field extraction method provided in the embodiments of this disclosure.

[0023] Figure 7 This is a flowchart illustrating the method for generating audit results provided in this embodiment of the disclosure.

[0024] Figure 8 This is a flowchart illustrating the method for generating audit conclusions provided in this embodiment of the disclosure.

[0025] Figure 9 This is a schematic flowchart of the model training method provided in the embodiments of this disclosure.

[0026] Figure 10 This is a schematic block diagram of a document review device according to one embodiment of the present disclosure.

[0027] Figure 11 This is a schematic block diagram of an electronic device according to one embodiment of the present disclosure. Detailed Implementation

[0028] The present disclosure will now be described in further detail with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are for illustrative purposes only and are not intended to limit the scope of the disclosure. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the accompanying drawings.

[0029] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a flowchart illustrating the document review method provided in an embodiment of this disclosure. Figure 1 The method shown includes steps S101 to S105. This method can be executed by the server.

[0031] Specifically, Figure 1 The method shown includes step S101: in response to an audit request for a target file, extracting audit information and key fields from the target file, and using the audit information as an audit node and the key fields as field nodes.

[0032] It should be noted that the "approval node" mentioned here refers to a data unit used to represent information about the reviewer or review process, storing relevant attributes of the review subject. The "field node" refers to a data unit used to represent key fields in the approval form (such as identity information, transaction amount, and document number), carrying the specific content of the business elements.

[0033] In practical applications, when a user submits a target file (such as a real estate transaction review form), the review information (such as the name of the initial reviewer, the ID of the secondary reviewer, and the review opinions at each stage) and key fields (such as the buyer's ID number, property certificate number, transaction amount, and fund flow records) can be identified and extracted. Natural Language Processing (NLP), regular expression matching, template recognition, or Optical Character Recognition (OCR) can be used to extract the review information and key fields from the target file, ensuring that key information from different types of files can be accurately extracted and classified into the corresponding node types.

[0034] Step S102: Determine directed edges based on the logical relationships between audit nodes, between field nodes, and between audit nodes and field nodes.

[0035] A directed edge is a connecting line that represents the logical dependency between nodes. It has a clear directionality and is used to express the influence or constraint of one node on another node.

[0036] By analyzing correlation patterns in historical audit data, the system automatically identifies which fields have business dependencies (such as the matching relationship between "bank statement amount" and "income certificate"), which audit stages have constraints (such as "preliminary review passed" being a prerequisite for "secondary review"), and which audit operations are directly based on specific key fields. These logical relationships are quantified as the weights and directions of directed edges, representing the actual logical relationships between different nodes. In determining the directed edges, the system fully reflects the logical relationships between multiple audit nodes, multiple field nodes, and between audit nodes and field nodes, expressing multi-level, non-linear audit logical dependencies and providing detailed evidence for subsequent audits of target documents.

[0037] For example: a directed edge A→B is established between field node A (contract amount) and audit node B (amount threshold check), indicating that "the amount field must undergo threshold audit"; a directed edge C→D is established between audit node C (date format check) and audit node D (date validity check), indicating that "validity can only be checked after the format is correct"; a directed edge E→F is established between field node E (start date) and field node F (end date), indicating the dependency constraint that "the start date must be earlier than the end date". These directed edges can be determined through a pre-configured rule base, business templates, or semantic analysis, ensuring that all logical dependencies are explicitly represented through directed edges.

[0038] Step S103: Construct a graph structure using audit nodes, field nodes, and directed edges.

[0039] It should be noted that the graph structure mentioned here refers to a directed graph data model consisting of review nodes, field nodes, and the directed edges between them, used to express the topological relationships of the review logic.

[0040] After determining the directed edges between each node, the field nodes and review nodes can be further integrated into a single directed graph. The presence of field nodes allows for fine-grained review down to individual data items, while review nodes represent key nodes in the review process; the two form a topological structure through directed edges. This graph structure effectively overcomes the limitations of treating the target document as a flat table or linear text, fully preserving and expressing the complex interactions between various element fields and review nodes during the review process. Field nodes in the graph structure are no longer isolated data points but form a unified network with other field nodes and review nodes through directed edges, enabling the system to understand complex business logic. For example, when field A is abnormal, a comprehensive judgment needs to be made by combining field B and field C.

[0041] Step S104: Input the graph structure into the neural network model for review, and generate the review results corresponding to the field nodes and review nodes respectively.

[0042] During the review process, each node is reviewed separately, taking into full account the logical relationships between different nodes (i.e., the logical relationships represented by directed edges). In one alternative approach, a graph analysis algorithm can be used to traverse the graph structure and evaluate the status of each node from multiple dimensions: for field nodes, the system evaluates their data integrity, format compliance, business rationality, and the logical relationships (constraints, dependencies) between upstream and downstream nodes; for review nodes, based on the status of associated field nodes and historical review patterns, it predicts the judgment that should be made in that review stage. The system does not evaluate each node in isolation, but rather transmits information through directed edges in the graph structure, enabling mutual influence and collaborative judgment between nodes. This graph-based collaborative evaluation method allows for full attention to each key field in the target file and the logical relationships between them, while also considering the role and impact of review nodes throughout the entire review process, rather than simply executing preset rules.

[0043] Step S105: Generate an audit conclusion based on multiple audit results.

[0044] The review conclusion mentioned here refers to the final review decision derived from the combined results of reviews at multiple nodes.

[0045] In practical applications, the review results of field nodes (reflecting the review status of specific business data) and review nodes (reflecting the judgment status of the review process) are comprehensively analyzed. The overall review status of the target document is determined through pre-defined decision logic. This approach considers not only the independent evaluation results of each node but also the logical chains formed between nodes through directed edges, thereby generating review conclusions that conform to business rules and possess context-aware capabilities. The implementation scheme for weighted aggregation of multiple review results will be described in subsequent embodiments and will not be repeated here.

[0046] Based on the publicly available solutions described above, the introduction of field nodes, which, together with review nodes, construct a directed graph structure, allows the review process to move beyond simply checking the entire document or isolated fields. Instead, it enables the precise capture of multi-level logical dependencies between field-level data and rules, significantly improving the accuracy and comprehensiveness of the review. The field node configuration allows the system to independently verify and track each key data item, avoiding missed detections or misjudgments caused by implicit dependencies between fields. Simultaneously, the topological characteristics of the directed graph support the natural expression of complex business rules (such as conditional branches, sequential dependencies, and cross-validation), effectively reducing the cost of manual review and minimizing human error. This ultimately improves the accuracy and efficiency of the review results.

[0047] In one or more embodiments of this disclosure, such as Figure 2 This is a flowchart illustrating the graph structure construction method provided in an embodiment of this disclosure. Figure 2 As shown, steps S102 and S103, which determine directed edges based on the logical relationships between review nodes, field nodes, and review nodes and field nodes, and construct a graph structure using review nodes, field nodes, and directed edges, include: Step S201: Determining a first directed edge based on the constraint relationships between multiple review nodes. Step S202: Determining a second directed edge based on the dependency and constraint relationships between multiple field nodes. Step S203: Determining a third directed edge based on the decision relationship between review nodes and field nodes. Step S204: Constructing a graph structure using the first directed edge, the second directed edge, the third directed edge, and the review nodes and field nodes.

[0048] In practical applications, by analyzing historical review process data, the temporal constraints and conditional dependencies between review nodes are automatically identified. In one optional approach, the constraint relationships between review nodes can be identified based on a pre-configured rule dependency library, business process diagram, or semantic analysis results, and a directed edge can be established for each pair of review nodes with constraints. The direction of the edge reflects the execution flow of the constraint or the direction of the conditional dependency. This determination of the first directed edge ensures that the graph structure can fully express the topological order and logical combination of review rules, avoiding errors caused by ignoring field nodes during independent item-by-item checks or by the mutual influence between review nodes.

[0049] By mining the relationships between fields in historical audit data, the system identifies how the state of one field affects the evaluation result of another for dependencies. For constraints, the system identifies the field conditions that must be met simultaneously. By establishing second directed edges between different field nodes, the system further enables previously independent field nodes to express dependencies and constraints through these second directed edges, building upon the existing semantic content of each node. This significantly improves the ability to identify hidden risks.

[0050] By analyzing operation logs in historical review records, the system accurately captures the relationships between review nodes and field nodes. For example, when the "ID photo clarity check" review operation frequently makes judgments based on the "ID card scan" field, the system establishes a third directed edge from the "ID photo clarity check" review node to the "ID card scan" field node, assigning corresponding weight values ​​based on historical judgment frequency and consistency. This judgment relationship includes not only direct associations (such as review operations directly targeting a specific field) but also indirect associations (such as review operations based on a comprehensive evaluation of multiple fields). The establishment of the third directed edge enables the system to reconstruct the decision path in real review scenarios, providing a basis for generating interpretable review conclusions subsequently.

[0051] It should be noted that the order of the execution steps in the process of establishing the first, second, and third directed edges mentioned above can be adjusted as needed, and the process of establishing all three directed edges can also be executed simultaneously.

[0052] After obtaining the three types of directed edges through the above steps, the data dependencies and constraints between fields, the constraints of the review rules, and the judgment relationship between fields and the review are all explicitly expressed in a topological structure by combining the three types of directed edges. This transforms the originally scattered and implicit business logic into a computable and traversable graph relationship. After the graph is constructed, it can be directly applied to subsequent review processes. For example, the execution order can be determined by topological sorting, the chain effect can be calculated by path propagation, and the root cause of the problem can be located by subgraph analysis, thereby achieving efficient and accurate automatic review.

[0053] Based on the publicly available solutions, by mapping constraints, dependencies, and decision relationships to first, second, and third directed edges respectively, and constructing a unified graph structure, the review process can comprehensively capture the multi-level logical relationships between field nodes, between review nodes, and between fields and review nodes. The direct connection between field nodes and review nodes (third directed edges) ensures the independent and accurate verification of each key data item, while the explicit modeling of dependencies between fields (second directed edges) and constraints between review nodes (first directed edges) makes the originally implicit dependencies in the target file quantifiable, significantly improving the accuracy, completeness, and efficiency of the review, and reducing compliance risks caused by logical omissions.

[0054] In one or more embodiments of this disclosure, such as Figure 3 This is a flowchart illustrating the first directed edge generation method provided in an embodiment of this disclosure. Figure 3As shown, determining the first directed edge based on the constraint relationship between multiple review nodes includes: Step S301: Determining the review order of each review node in the review process. Step S302: Determining the first directed edge representing the constraint relationship between each review node according to the review order.

[0055] It should be noted that the review process mentioned here refers to the logical execution order or dependency sequence that each review rule must follow when conducting a complete data review of the target file. This process is usually predefined by business specifications, compliance requirements, or risk control strategies. The review order refers to the sequence in which review nodes are executed throughout the entire review process, such as performing basic format checks first, then numerical threshold validation, and finally cross-consistency verification.

[0056] In practical applications, there are several ways to obtain the review order. For example, the review order can be determined by parsing the pre-set review process template of the business system and extracting the standard execution sequence of each review node. Alternatively, the review order can be determined by analyzing the operation logs in historical review records and statistically analyzing the actual execution path of the review nodes.

[0057] For each review node, the system analyzes its upstream and downstream nodes, establishing a first directed edge from the first execution node to the next execution node. Each first directed edge not only represents the execution order but also carries rich constraint information, such as conditional expressions ("When the data integrity check pass rate > 90%, some review steps can be skipped"), execution probability (path selection probability derived from historical data analysis), and constraint strength (the degree of mandatory constraint on the critical path). This refined method of constructing first directed edges enables the system to accurately express the complex logical dependencies in the review process, rather than simply the order of execution.

[0058] For example, in a multi-level review scenario, the "initial review" node must be completed before the "secondary review" node. Based on this, the system can establish a first directed edge from the "initial review" node to the "secondary review" node, indicating that "passing the initial review" is a prerequisite for executing the "secondary review." This constraint relationship includes not only simple sequential relationships but also complex conditional constraints. For instance, when the "risk rating" review node determines a risk level to be "high-risk," it must go through the "risk control expert" review node instead of a regular review node. The establishment of this first directed edge enables the system to accurately model the internal logic of the review process, ensuring that the review process complies with business specifications and compliance requirements.

[0059] Based on the publicly available solutions, by determining the review order and establishing the first directed edge accordingly, the constraint relationship between review nodes is accurately expressed, ensuring that the review process strictly follows the sequential logic defined by the business, avoiding computational redundancy, logical confusion, and potential risks caused by disordered execution; the existence of the first directed edge significantly improves the accuracy and efficiency of the review.

[0060] In one or more embodiments of this disclosure, such as Figure 4 This is a flowchart illustrating the second directed edge generation method provided in an embodiment of this disclosure. Figure 4 As shown, determining the second directed edge based on the dependencies and constraints between multiple field nodes includes: Step S401: Determining the matching frequency between each field node. Step S402: Determining the second directed edge representing the dependencies between field nodes based on the matching frequency. Alternatively, Step S403: Determining the business rules for each field node. Step S404: Determining the second directed edge representing the constraints between field nodes based on the business rules.

[0061] There are two types of dependencies between field nodes: one is the dependency relationship between field nodes, and the other is the constraint relationship between field nodes. The following will explain the implementation process of constructing the second directed edge for each of these two logical relationships.

[0062] When establishing the second directed edge based on dependencies, the system analyzes massive amounts of historical review data to calculate the co-occurrence frequency of each pair of field nodes during the review process. Specifically, a field association matrix can be constructed and recorded in historical review forms. This matrix tracks the frequency at which field B simultaneously exhibits a corresponding state when field A displays a specific state (e.g., abnormal, high risk). For example, in a financial loan review scenario, statistics show that when the "bank statement amount" field exhibits abnormal fluctuations, the "social security contribution base" field also shows an abnormality with a probability as high as 85%. This high matching frequency indicates a strong dependency between the two fields. Further statistical methods (such as Pearson correlation coefficient and mutual information) are used to quantify the matching frequency, eliminating the interference of random associations and ensuring that the identified dependencies are statistically significant. This data-driven approach enables the system to automatically discover implicit associations that business personnel might overlook.

[0063] Furthermore, a matching frequency threshold can be pre-set. When the matching frequency of two field nodes exceeds this threshold, a second directed edge is established from the "cause field" to the "result field," and the edge weight is assigned according to the strength of the matching frequency. For example, in payment transaction review, the system finds that the matching frequency between the "transaction time" field and the "user active period" field is 0.92 (far higher than the threshold of 0.7). Therefore, a second directed edge is established from "user active period" to "transaction time," and a weight value of 0.92 is assigned, indicating that when the transaction time deviates from the user active period, the transaction risk increases significantly. This dependency modeling not only considers the direct association between pairs of fields but also identifies indirect dependency chains through graph algorithms, such as the indirect transitive dependency relationship formed by "field A to field B and then to field C." The establishment of the second directed edge enables the system to understand the dynamic association patterns between fields. When an anomaly occurs in a field in the target file, the system can automatically assess its impact on related fields, achieving a more accurate risk assessment.

[0064] When establishing second directed edges based on constraint relationships, the business rules for field nodes can be analyzed and determined from a rule base, or business rules extracted from high-quality audit records using a model. For each business rule, the system parses the field nodes involved and the logical relationships between them, establishing the corresponding second directed edges. This business rule-based approach ensures that the audit process complies with industry standards and regulatory requirements, avoiding compliance risks that may arise from a purely data-driven approach.

[0065] The dual construction mechanism of the second directed edge (dependency based on matching frequency and constraint based on business rules) has significant technical advantages and is necessary. In this disclosed solution, the establishment of a second directed edge ensures that the relationships between fields are no longer simplified to isolated fields. Through the second directed edge, a complete model of the field network is achieved, enabling the system to understand complex logic. At the same time, the constraint relationship driven by business rules ensures that core compliance requirements are not interfered with by data noise.

[0066] The two methods described above can be used independently or in combination (e.g., first using matching frequency to discover candidate dependencies, then using business rules to verify and refine them). Both ultimately generate a second directed edge, transforming the dependencies and constraints between field nodes from an implicit or scattered state into explicit topological connections in a graph structure. This avoids ignoring the relationships between fields and directly validating individual fields independently, which could lead to missed cross-constraint checks or cascading errors.

[0067] Based on the publicly available solutions, by determining the second directed edge based on matching frequency or business rules, accurate and flexible modeling of dependencies and constraints between field nodes is achieved, enabling the graph structure to comprehensively capture implicit statistical associations and explicit business constraints in the data layer. The matching frequency approach enhances the system's adaptability and discovery capabilities, making it suitable for dynamic or rule-incomplete scenarios. The business rule approach ensures the strictness and controllability of constraints, avoiding misjudgments. The existence of the second directed edge significantly improves the completeness and accuracy of the audit, effectively reducing the risk of missed detections and improving audit efficiency.

[0068] In one or more embodiments of this disclosure, such as Figure 5 This is a flowchart illustrating the third directed edge generation method provided in an embodiment of this disclosure. Figure 5 As shown, determining the third directed edge based on the judgment relationship between the audit node and the field node includes: Step S501: Based on the judgment log in the historical audit records, determine the judgment record between the audit information and the field information. Step S502: Based on the judgment record, determine the third directed edge representing the judgment relationship between the audit node and the field node.

[0069] In practical applications, the system analyzes a structured historical audit database to extract detailed correlation information between each audit operation and the audited fields. Specifically, the system performs data analysis from different dimensions, such as analyzing the type and content of audit operations (e.g., "ID photo clarity check," "income verification"). It also analyzes the field nodes actually referenced during the audit process (e.g., "ID card scan," "bank statement"). Furthermore, it analyzes the correlation strength between audit decisions and referenced fields (e.g., "when the ID card scan's blurriness is >0.6, 90% of auditors will request supplementary materials"). For example, in historical data on financial loan audits, the system found that the "risk assessment" audit node references the "credit score" field in 85% of cases and the "debt ratio" field in 70% of cases; these statistics are recorded as judgment records. The system further employs association rule mining algorithms to identify the conditional probability relationship between audit decisions and field states, thereby quantifying the judgment correlation between audit information and field information. This extraction of judgment records based on historical data enables the system to objectively reconstruct the decision logic in real audit scenarios, rather than relying on subjective experience or simplified assumptions.

[0070] After extracting the judgment records, a weighted directed graph model can be used to construct a third directed edge. In one optional approach, the specific implementation is as follows: Calculate the association weight for each judgment record, taking into account three factors: reference frequency, decision impact, and historical accuracy. When the association weight exceeds a preset threshold, establish a third directed edge between the corresponding review node and the field node; label each third directed edge with judgment rules, including conditional expressions and decision impact patterns. This refined construction of the third directed edge not only captures explicit associations (such as directly referenced fields) but also identifies implicit associations (such as fields indirectly influencing decisions) through graph neural networks, enabling the system to fully express the complex network of criteria for review decisions.

[0071] By using dynamic modeling driven by historical judgment records, the actual judgment relationship between review information and field information is explicitly transformed into a third directed edge in the graph structure. This allows the graph structure to accurately reflect historical review behavior patterns and support precise matching of review rules and fields. Especially in complex scenarios with a large rule base and diverse fields, it avoids invalid judgments or key omissions.

[0072] Based on the publicly available solutions, by determining the judgment records based on historical review records and judgment logs, and establishing a third directed edge accordingly, the dynamic determination and adjustment of the judgment relationship between review nodes and field nodes is realized, enabling the graph structure to more accurately reflect the actual business review relationship; the existence of the third directed edge significantly improves the accuracy of the review.

[0073] In one or more embodiments of this disclosure, such as Figure 6 This is a flowchart illustrating the key field extraction method provided in the embodiments of this disclosure. Figure 6 As shown, step S101, in response to a review request for the target file, extracts review information and key fields from the target file, and uses the review information as a review node and the key fields as field nodes, including: Step S601: Extracting review information and key fields from the target file. Step S602: Performing feature encoding on the extracted review information and key fields, including converting text to vectors, standardizing numerical values, and encoding categories. Step S603: Based on the encoding results, creating review information as a review node and key fields as field nodes.

[0074] In practical applications, parsing the data structure of target files includes identifying key-value pairs in structured formats. A multimodal data parsing model can be employed to handle target files in various formats such as JSON, XML, PDF, and images. For structured data (such as JSON returned by an API), field paths and hierarchical relationships can be identified through pattern matching. For semi-structured data (such as PDF documents), key fields and review information can be extracted using layout analysis and semantic recognition techniques. For unstructured data (such as scanned images), content parsing can be performed by combining OCR and deep learning models. This multimodal parsing capability ensures that the system can accurately extract review information and key fields from target files of various complex formats, providing a high-quality data foundation for subsequent node creation.

[0075] Next, the system extracts review information from the target file, including reviewer information and review conclusion information (it should be noted that the new target file does not yet have review conclusions, so blank spaces can be reserved for storing predicted review conclusions later). Specifically, a predefined review information pattern library can be used to match review-related data in the target file. For reviewer information, the system extracts data such as reviewer role, reviewer ID, review timestamp, and historical review performance; for review conclusion information, the system extracts data such as review conclusion type (pass / reject / review), reason for conclusion, and confidence level. In an optional solution, context-aware technology can also be used to distinguish between historical review records and current pending review information, ensuring that only review information related to the current review request is extracted. This precise extraction of review information enables the system to accurately reconstruct the context of the review process, providing a complete information foundation for constructing review nodes.

[0076] Simultaneously, the system extracts key fields from the target file, including identity information, transaction elements, and risk tags. In one optional approach, a domain-specific field recognition model can be used to accurately locate core business data. For identity information, personal identifiers such as ID number, name, and contact information can be extracted; for transaction elements, core business parameters such as transaction amount, time, and location can be extracted; and for risk tags, risk indicators automatically generated by the system or manually labeled can be extracted. Furthermore, entity recognition and relationship extraction technologies can be used to understand implicit relationships between fields, such as identifying that "bank statements" and "income certificates" belong to the same verification topic. Through comprehensive and accurate extraction of key fields, it ensures that the subsequently constructed field nodes fully reflect the core field information required for business review.

[0077] Furthermore, the extracted audit information and key fields can be feature-encoded. In one alternative approach, differentiated encoding strategies can be used to process different types of data: for text data (such as audit reasons, document types), a pre-trained language model (such as BERT) can be applied to convert it into a fixed-dimensional vector representation, preserving semantic information; for numerical data (such as transaction amount, age), standardization can be performed to map the original values ​​to the [0,1] interval, eliminating the influence of units; for categorical data (such as document types, transaction channels), discrete values ​​can be converted into continuous vectors through embedding encoding or one-hot encoding.

[0078] After encoding is completed and the encoding results are obtained, the audit information and key fields are created as audit nodes and field nodes, respectively. Specifically, an audit node can be created for each instance of audit information, with node attributes including the encoded feature vector and metadata (such as audit stage type and auditor role). A field node is also created for each instance of key field, with node attributes including the encoded feature vector and business semantic tags (such as field type and validation rules). For example, the system creates an audit node for the encoded result of "Initial Auditor: Zhang San" and a field node for the encoded result of "ID Number: 1101011990 XXXXXXXX". The system also adds rich contextual information to the nodes, such as associating field nodes with business rule tags and audit nodes with historical accuracy metrics. This node creation process not only completes the data representation transformation but also preserves the key semantics of the original information through feature encoding, making the nodes effective units for subsequent graph structure analysis.

[0079] By using feature encoding, the audit information and key fields are transformed into a unified, computable vector representation, which upgrades the extraction process from static parsing to dynamic, semantically rich node creation, supporting the efficient construction of subsequent graph structures and complex logical reasoning.

[0080] Based on the publicly available solutions mentioned above, by parsing the data structure and accurately extracting audit information (auditor information, audit conclusion information) and key fields (identity information, transaction elements, risk tags), combined with feature encoding (text vectors, numerical standardization, category encoding), a high-quality mapping from the original file to graph nodes is achieved. The introduction of feature encoding significantly improves the comparability of the data and computational efficiency, ensuring that audit nodes and field nodes have semantic consistency and numerical compatibility in the graph structure, and avoiding logical errors or loss of accuracy caused by data heterogeneity.

[0081] In one or more embodiments of this disclosure, such as Figure 7 This is a flowchart illustrating the method for generating audit results provided in this embodiment of the disclosure. Figure 7As shown, step S104, which involves inputting the graph structure into a neural network model and using the neural network model to review the review nodes, field nodes, and directed edges contained in the graph structure to generate review results for each node, includes: Step S701: Inputting the graph structure into a pre-trained neural network model for aggregation processing. Step S702: Determining the first review result for the field nodes; wherein the first review result includes: valid verification result and risk assessment result. Step S703: Determining the second review result for the review nodes; wherein the second review result includes: judgment conclusion.

[0082] In practical applications, this pre-trained neural network model can be a graph neural network model, which contains a multi-layer graph convolutional network architecture. Each layer includes two core operations: message passing and node updating. The node features in the graph structure (including the encoded representations of review nodes and field nodes) serve as model input, and the edge connections define the information transmission paths. This graph neural network model can directly process non-Euclidean data, preserving the topological relationships between nodes, which is crucial for capturing the complex dependencies in the review logic. This enables the model to understand the interrelationships between these elements, rather than processing each field in isolation.

[0083] Next, node information aggregation is performed through a multi-layer propagation mechanism of the graph neural network model. Each layer of propagation includes two key steps: message passing and node updating. In the message passing phase, each node collects feature information from its neighboring nodes, forming a message set. In the node updating phase, learnable aggregation functions (such as weighted summation, max pooling, etc.) can be applied to combine the collected messages with the node's own features to generate an updated node representation. This process iterates through the multi-layer network, enabling each node to capture increasingly broader contextual information.

[0084] Furthermore, the aggregated node representations are input into the classification layer to calculate the probability distribution of the node's review status. For field nodes, the system outputs valid verification results (e.g., correct format, complete information) and their confidence scores; for review nodes, the system outputs a judgment conclusion (e.g., passed, requires manual review) and its confidence score. The confidence score reflects the model's certainty about the review result and is calculated based on the matching degree between the node representation and historical review patterns. The calculation of local review status not only considers the characteristics of the node itself but also integrates the contextual information passed in the graph structure, making the evaluation more comprehensive and accurate.

[0085] Based on the aggregated node representation, the system predicts the review results of field nodes, including validation validity or risk level. A dual-task learning framework can be adopted: one branch predicts valid validation results, assessing whether the field meets basic format and completeness requirements; the other branch predicts risk assessment results, assessing whether the field poses business risks. For valid validation results, the system checks whether the field data conforms to predefined rules (such as ID number format, whether required fields exist); for risk assessment results, the system analyzes the degree of anomaly of the field in the business context (such as whether the transaction amount deviates from the user's historical patterns).

[0086] By combining the representations of review nodes with those of associated field nodes, the system simulates the decision-making logic in a real review process. The judgment of a review node depends not only on its directly associated fields but also on the status of other nodes along the entire review path. In one alternative approach, an attention mechanism can be used to automatically learn the contribution weights of different fields to the review decision, making the judgment more consistent with business logic. When the representation of a review node shows high uncertainty, a result of "Process Status: Requires Manual Review" can be generated to avoid making incorrect decisions under low confidence conditions. This graph-based review node prediction enables the system to recreate the complex decision-making processes in real review scenarios, rather than simply executing preset rules.

[0087] Based on the publicly available solutions described above, the node information aggregation mechanism of the graph neural network model can effectively capture the complex dependencies between field nodes. The dual-dimensional evaluation of field nodes (valid verification results and risk assessment results) enables the system to distinguish between technical compliance and business compliance. Furthermore, the judgment conclusion generation process of the review node incorporates the contextual information of related fields, enabling the system to make a comprehensive judgment rather than simply executing preset rules.

[0088] In one or more embodiments of this disclosure, such as Figure 8 This is a flowchart illustrating the method for generating audit conclusions provided in this embodiment of the disclosure. Figure 8 As shown, step S105, which generates an audit conclusion based on multiple audit results, includes: Step S801: Weighting and summing the valid verification results and risk assessment results according to preset weights to generate an overall field evaluation score. Step S802: Weighting and fusing the overall field evaluation score with the judgment conclusion according to business priority to generate a final comprehensive audit score. Step S803: Determining the audit conclusion based on the final comprehensive audit score; wherein, the audit conclusion includes: pass, reject, and manual review.

[0089] After generating the review results for each node, the valid verification results and risk assessment results are weighted and summed according to preset weights to generate an overall field evaluation score. Specifically, weight configurations can be defined for each field node to reflect the relative importance of valid verification and risk assessment in the review of that field. For example, for the "ID number" field, the valid verification weight may be higher (0.7) because format correctness is crucial; while for the "transaction amount" field, the risk assessment weight may be higher (0.6) because business risk is more critical. The weighted sum is calculated as follows: Overall field evaluation score = w1 × valid verification score + w2 × risk assessment score, where w1 and w2 are preset weights, satisfying w1 + w2 = 1. This weighted fusion method allows the system to flexibly adjust the evaluation focus according to the business characteristics of different fields, generating a comprehensive score that better meets actual business needs.

[0090] Furthermore, the overall field evaluation score and the judgment conclusion are weighted and fused to generate the final comprehensive review score. In one optional approach, a hierarchical fusion method can be adopted. First, the overall field evaluation scores within the same review stage are aggregated to generate a stage-level evaluation score; then, the stage-level evaluation score is fused with the judgment conclusion of the corresponding review node. The fusion weight is dynamically adjusted according to business rules, with key fields and core review stages receiving higher weights. For example, in financial loan reviews, the "income verification" field and the "risk assessment" stage may receive higher weights. The final comprehensive review score is calculated as: α × field aggregation score + β × review node judgment score, where α and β are weighting coefficients determined by business priority. This hierarchical weighted fusion mechanism ensures that key business elements are given sufficient consideration in the final decision-making process, while maintaining the transparency and explainability of the review logic.

[0091] Finally, the system determines the review conclusion based on the final comprehensive review score. Technically, the system sets multiple threshold ranges to map the final comprehensive review score to specific review conclusions: when the score is higher than the high threshold, a pass conclusion is generated; when the score is lower than the low threshold, a rejection conclusion is generated; and when the score falls between the two, a manual review conclusion is generated. It should be noted that the threshold settings take into account business risk preferences and historical review data, and can be dynamically adjusted to adapt to different business scenarios.

[0092] Based on the publicly available solutions described above, by inputting the graph structure into a pre-trained graph neural network model for multi-layer aggregation processing, joint representation and prediction of field nodes and review nodes in the topological context are achieved. This allows the review results to fully utilize the dependencies between nodes and global structural information, avoiding the low accuracy caused by simply reviewing based on the semantics of the target file. Furthermore, the separate generation of the first review result (valid verification + risk assessment) and the second review result (judgment conclusion), along with the subsequent weighted fusion mechanism, provides fine-grained, highly interpretable multi-level output. The weighted fusion of the overall field evaluation score and the business priority of the judgment conclusion further enhances the comprehensiveness and practicality of the review conclusion. This enables a balance between risk sensitivity and pass rate in complex business scenarios, significantly reducing the risk of misjudgment and the need for manual intervention.

[0093] In one or more embodiments of this disclosure, such as Figure 9 This is a schematic flowchart illustrating the model training method provided in an embodiment of this disclosure. Figure 9 As shown, the neural network model training method includes: Step S901: Collect historical review form data, including review node information, key fields, and historical review conclusions. Step S902: Construct a graph structure sample from the historical review form data, including extracting field sample nodes, review sample nodes, and review conclusion sample nodes, as well as directed edges representing the logical relationships between nodes. Step S903: Input the graph structure sample into the model to be trained for training, and obtain the trained neural network model.

[0094] The model mentioned above is a graph neural network model obtained after training with specific training samples. The specific training process is as follows: Completed audit records can be extracted from the historical database of the business system, ensuring data coverage of diverse business scenarios and audit results. The data to be collected mainly includes: audit node information (such as auditor ID, audit stage, audit time, etc.), key fields (such as identity information, transaction elements, risk tags, etc.), and historical audit conclusions (such as approved, rejected, manually reviewed, etc.). Furthermore, the collected data undergoes quality assessment and cleaning to remove incomplete or abnormal records, ensuring the representativeness and reliability of the training data. This large-scale, high-quality historical data collection provides a solid foundation for the model to learn realistic audit logic, avoiding performance degradation due to insufficient or biased data.

[0095] After data extraction, a graph structure sample can be constructed to train the model. Specifically, field sample nodes can be extracted from historical audit forms, corresponding to key fields in the business, such as ID number, transaction amount, and risk score. Each field sample node stores the field value and its metadata. Next, audit sample nodes are extracted, corresponding to each stage and auditor in the audit process, such as initial review, secondary review, and risk control experts. Each audit sample node stores the audit operation and opinions. Finally, audit conclusion sample nodes are extracted, corresponding to the final audit decision, such as pass or rejection. After extraction, directed edges representing the logical relationships between nodes are constructed, including directed edges representing dependencies between field sample nodes, directed edges representing constraints between audit sample nodes, and directed edges representing judgment relationships between audit sample nodes and field sample nodes. The graph structure sample fully expresses the logical structure of the audit case. This construction of the graph structure sample transforms historical audit forms from flat records into structured knowledge, providing a suitable representation for the model to learn the audit logic.

[0096] Historical review conclusions are used as global labels for graph structure samples, while local supervision signals are generated for key nodes. For review conclusion sample nodes, historical review conclusions are directly used as supervision labels; for key field sample nodes and review sample nodes, auxiliary supervision signals are generated based on their roles and influences in historical reviews. Furthermore, node features are encoded, converting the raw data into numerical representations suitable for model processing: text features are converted into vectors through a pre-trained language model, numerical features are standardized, and categorical features are embedded and encoded. The use of supervision signals and feature encoding enables historical review experience to effectively guide model learning.

[0097] During training, the graph structure samples prepared in the previous embodiments are input into the graph neural network model to be trained in batches. In an optional approach, a dynamic batching strategy can be used to group graph structure samples with similar structures to improve training efficiency. For each batch of graph structure samples, they are converted into adjacency matrices and node feature matrices, which are then used as input to the graph neural network model. Considering that the size and complexity of graph structure samples may vary significantly, graph filling and pruning techniques can be used to ensure that samples within a batch have compatible dimensions. For example, small graph structure samples containing 5-20 nodes can be grouped together, while medium-sized graph structures containing 20-50 nodes can be grouped separately. This helps improve the stability and efficiency of the training process.

[0098] During training, the graph neural network model progressively aggregates neighbor information of nodes through multi-layer graph convolution operations. At each layer, each node collects feature information from its direct neighbors and updates its representation using learnable aggregation functions (such as weighted summation and max pooling). As the number of layers increases, nodes can capture increasingly broader contextual information, forming richer representations that include multi-hop dependencies. This multi-layer information aggregation enables the model to make comprehensive judgments and learn complex implicit dependencies between fields, rather than just surface linear rules.

[0099] During training, a multi-task loss function can be employed to simultaneously optimize the prediction of global review conclusions and key nodes. For review conclusion sample nodes, cross-entropy loss can be calculated to measure the difference between the model's predicted review conclusions and historical review conclusions. For key field sample nodes and review sample nodes, auxiliary losses are calculated to guide the model in learning node-level review logic. Furthermore, a regularization term can be introduced to prevent the model from overfitting to noise in historical data. This multi-task loss design enables the model to learn both global review strategies and local review rules simultaneously, improving the overall learning performance.

[0100] During training, gradient descent can be employed to calculate the gradient of the loss function with respect to the model parameters using the chain rule, and then the parameters are updated to minimize the loss. Considering that the deep structure of graph neural networks may lead to the vanishing gradient problem, the system employs gradient pruning and adaptive learning rate strategies (such as the Adam optimizer) to ensure stable and efficient training. As an alternative, an early stopping strategy can be used, stopping training when validation set performance no longer improves to prevent overfitting. This meticulous training control ensures that the model learns validation rules with strong generalization ability, rather than simply memorizing historical data.

[0101] It's important to note that the model training process is an iterative process that continues until the model converges or reaches the preset number of training rounds, resulting in a trained graph neural network model. For example, a progressive training strategy can be employed: initially using a larger learning rate for rapid convergence, and later using a smaller learning rate for fine-tuning; simultaneously, the system periodically evaluates the model's performance on the validation set, selecting the best-performing checkpoint as the final model. This complete training process ensures that the model can automatically extract implicit review patterns from historical review forms, achieving the core objective of automatically learning review rules and applying them to new review forms (i.e., target documents).

[0102] Based on the publicly available solutions described above, by constructing historical review data into graph structure samples, structural information (logical relationships such as fields and review nodes) from the review process can be preserved and utilized. Secondly, the multi-layered information aggregation mechanism of the graph neural network model enables the model to automatically learn implicit dependencies between fields, eliminating the need for manual maintenance of numerous rules and significantly reducing operational costs. Furthermore, the multi-task loss function design allows the model to simultaneously learn global review strategies and local review rules, improving the accuracy and adaptability of the review process.

[0103] Based on any of the above embodiments, this disclosure also provides a document review device. Figure 10 This is a schematic block diagram of a document review device according to one embodiment of the present disclosure. Figure 10 As shown, the document review device includes: an extraction module 1001, used to extract review information and key fields from the target file in response to a review request for the target file, and to use the review information as review nodes and the key fields as field nodes. A determination module 1002 is used to determine a first directed edge based on the constraint relationships between multiple review nodes; determine a second directed edge based on the dependency and constraint relationships between multiple field nodes; and determine a third directed edge based on the decision relationships between review nodes and field nodes. A graph construction module 1003 is used to construct a graph structure using the review nodes, field nodes, and directed edges. A generation module 1004 is used to input the graph structure into a neural network model, and to review the review nodes, field nodes, and directed edges contained in the graph structure through the neural network model, generating review results corresponding to the field nodes and review nodes respectively. Finally, it generates a review conclusion based on multiple review results.

[0104] The determination module 1002 is used to determine the review order of each review node in the review process; based on the review order, the first directed edge representing the constraint relationship between each review node is determined.

[0105] The determination module 1002 is used to determine the matching frequency between each field node; determine the second directed edge representing the dependency relationship between the field nodes based on the matching frequency; and / or determine the business rules of each field node; and determine the second directed edge representing the constraint relationship between the field nodes based on the business rules.

[0106] The determination module 1002 is used to determine the determination record between the audit information and the field information based on the determination log in the historical audit record; and to determine the third directed edge representing the determination relationship between the audit node and the field node based on the determination record.

[0107] The extraction module 1001 is used to extract audit information and key fields from the target file; to perform feature encoding on the extracted audit information and key fields, including converting text into vectors, standardizing numerical values, and encoding categories; and to create audit nodes for the audit information and field nodes for the key fields based on the encoding results.

[0108] The generation module 1004 is used to input the graph structure into a pre-trained neural network model for aggregation processing; determine the first review result of the field node; wherein the first review result includes: valid verification result and risk assessment result; determine the second review result of the review node; wherein the second review result includes: judgment conclusion.

[0109] The generation module 1004 is used to perform weighted summation of valid verification results and risk assessment results according to preset weights to generate an overall field evaluation score; according to business priority, the overall field evaluation score is weighted and integrated with the judgment conclusion to generate a final comprehensive review score; the review conclusion is determined based on the final comprehensive review score; the review conclusion includes: pass, reject, and manual review.

[0110] Training module 1005 is used to collect historical audit form data, including audit node information, key fields and historical audit conclusions; construct graph structure samples from the historical audit form data, including field sample nodes, audit sample nodes and audit conclusion sample nodes, as well as directed edges representing the logical relationships between nodes; input the graph structure samples into the model to be trained for training, and obtain the trained neural network model.

[0111] The specific implementation process of the functions and roles of each module in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0112] The executing entity of the document review method in the specific embodiments of this disclosure can be an electronic device such as a server (including a local server or a cloud computing platform).

[0113] Therefore, based on any of the above embodiments, this disclosure also provides an electronic device that can execute the document review method of any of the embodiments described above.

[0114] Figure 11 This is a schematic block diagram of an electronic device according to one embodiment of the present disclosure.

[0115] The hardware architecture of the electronic device 1000 can be implemented using a bus architecture. The bus architecture can include any number of interconnect buses and bridges, depending on the specific application of the hardware and overall design constraints. Bus 1100 connects various circuits, including one or more processors 1200, memory 1300, and / or hardware modules. Bus 1100 can also connect various other circuits 1400, such as peripheral devices, voltage regulators, power management circuits, external antennas, etc.

[0116] Bus 1100 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, this diagram uses only one connection line, but this does not imply that there is only one bus or one type of bus.

[0117] This disclosure also provides a readable storage medium storing a computer program that, when executed by a processor, is used to implement the methods described above. A "readable storage medium" can be any means capable of containing, storing, communicating, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples of a readable storage medium include: an electrical connection with one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM or flash memory), fiber optic devices, and portable read-only memory (CDROM), etc.

[0118] This disclosure also provides a computer program product, the methods of which can be implemented wholly or partially through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented wholly or partially as a computer program product. The computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, all or part of the processes or functions of this disclosure are performed.

[0119] Computer programs or instructions can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any available medium capable of access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0120] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0122] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0123] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0124] In the description of this specification, the references to terms such as "one embodiment / mode," "some embodiments / modes," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, or characteristic described in connection with that embodiment / mode or example is included in at least one embodiment / mode or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment / mode or example. Moreover, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments / modes or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments / modes or examples described in this specification, as well as the features of different embodiments / modes or examples.

[0125] Furthermore, 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 at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0126] Those skilled in the art should understand that the above embodiments are merely for illustrating the present disclosure and are not intended to limit the scope of the disclosure. Those skilled in the art can make other changes or modifications based on the above disclosure, and these changes or modifications still fall within the scope of the present disclosure.

Claims

1. A document review method, characterized in that, The method includes: In response to a request to review a target file, review information and key fields are extracted from the target file, and the review information is used as a review node, and the key fields are used as field nodes. The first directed edge is determined based on the constraint relationships between the multiple audit nodes; The second directed edge is determined based on the dependencies and constraints between the multiple field nodes; The third directed edge is determined based on the judgment relationship between the audit node and the field node; A graph structure is constructed using the audit node, the field node, the first directed edge, the second directed edge, and the third directed edge; The graph structure is input into a neural network model, and the neural network model reviews the review nodes, field nodes, and directed edges contained in the graph structure, generating review results corresponding to the field nodes and review nodes respectively; and, An audit conclusion is generated based on multiple audit results.

2. The document review method according to claim 1, characterized in that, The determination of the first directed edge based on the constraint relationships between the multiple audit nodes includes: Determine the review order of each review node in the review process; Based on the review order, the first directed edge representing the constraint relationship between each of the review nodes is determined.

3. The document review method according to claim 1, characterized in that, The determination of the second directed edge based on the dependency and constraint relationships between multiple field nodes includes: Determine the matching frequency between each of the aforementioned field nodes; The second directed edge representing the dependency relationship between the field nodes is determined based on the matching frequency; and / or, Determine the business rules for each of the aforementioned field nodes; The second directed edge representing the constraint relationship between the field nodes is determined based on the business rules.

4. The document review method according to claim 1, characterized in that, The determination of the third directed edge based on the judgment relationship between the audit node and the field node includes: Based on the judgment log in the historical review records, determine the judgment records between the review information and the field information; Based on the judgment record, the third directed edge representing the judgment relationship between the audit node and the field node is determined.

5. The document review method according to claim 1, characterized in that, The step of responding to an audit request for a target file by extracting audit information and key fields from the target file, and using the audit information as an audit node and the key fields as field nodes, includes: Extract the audit information and the key fields from the target file; The extracted audit information and key fields are feature-encoded, including converting text into vectors, standardizing numerical values, and encoding categories. Based on the encoding results, the audit information is created as the audit node, and the key fields are created as the field nodes.

6. The document review method according to claim 1, characterized in that, The step of inputting the graph structure into a neural network model and using the neural network model to review the review nodes, field nodes, and directed edges contained in the graph structure, generating review results corresponding to each node, includes: The graph structure is input into a pre-trained neural network model for aggregation processing; Determine the first review result of the field node; wherein, the first review result includes: valid verification result and risk assessment result; Determine the second audit result of the audit node; wherein the second audit result includes: a judgment conclusion.

7. The document review method according to claim 6, characterized in that, The process of generating an audit conclusion based on multiple audit results includes: The valid verification results and the risk assessment results are weighted and summed according to preset weights to generate an overall field assessment score. Based on business priorities, the overall evaluation score of the fields is weighted and fused with the judgment conclusion to generate the final comprehensive review score; The review conclusion is determined based on the final comprehensive review score; wherein the review conclusion includes: pass, rejection, and manual review.

8. The document review method according to claim 1, characterized in that, Neural network model training methods include: Collect historical audit data, including audit node information, key fields, and historical audit conclusions; A graph structure sample is constructed for the historical audit order data, including field extraction sample nodes, audit sample nodes, and audit conclusion sample nodes, as well as directed edges representing the logical relationships between nodes; The graph structure samples are input into the model to be trained to obtain the trained neural network model.

9. An electronic device, characterized in that, include: The memory stores execution instructions; as well as A processor that executes execution instructions stored in the memory, causing the processor to perform the method of any one of claims 1 to 8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images