AI model assisted material auditing method, system, device and medium

Abstract

The application relates to the technical field of data processing, and in particular to an AI model auxiliary-based material auditing method, system, device and medium. The application first extracts text from the material to be audited through an OCR module, and then performs semantic analysis on the unstructured text by using an NLP module to convert the unstructured text into structured features; the structured features are processed based on a preset knowledge base to generate an auditing feature vector containing business semantics; then the feature vector is input into a machine learning model composed of an anomaly detection model and a risk classification model to generate a preliminary auditing conclusion from three dimensions of basic rules, business rules and risk rules; finally, the conclusions of the three dimensions are evaluated for confidence, and corresponding processing procedures are triggered according to the evaluation results; through the cooperation of multiple AI models, the transformation from surface text recognition to deep semantic understanding is realized, a comprehensive evaluation mechanism based on multiple dimensions is established, and the accuracy and efficiency of the auditing are improved.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to AI model-assisted document review methods, systems, devices and media. Background Technology

[0002] With the deepening of enterprise digital transformation, the volume of documents and data has exploded. Enterprises need to process a large number of business documents such as invoices, purchase orders, and expense reports every day. These documents not only involve monetary calculations but also require comprehensive assessments of compliance, authenticity, and risk. How to efficiently and accurately complete document review has become a significant challenge for enterprise operations.

[0003] Existing automated review systems, such as the patent with publication number CN119091460A, provide a text recognition and automated review method. This method targets semi-structured and unstructured document data, selects a corresponding OCR engine for text recognition, and uses rule extraction and model extraction to obtain structured text data. The method determines compliance by comparing the data with preset target structured text data and dynamically adjusts the confidence threshold based on document complexity, ultimately outputting a compliance rate.

[0004] However, existing technologies only focus on compliance judgments from a single dimension and lack a comprehensive evaluation mechanism for multi-dimensional audit results, making it unable to effectively handle complex business scenarios. Summary of the Invention

[0005] To address the shortcomings of existing review systems that focus solely on single-dimensional compliance assessments and lack a comprehensive evaluation mechanism for multi-dimensional review results, thus failing to effectively handle complex business scenarios, this application provides an AI model-assisted document review method, system, equipment, and media, employing the following technical solution:

[0006] Firstly, this application provides an AI model-assisted document review method, including the following steps:

[0007] Receive pending review materials obtained from the enterprise information system;

[0008] The OCR module is used to extract the text of the documents to be reviewed.

[0009] The NLP module is used to perform semantic parsing on the data text to generate structured features;

[0010] The structured features are processed based on a pre-defined knowledge base to generate an audit feature vector;

[0011] The audit feature vector is input into a preset machine learning model to generate a preliminary audit conclusion. The machine learning model includes an anomaly detection model and a risk classification model. The preliminary audit conclusion includes basic rule check results, business rule matching results, and risk rule assessment results.

[0012] A confidence level assessment is performed based on the results of the basic rule check, the business rule matching, and the risk rule evaluation, and the corresponding processing flow is triggered based on the assessment results.

[0013] By adopting the above technical solution, this application first extracts text from the materials to be reviewed using an OCR module, then performs semantic parsing using an NLP module to convert unstructured text into structured features. Based on a preset knowledge base, the structured features are processed to generate review feature vectors containing business semantics. Then, the feature vectors are input into a machine learning model composed of an anomaly detection model and a risk classification model to generate preliminary review conclusions from three dimensions: basic rules, business rules, and risk rules. Finally, the confidence levels of the conclusions in these three dimensions are evaluated, and corresponding processing procedures are triggered based on the evaluation results. Through the collaborative cooperation of multiple AI models, the transformation from surface text recognition to deep semantic understanding is realized, and a comprehensive evaluation mechanism based on multiple dimensions is established, improving the accuracy and efficiency of the review.

[0014] Optionally, the preset knowledge base integrates optical character recognition data and natural language processing data through an entity association interface to construct an entity association network of suppliers, customers, contracts, and product specifications.

[0015] By adopting the above technical solutions, the knowledge base design based on entity association networks in this application not only improves the integrity and consistency of information, but also supports more complex business scenario analysis, providing knowledge support for subsequent feature processing and risk assessment.

[0016] Optionally, the anomaly detection model employs a combination of the Isolation Forest algorithm and the Long Short-Term Memory network. The Isolation Forest algorithm is used to quickly identify obvious abnormal data, while the Long Short-Term Memory network is used to learn the time-series features of the document data.

[0017] By adopting the above technical solution, this application organically combines the Isolation Forest algorithm with the Long Short-Term Memory network. The Isolation Forest algorithm detects based on the isolation characteristics of the data, while the LSTM network is responsible for capturing long-term dependencies in time series data. This not only ensures the ability to respond quickly to obvious anomalies, but also enables the mining of potential abnormal patterns in time series data through deep learning networks.

[0018] Optionally, the method further includes the following steps:

[0019] Priority thresholds are set for the basic rule check results, business rule matching results, and risk rule assessment results, respectively;

[0020] The system checks whether there are rule conflicts among the basic rule check results, business rule matching results, and risk rule assessment results.

[0021] When a rule conflict is detected, the priority threshold corresponding to the conflicting rule is obtained;

[0022] The higher priority rule is executed according to the relationship between the priority thresholds, and the rule processing result is generated.

[0023] By adopting the above technical solution, this application sets priority thresholds for the basic rule check results, business rule matching results, and risk rule assessment results, respectively. These thresholds can be configured according to the business importance and risk level. When performing rule checks, the system automatically detects whether there are conflicts in the judgment results of different rules. When a rule conflict is found, the priority thresholds corresponding to the relevant rules are obtained, and the execution order is determined by comparing the threshold sizes. This achieves the orderliness and consistency of rule judgment and avoids the review chaos caused by rule conflicts.

[0024] Optionally, the risk classification model divides the audit results into three levels: high risk, medium risk, and low risk.

[0025] When the review result is high risk, an automatic rejection process is triggered and a risk report is generated.

[0026] When the review result is medium risk, a manual review process is triggered and key points are highlighted.

[0027] If the audit result is low risk, it will be archived and stored directly.

[0028] By adopting the above technical solution, the audit results are divided into three levels through a risk classification model, and differentiated processing procedures are designed for different risk levels; this not only ensures a rapid response to high-risk situations, but also improves the overall audit efficiency.

[0029] Optionally, the confidence assessment specifically includes the following steps:

[0030] Assign corresponding weight values ​​to the basic rule check results, business rule matching results, and risk rule assessment results respectively;

[0031] The weight values ​​are dynamically adjusted according to the rule type, wherein the rule type includes mandatory field check rules, budget matching rules, and supplier qualification verification rules;

[0032] The confidence assessment result is obtained by weighting the basic rule check result, business rule matching result and risk rule assessment result with the adjusted weight values.

[0033] By adopting the above technical solution, this application first sets initial weight values ​​for the basic rule check results, business rule matching results, and risk rule assessment results. The system then dynamically adjusts the weights according to the current rule type being processed. Rule types include mandatory field check rules, budget matching rules, and supplier qualification verification rules. For example, when processing mandatory field checks, the system increases the weight of the basic rule check results; when performing budget matching, it increases the weight ratio of the business rule matching results. Finally, the system uses the adjusted weight values ​​to perform a weighted calculation on each rule result to obtain the final confidence assessment result. This achieves flexible adjustment of the assessment criteria, enabling the system to optimize assessment strategies according to different business scenarios.

[0034] Optionally, the processing flow specifically includes the following steps:

[0035] Obtain the output results of the risk classification model and the determination results of the confidence assessment;

[0036] Based on the output and the judgment results, you can choose to activate the automatic review channel or the manual review channel.

[0037] By adopting the above technical solution, the system first obtains the risk rating results of the risk classification model for the documents, and at the same time obtains the judgment results of the confidence assessment mechanism for comprehensive analysis, and selects the most suitable processing channel.

[0038] Secondly, this application provides an AI model-assisted document review system, including:

[0039] The data receiving module is used to receive data to be reviewed obtained from the enterprise information system;

[0040] The OCR processing module is used to extract the text from the documents to be reviewed.

[0041] The NLP processing module is used to perform semantic parsing on the data text and generate structured features;

[0042] The knowledge base processing module is used to process the structured features based on a preset knowledge base to generate audit feature vectors;

[0043] The machine learning module is used to input the audit feature vector into a preset machine learning model to generate a preliminary audit conclusion. The machine learning model includes an anomaly detection model and a risk classification model. The preliminary audit conclusion includes basic rule check results, business rule matching results, and risk rule assessment results.

[0044] The assessment and processing module is used to assess confidence based on the basic rule check results, business rule matching results, and risk rule assessment results, and to trigger the corresponding processing flow based on the assessment results.

[0045] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-mentioned AI model-assisted data review method.

[0046] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-mentioned AI model-assisted document review method.

[0047] In summary, this application includes at least one of the following beneficial technical effects:

[0048] This application first extracts text from the materials to be reviewed using an OCR module, then performs semantic parsing using an NLP module to convert unstructured text into structured features. Based on a pre-defined knowledge base, the structured features are processed to generate review feature vectors containing business semantics. These feature vectors are then input into a machine learning model composed of an anomaly detection model and a risk classification model to generate preliminary review conclusions from three dimensions: basic rules, business rules, and risk rules. Finally, the confidence levels of these three conclusions are assessed, and corresponding processing flows are triggered based on the assessment results. Through the collaborative work of multiple AI models, a transformation from surface text recognition to deep semantic understanding is achieved, establishing a multi-dimensional comprehensive evaluation mechanism that improves the accuracy and efficiency of the review process.

[0049] This application organically combines the Isolation Forest algorithm with the Long Short-Term Memory network. The Isolation Forest algorithm detects based on the isolation characteristics of data, while the LSTM network is responsible for capturing long-term dependencies in time-series data. This ensures both a rapid response to obvious anomalies and the ability to mine potential abnormal patterns in time-series data through deep learning networks.

[0050] This application sets priority thresholds for the basic rule check results, business rule matching results, and risk rule assessment results. These thresholds can be configured according to the business importance and risk level. When performing rule checks, the system automatically detects whether there are conflicts in the judgment results of different rules. When a rule conflict is found, the priority thresholds corresponding to the relevant rules are obtained, and the execution order is determined by comparing the threshold sizes. This achieves the orderliness and consistency of rule judgment and avoids the review chaos caused by rule conflicts. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the document review method based on an AI model, as described in this application.

[0052] Figure 2 This is a flowchart illustrating the generation of rule processing results in the AI ​​model-assisted document review method of this application embodiment;

[0053] Figure 3 This is a flowchart illustrating the confidence assessment process in the AI ​​model-assisted document review method of this application embodiment;

[0054] Figure 4 This is a schematic diagram of the modules of the document review system based on AI model assistance in an embodiment of this application;

[0055] Figure 5 This is an internal structural diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0056] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0057] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0058] The embodiments of this application will now be described in further detail with reference to the accompanying drawings.

[0059] Firstly, this application provides a document review method based on an AI model, referring to... Figure 1 It includes the following steps:

[0060] S110. Receive the documents to be reviewed from the enterprise information system.

[0061] In this embodiment, the enterprise information system refers to the internal information system used for daily operation and management within an enterprise, including a business processing system and a document management system. The documents to be reviewed mainly consist of various business documents generated during the enterprise's daily operations.

[0062] Specifically, the system retrieves pending documents from the enterprise information system via an interface. These documents are stored in electronic document format, including purchase contracts, payment applications, and expense reports. The system places the retrieved documents into a pending review queue and assigns them a unique task identifier. For example, in an invoice review scenario, the enterprise information system will retrieve VAT invoices issued by suppliers, corresponding purchase orders, contracts, and other related documents. In a qualification review scenario, the system will receive qualification documents provided by suppliers, such as business licenses, tax registration certificates, and production licenses. In an insurance claims scenario, the system will retrieve claims-related documents such as claims applications, accident reports, and loss lists.

[0063] S120. Extract the document text from the documents to be reviewed using the OCR module.

[0064] In this embodiment, the OCR module is used to convert images or PDF documents into processable text content. It adopts a deep learning text recognition algorithm and supports the processing of multiple document formats.

[0065] Specifically, the OCR module first preprocesses the input document, including image enhancement, layout analysis, and text localization. Then, using a trained recognition model, it converts the text areas in the document into machine-readable text format. For table areas, the system extracts cell content using a table structure recognition algorithm. For example, when processing VAT invoices, the OCR module can accurately recognize fields such as invoice code, invoice number, invoice date, buyer and seller information, product name, specifications, quantity, unit price, and amount. When processing business licenses, it can accurately extract key information such as company name, unified social credit code, legal representative, registered address, and business scope. When processing bank statements, it can recognize fields such as transaction date, transaction amount, counterparty, and account balance.

[0066] S130. Use the NLP module to perform semantic parsing on the data text and generate structured features.

[0067] In this embodiment, the NLP module refers to the functional module used to understand and process natural language text. Structured features refer to the normalized information extracted from unstructured text, including field attributes, business elements, and semantic relationships.

[0068] Specifically, the NLP module uses a pre-trained language model to process the text. First, it performs word segmentation and part-of-speech tagging to identify entity names, numerical information, and time information. Then, it extracts semantic relationships from the text, such as rights and obligations in contracts and payment terms in transactions. Finally, the extracted information is converted into structured feature vectors for subsequent rule matching and risk assessment. For example, in invoice review, the NLP module can parse information such as contract numbers and project names contained in invoice remarks; in business license review, it can perform semantic understanding of the business scope text to determine whether it covers relevant business content; and in insurance claims review, it can extract key information such as the cause of the accident and the extent of the loss from the accident description.

[0069] Furthermore, in some possible embodiments, when the OCR module's confidence level for recognizing a specific region is below 0.8, relevant contextual information is passed to the NLP module for semantic assistance in judgment. For example, when processing invoice amounts, if the OCR confuses the numbers "8" and "3", the NLP module will combine contextual information such as contract amount and historical transaction records to provide the most likely value, thereby improving recognition accuracy.

[0070] S140. Based on the preset knowledge base, the structured features are processed to generate audit feature vectors.

[0071] The pre-defined knowledge base integrates optical character recognition data and natural language processing data through an entity association interface to construct an entity association network of suppliers, customers, contracts, and product specifications. The audit feature vector refers to the standardized set of features processed by the knowledge base.

[0072] Specifically, the knowledge base integrates OCR recognition data and NLP processing results through an entity association interface. The entity association network comprises four core entities: supplier profiles, customer information, contract templates, and product specifications. The system establishes an association network based on the business relationships between entities, supporting cross-document information retrieval and association analysis. During the feature processing stage, the system matches structured features with entity information in the knowledge base, supplementing relevant business attributes and association information, ultimately generating an audit feature vector containing complete business semantics. For example, in invoice and contract association analysis, the knowledge base can quickly retrieve information such as transaction amount and product specifications for the corresponding contract, achieving cross-validation. In customer access auditing, the entity association network can quickly obtain multi-dimensional information such as the supplier's historical transaction records, credit rating, and cooperative projects.

[0073] In some possible implementations, the knowledge base integrates OCR recognition data and NLP processing results through an entity association interface. Employing dynamic knowledge graph technology, the knowledge base automatically extracts new entity relationships and updates the knowledge network after each review task is completed. Simultaneously, the system incorporates temporal feature analysis to record the evolution of entity relationships over time. For example, by tracking changes in suppliers' historical transaction patterns, the system can automatically discover potential business rules and use them to optimize review strategies.

[0074] S150. Input the audit feature vector into the preset machine learning model to generate a preliminary audit conclusion.

[0075] The machine learning models include anomaly detection models and risk classification models, and the preliminary review conclusions include basic rule check results, business rule matching results, and risk rule assessment results.

[0076] Specifically, the anomaly detection model uses the Isolation Forest algorithm to identify obvious data anomalies, while simultaneously employing Long Short-Term Memory (LSTM) networks to analyze time-series features. The risk classification model, trained on historical labeled data, is used to assess the risk level of documents. The system inputs the audit feature vectors into both models, generating a comprehensive preliminary audit conclusion encompassing multiple dimensions. For example, in invoice audits, the anomaly detection model can identify anomalies in invoice amounts and issuance times; the risk classification model can assess transaction risk levels based on supplier historical performance and transaction frequency. In bank statement audits, it can identify abnormal fund transfer patterns, such as frequent large transfers and suspicious fund circulation. Anomaly samples detected by the anomaly detection model are automatically added to the training set of the risk classification model; the risk classification model's judgment results, in turn, adjust the anomaly detection threshold parameters. The system also introduces a dynamic confidence weight mechanism, automatically adjusting the weight ratio of each model in the final decision based on their historical accuracy in different business scenarios.

[0077] S160. Based on the results of the basic rule check, the business rule matching results, and the risk rule assessment results, a confidence level assessment is performed, and the corresponding processing flow is triggered based on the assessment results.

[0078] In this embodiment, confidence assessment refers to the quantitative evaluation of the credibility of the audit conclusion. The processing flow refers to the subsequent processing methods selected based on the assessment results.

[0079] Specifically, the system calculates weights for the results of basic rule checks, business rule matching, and risk rule assessments. By setting weight coefficients for different rule types, a comprehensive evaluation of the audit conclusions is achieved. Based on the evaluation score, the system selects to trigger either automatic auditing or manual review to ensure the accuracy and reliability of the audit results.

[0080] Furthermore, the risk classification model categorizes review results into three levels: high-risk, medium-risk, and low-risk. When the review result is high-risk, an automatic rejection process is triggered, and a risk report is generated. When the review result is medium-risk, a manual review process is triggered, and key points are highlighted. When the review result is low-risk, the result is directly archived. The processing flow specifically includes obtaining the output results of the risk classification model and the confidence assessment results; based on the output results and the assessment results, the corresponding process is selected and initiated, such as an automatic review channel, a manual review channel, or direct archiving. For example, for high-risk invoices (such as amounts significantly different from the contract), the system directly triggers a rejection process and generates a risk report; for medium-risk cases (such as minor discrepancies in amount), a manual review is triggered, and key points are highlighted; for low-risk cases (such as slight differences in remarks), the system directly archives the result. In insurance claims review, the system can select the appropriate processing flow based on characteristics such as the claim amount and accident type.

[0081] In one possible implementation, the system first categorizes audit tasks using a scenario recognition model, such as invoice audit, contract audit, and qualification audit. For different scenarios, the system automatically adjusts the weighting of confidence assessment. For example, when processing large invoices, the system increases the weighting of the amount matching rule; when auditing the qualifications of new suppliers, it increases the weighting of the risk assessment rule. The system also continuously optimizes the weighting parameters for each scenario based on historical audit results, achieving adaptive evolution of confidence assessment.

[0082] In one embodiment, refer to Figure 2 The method also includes the following steps:

[0083] S210. Set priority thresholds for the basic rule check results, business rule matching results, and risk rule assessment results respectively.

[0084] In this embodiment, the rule priority threshold is used to determine the execution order and conflict handling strategy for different types of rules. Basic rules mainly include mandatory field checks, field format validation, and document integrity checks; business rules include budget matching, approval authority verification, and business process compliance checks; risk rules include supplier qualification verification, transaction anomaly identification, and related transaction checks.

[0085] Specifically, the system sets different priority ranges for three types of rules: the priority threshold for basic rules is set in the range of 10 to 30, for business rules in the range of 31 to 60, and for risk rules in the range of 61 to 90. Within each range, the priority is further subdivided according to the importance of the rule. For example, in the basic rules, the priority of required field checks is set to 25, field format validation is set to 20, and document integrity checks are set to 15.

[0086] S220, Check whether there are rule conflicts in the results of the basic rule check, the business rule matching results, and the risk rule assessment results.

[0087] In this embodiment, rule conflict refers to contradictory judgments made by different rules regarding the same audit object. For example, in the amount format rules, the basic rule requires the amount to be an integer, while the business rule allows the amount to contain two decimal places; or the business rule determines that the transaction amount is within the budget range, while the risk rule determines that the transaction is abnormal.

[0088] Specifically, the system establishes a rule conflict detection matrix to record the constraint relationships between rules. When performing rule checks, the system automatically compares the results against the detection matrix to identify any logical conflicts. If a conflict is found, the system records the conflicting rule number and the specific content of the conflict.

[0089] S230. When a rule conflict is detected, obtain the priority threshold corresponding to the conflicting rule.

[0090] In this embodiment, the system needs to obtain the priority thresholds corresponding to each of the conflicting rules in order to determine which rule to use in the final judgment.

[0091] Specifically, when a rule conflict is detected, the system reads the priority information of the conflicting rules from the rule configuration library. For each pair of conflicting rules, the system also obtains their priority thresholds to prepare for subsequent priority comparisons.

[0092] S240. Execute the high-priority rule according to the priority threshold relationship and generate the rule processing result.

[0093] In this embodiment, the system selects the rule with the highest priority as the final decision criterion based on the priority thresholds. The rule processing result includes the rule number, execution result, and processing description.

[0094] Specifically, the system compares the priority thresholds of conflicting rules and executes the rule with the highest priority. For example, when there is a conflict between the basic rule (priority 25) and the business rule (priority 35) regarding the amount format requirements, the system adopts the judgment criteria of the business rule. After execution, the system generates a rule processing result report, recording the rule basis and specific processing procedures used.

[0095] In one embodiment, refer to Figure 3 The confidence assessment specifically includes the following steps:

[0096] S310. Set corresponding weight values ​​for the basic rule check results, business rule matching results, and risk rule assessment results, respectively.

[0097] In this embodiment, the rule weight value refers to the degree of influence of different rule judgment results on the confidence level assessment. The basic rule check result reflects the basic compliance of the document, the business rule matching result reflects the compliance of the business logic, and the risk rule assessment result indicates the degree of control over potential risks.

[0098] Specifically, the initial weight value for the basic rule check results is set to 0.2, mainly considering the completeness of basic elements; the initial weight value for the business rule matching results is set to 0.3, focusing on the compliance of business processes; and the initial weight value for the risk rule assessment results is set to 0.5, focusing on assessing the effectiveness of risk control. The system stores these weight values ​​in a weight configuration table as the basic parameters for the assessment.

[0099] S320. Dynamically adjust the weight values ​​according to the rule type, where the rule type includes mandatory field check rules, amount budget matching rules, and supplier qualification verification rules.

[0100] In this embodiment, the system dynamically adjusts the weight values ​​based on the specific rule type. Rule types include mandatory field check rules, budget matching rules, and supplier qualification verification rules; the importance of different types of rules varies across different business scenarios.

[0101] Specifically, the system establishes a mapping relationship between rule types and weight adjustment coefficients. When processing rules for checking required fields, the system increases the weight value of the basic rule; when processing rules for matching budget amounts, it increases the weight percentage of the business rule; and when executing rules for verifying supplier qualifications, it increases the weight coefficient of the risk rule. The adjusted weight values ​​still need to meet the constraint that the sum of the weights is 1.

[0102] S330. The confidence assessment result is obtained by weighting the basic rule check result, business rule matching result and risk rule assessment result with the adjusted weight values.

[0103] In this embodiment, the confidence assessment result refers to the comprehensive score obtained through weighted calculation, which is used to measure the credibility of the audit results. The system uses adjusted weight values ​​to perform weighted calculations on the three types of rule results.

[0104] Specifically, the system first standardizes the judgment results of various rules into a score format. Then, it multiplies the adjusted weight value by the corresponding rule score and sums them to obtain the final confidence score.

[0105] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0106] Secondly, this application provides an AI model-assisted document review system. The AI ​​model-assisted document review system of this application will be described below in conjunction with the above-mentioned AI model-assisted document review method.

[0107] Reference Figure 4 An AI-model-assisted document review system includes:

[0108] The data receiving module is used to receive data to be reviewed obtained from the enterprise information system;

[0109] The OCR processing module is used to extract the text from the documents to be reviewed.

[0110] The NLP processing module is used to perform semantic parsing on the data text and generate structured features;

[0111] The knowledge base processing module is used to process structured features based on a preset knowledge base and generate audit feature vectors.

[0112] The machine learning module is used to input the audit feature vector into a preset machine learning model to generate preliminary audit conclusions. The machine learning model includes an anomaly detection model and a risk classification model. The preliminary audit conclusions include basic rule check results, business rule matching results, and risk rule assessment results.

[0113] The assessment and processing module is used to assess confidence based on the results of basic rule checks, business rule matching, and risk rule assessments, and to trigger the corresponding processing flow based on the assessment results.

[0114] In one embodiment, this application provides an electronic device, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, this electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. The database stores data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements an AI model-assisted document review method.

[0115] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0116] In one embodiment, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0117] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0118] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A document review method based on AI model assistance, characterized in that, Includes the following steps: Receive pending review materials obtained from the enterprise information system; The OCR module is used to extract the text of the documents to be reviewed. The NLP module is used to perform semantic parsing on the data text to generate structured features; The structured features are processed based on a preset knowledge base to generate audit feature vectors. The preset knowledge base integrates optical character recognition data and natural language processing data through an entity association interface to construct an entity association network of suppliers, customers, contracts and product specifications. Dynamic knowledge graph technology is used so that the knowledge base automatically extracts new entity relationships and updates the network after each audit task is completed. The audit feature vector is input into a preset machine learning model to generate preliminary audit conclusions. The machine learning model includes an anomaly detection model and a risk classification model. The preliminary audit conclusions include basic rule check results, business rule matching results, and risk rule assessment results. Anomaly samples detected by the anomaly detection model are automatically added to the training set of the risk classification model. The judgment results of the risk classification model, in turn, adjust the threshold parameters of the anomaly detection model. The anomaly detection model uses a combination of the Isolation Forest algorithm and a Long Short-Term Memory (LSTM) network. The Isolation Forest algorithm is used to quickly identify obvious abnormal data, and the LTM network is used to learn the time-series features of the document data. Confidence assessment is performed based on the results of the basic rule check, the business rule matching, and the risk rule evaluation, and the corresponding processing flow is triggered based on the evaluation results. The method further includes the following steps: Priority thresholds are set for the basic rule check results, business rule matching results, and risk rule assessment results, respectively. Corresponding weight values ​​are also set for each of these three results. These weight values ​​are dynamically adjusted based on the rule type, which includes mandatory field check rules, budget matching rules, and supplier qualification verification rules. The adjusted weight values ​​are then used to perform a weighted calculation on the basic rule check results, business rule matching results, and risk rule assessment results to obtain the confidence level assessment result. The system detects whether there are rule conflicts among the basic rule check results, business rule matching results, and risk rule assessment results. Specifically, a rule conflict detection matrix is ​​established to record the constraint relationships between rules. When a rule check is performed, the system automatically compares the results with the detection matrix to identify whether there are conflicts between the judgment results. When a rule conflict is detected, the priority threshold corresponding to the conflicting rule is obtained; The higher priority rule is executed according to the relationship between the priority thresholds, and the rule processing result is generated.

2. The document review method based on AI model assistance according to claim 1, characterized in that, The risk classification model categorizes the audit results into three levels: high risk, medium risk, and low risk. When the review result is high risk, an automatic rejection process is triggered and a risk report is generated. When the review result is medium risk, a manual review process is triggered and key points are highlighted. If the audit result is low risk, it will be archived and stored directly.

3. The document review method based on AI model assistance according to claim 1, characterized in that, The processing flow specifically includes the following steps: Obtain the output results of the risk classification model and the determination results of the confidence assessment; Based on the output and the judgment results, you can choose to activate the automatic review channel or the manual review channel.

4. A document review system based on AI model assistance, characterized in that: The document review method based on AI model assistance according to any one of claims 1-3 includes: The data receiving module is used to receive data to be reviewed obtained from the enterprise information system; The OCR processing module is used to extract the text from the documents to be reviewed. The NLP processing module is used to perform semantic parsing on the data text and generate structured features; The knowledge base processing module is used to process the structured features based on a preset knowledge base to generate audit feature vectors; The machine learning module is used to input the audit feature vector into a preset machine learning model to generate a preliminary audit conclusion. The machine learning model includes an anomaly detection model and a risk classification model. The preliminary audit conclusion includes basic rule check results, business rule matching results, and risk rule assessment results. The assessment and processing module is used to assess confidence based on the basic rule check results, business rule matching results, and risk rule assessment results, and to trigger the corresponding processing flow based on the assessment results.

5. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the AI ​​model-assisted document review method according to any one of claims 1-3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the document review method based on AI model assistance as described in any one of claims 1-3.

Images