A document risk identification method, apparatus, device, medium, and product.

By using neural network models to automatically identify general and specific risks in electronic documents, and combining risk knowledge graphs and discriminant sub-models, the problem of low efficiency in manual identification is solved, achieving efficient and accurate document risk identification and potential risk identification.

CN123068627APending Publication Date: 2026-11-10INDUSTRIAL AND COMMERCIAL BANK OF CHINA
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611025400.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Current technologies for identifying risks in electronic documents mainly rely on manual methods, which are inefficient.

Method used

An automated risk identification method based on a neural network model is adopted to improve identification efficiency and accuracy by identifying general and specific risks in electronic documents and combining risk knowledge graphs and risk discrimination sub-models.

Benefits of technology

It has achieved automation and high efficiency in electronic document risk identification, improved the accuracy and comprehensiveness of risk identification, and can identify potential risks and perform document repair or alerts.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123068627A_ABST
    Figure CN123068627A_ABST
Patent Text Reader

Abstract

This application provides a document risk identification method, apparatus, device, medium, and product, which can be applied to the fields of artificial intelligence technology and fintech. The method includes: determining a target document type for a target electronic document from N preset document types; each of the N preset document types corresponds to one of N specific risk identification models; each specific risk identification model is used to identify specific risks in documents of the corresponding preset document type; N is a positive integer; based on the specific risk identification model corresponding to the target document type, identifying the target specific risk of the target electronic document according to the target document characteristics; based on a general risk identification model, identifying the target general risk of the target electronic document according to the target document characteristics; and determining the risk identification result of the target electronic document based on the identified target general risk and target specific risk.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence technology and fintech, specifically to a document risk identification method, apparatus, device, medium, and product. Background Technology

[0002] With the popularization of information technology, electronic documents are widely used as a convenient tool in daily life and work. For example, they are used to edit reports that need to be printed or to transmit work information.

[0003] However, electronic documents themselves also carry risks. For example, there may be typos, layout issues that don't conform to the preset format, or missing images. Therefore, risk identification is usually necessary for electronic documents.

[0004] Currently, risk identification of electronic documents is usually done manually, which is relatively inefficient. Summary of the Invention

[0005] In view of the above problems, embodiments of this application provide a document risk identification method, apparatus, device, medium, and product that can improve the efficiency of document risk identification.

[0006] According to a first aspect of this application, a document risk identification method is provided, comprising: determining a target document type of a target electronic document from N preset document types; each of the N preset document types corresponds to one of N specific risk identification models; each specific risk identification model is used to identify specific risks in a document of the corresponding preset document type; N is a positive integer; based on the specific risk identification model corresponding to the target document type, identifying the target specific risk of the target electronic document according to the target document characteristics of the target electronic document; based on a general risk identification model, identifying the target general risk of the target electronic document according to the target document characteristics of the target electronic document; and determining the risk identification result of the target electronic document based on the identified target general risk and target specific risk.

[0007] According to an embodiment of this application, the target document features include at least one of the following: document content features, document layout features, and document fusion features; the method for determining the document fusion features includes: for the target electronic document, determining a document layout region; for any document layout region, determining the layout semantic features of the targeted document layout region, and determining the document content features corresponding to the document content located in the targeted document layout region, fusing the determined layout semantic features and the determined document content features to obtain the document fusion features of the targeted document layout region.

[0008] According to an embodiment of this application, the document content features include document content semantic features; the method for determining the document content semantic features includes: for any document content, extracting document content semantic features based on the document content and the context information of the document content.

[0009] According to an embodiment of this application, the document content features include document content editing features; the method for determining the document content editing features includes: for any document content, extracting document content editing features based on the document content and the editing operation information of the document content.

[0010] According to an embodiment of this application, the method further includes: identifying target potential risks of the target electronic document based on the identified target general risks and target specific risks; the target potential risks satisfying preset potential conditions; the preset potential conditions including at least one of the following: the degree of risk correlation between the target potential risks and the identified target general risks is greater than a first correlation threshold, and the degree of risk correlation between the target potential risks and the identified target specific risks is greater than a second correlation threshold; the step of determining the risk identification result of the target electronic document based on the identified target general risks and target specific risks includes: determining the risk identification result of the target electronic document based on the identified target general risks, target specific risks, and target potential risks.

[0011] According to an embodiment of this application, identifying the target potential risk of the target electronic document based on the identified target general risk and target specific risk includes: determining a target risk knowledge graph; the target risk knowledge graph contains risk nodes for characterizing document risks, and the edges connecting different risk nodes in the target risk knowledge graph are used to characterize the association relationship between the connected risk nodes; for any risk node in the target risk knowledge graph, determining the degree of risk association between the risk represented by the targeted risk node and the identified target general risk based on the path between the targeted risk node and the risk node for characterizing the identified target general risk, and determining the degree of risk association between the risk represented by the targeted risk node and the identified target specific risk based on the path between the targeted risk node and the risk node for characterizing the identified target specific risk; and determining the risk represented by the risk node whose determined degree of risk association satisfies the preset potential condition as the target potential risk of the target electronic document.

[0012] According to embodiments of this application, the N preset document types each correspond to N different risk knowledge graphs; each risk knowledge graph includes risk nodes for representing document-specific risks of the corresponding preset document type, and risk nodes for representing general document risks; determining the target risk knowledge graph includes: determining the risk knowledge graph corresponding to the target document type as the target risk knowledge graph.

[0013] According to embodiments of this application, any specific risk identification model includes: P specific risk discrimination sub-models; different specific risk discrimination sub-models are used to identify different specific risks in documents of corresponding preset document types; where P is a positive integer; the general risk identification model includes: Q general risk discrimination sub-models; different general risk discrimination sub-models are used to identify different general risks in documents; where Q is a positive integer; the method further includes: when a new specific risk is added for any preset document type, determining a new specific risk discrimination sub-model for identifying the new specific risk, and adding the determined new specific risk discrimination sub-model to the specific risk identification model corresponding to the preset document type; when a new general risk is added, determining a new general risk discrimination sub-model for identifying the new general risk, and adding the determined new general risk discrimination sub-model to the general risk identification model.

[0014] According to a second aspect of this application, a document risk identification device is provided, comprising: a type determination module, configured to determine a target document type of a target electronic document from N preset document types; the N preset document types each correspond to N specific risk identification models; each specific risk identification model is used to identify specific risks in documents of the corresponding preset document type; N is a positive integer; a specific risk module, configured to identify target specific risks of the target electronic document based on the specific risk identification model corresponding to the target document type and according to the target document characteristics of the target electronic document; a general risk module, configured to identify target general risks of the target electronic document based on the general risk identification model and according to the target document characteristics of the target electronic document; and a result determination module, configured to determine the risk identification result of the target electronic document based on the identified target general risks and target specific risks.

[0015] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0016] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0017] According to a fifth aspect of this application, a computer program product is also provided, comprising a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0018] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0019] Figure 1 This illustration schematically depicts an application scenario of a document risk identification method according to an embodiment of this application.

[0020] Figure 2 A flowchart illustrating a document risk identification method according to an embodiment of this application is shown schematically;

[0021] Figure 3 This schematic diagram illustrates a structural block diagram of a document risk identification device according to an embodiment of this application;

[0022] Figure 4 A block diagram schematically illustrates an electronic device suitable for implementing a document risk identification method according to an embodiment of this application. Detailed Implementation

[0023] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0024] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0025] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0026] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0027] With the popularization of information technology, electronic documents are widely used as a convenient tool in daily life and work. For example, they are used to edit reports for printing or to transmit work information. However, electronic documents themselves also pose risks. For instance, electronic documents may contain typos, have layouts that do not conform to preset formats, or have missing images. Therefore, risk identification of electronic documents is usually necessary. Currently, risk identification of electronic documents is typically done manually, which is relatively inefficient.

[0028] To address the aforementioned technical problems, embodiments of this application provide a document risk identification method. This method utilizes a risk identification model to automatically identify risks in electronic documents, thereby improving the efficiency of document risk identification. The model architecture of the risk identification model can specifically be a neural network model or a large-scale deep learning model.

[0029] Furthermore, electronic documents typically include different document types, such as contracts, invoices, and reports. There can be common document risks (referred to as general risks for easy distinction) across different document types, such as typos or missing key information. Different document types can also have unique, specific document risks (referred to as document type-specific risks for easy distinction), such as inconsistencies in contract clauses or missing signature images in invoice documents.

[0030] General risks can be risks present in documents of multiple document types, or risks present in documents of various document types. Specific risks, also known as particular risks, can be risks unique to a specific document type. Specifically, they can be risks that exist only in documents of that specific document type, risks whose probability of occurrence in documents of that specific document type is higher than a probability threshold, or risks whose probability of occurrence in documents of that specific document type is higher than an upper probability threshold but whose probability of occurrence in documents of other document types is lower than a lower probability threshold. Different document types can each correspond to different specific risks.

[0031] Accordingly, the above methods can employ different risk identification models to identify the specific and general risks of electronic documents. Since different document types correspond to different specific risks, risk identification models can be trained separately for each document type to identify the corresponding specific risks. For ease of differentiation, a general risk identification model can be determined to identify general document risks, and specific risk identification models can be determined for each document type to identify its specific risks. Different document types can each correspond to a different specific risk identification model.

[0032] Therefore, the above methods can identify general risks of electronic documents based on a general risk identification model; alternatively, they can determine the document type of the electronic document and identify specific risks based on a specific risk identification model corresponding to that document type. By utilizing different risk identification models to identify both general and specific risks of electronic documents, different models can focus on identifying either general or specific risks, thus improving the predictive accuracy of the risk identification models and consequently enhancing the accuracy and comprehensiveness of document risk identification.

[0033] In the methods described above, the identified general and specific risks can be directly used as the risk identification results for electronic documents. These risks can include information such as risk location, risk type, and risk severity. For example, a risk might exist in a certain part of the electronic document, such as a typo, or an image might be missing. Alternatively, the identified general and specific risks can be combined to further determine the risk identification results for the electronic document. Specifically, this can involve comprehensively determining the risk severity of the electronic document. If the risk severity exceeds a threshold, document repair operations can be triggered or a document risk warning message can be generated. Alternatively, the severity of both general and specific risks can be determined, and the subsequent document repair direction can be determined based on these determined severity levels.

[0034] In a specific example, if it is determined that the severity of general risks identified for electronic documents is high, while the severity of specific risks is low, then it can be determined that subsequent remedial operations should be performed more on the general risks for electronic documents.

[0035] In the methods described above, the document type of an electronic document can be determined using its document characteristics, such as its content, filename, and layout. Similarly, the risk identification model described above can also perform risk identification and prediction based on the document characteristics of electronic documents.

[0036] Document features for electronic documents can specifically include: document content features and document layout features. Document content includes, for example, document images and text; document content features can be extracted from the document's content. Document layout includes, for example, layout areas such as headers, body text, signatures, and seals; document layout features can be extracted from information about these layout areas.

[0037] In a specific example, documents of different types typically have different document layouts. Contract documents and invoices differ significantly in their layouts. Therefore, document layout characteristics can help identify document types, improving the accuracy of document type identification. Document risks can sometimes also be reflected in document layout, such as missing headings, missing headers, or missing seals. Thus, document layout characteristics can help identify document risks, improving the accuracy of document risk identification.

[0038] Regarding document features, document content features and document layout features can be further fused to obtain document fusion features, which can be used as input features for document type determination or document risk identification. Specifically, for any document layout region, the corresponding layout semantic features can be determined, and the content features of the document content located within that layout region can be fused with these layout semantic features to obtain the document fusion features. By leveraging the positional association between the layout region and the document content (i.e., the document content located within the layout region), the document fusion features can be extracted and used accordingly to determine the document type or identify document risks.

[0039] In a specific example, documents of different types may contain different content within the same semantically defined document layout area. For instance, the content of contract documents and report documents in their body heading layout areas often differs significantly. Therefore, document fusion features can help identify document types, improving the accuracy of document type identification. Document type can also sometimes be reflected in a mismatch between document layout and content. For example, for a layout area semantically designated as a "seal area," if the document content within that area is a signature image, there is a risk of a mismatch between document layout and content. Document fusion features can then help identify this document risk, improving the accuracy of document risk identification.

[0040] Regarding document content features, one can either treat the content of an electronic document as a whole and extract its features, or divide the content into one or more sections and extract features from each section separately. For example, an electronic document may include multiple paragraphs of text and multiple images. Therefore, the text and images can be divided into paragraphs to obtain multiple sections of content, from which features can be extracted. Of course, other methods can also be used for content segmentation. Specifically, semantic features of the document content can be extracted to determine the document type or identify document risks.

[0041] Furthermore, for any document content, its semantic features can be determined by combining the document's content with its context. For example, for an image in an electronic document, its semantic features can be extracted by combining the textual context, including the image's name, preceding description, or following explanation, thereby improving the accuracy of the semantic features.

[0042] Regarding document features, editing operations on electronic documents can be combined to extract relevant content editing features, which can be used to determine document type or identify document risks. Specific examples of content editing features include features indicating whether document content is copied, features indicating the number of edits to a document, features indicating the duration of editing, features indicating the user who edited the document, and features indicating the historical edits of a document. It is understandable that content editing features, such as historical edits, can help determine document type and identify document risks. Of course, content editing features can include features of the editing operation itself, as well as features resulting from the fusion of editing operations and document content. Specifically, this involves fusing the editing operation features with the content features of the document to obtain fused features.

[0043] In a specific example, based on content editing characteristics, document content that has been copied but not modified, document content edited by users whose historical editing error rate exceeds a threshold, and document content whose number of edits exceeds a threshold can be identified. The identified document content can be regarded as content that is easy to identify as risky. Therefore, in the process of document risk identification, the identification weight of the identified document content can be increased, which can improve the accuracy of document risk identification.

[0044] In the methods described above, risks within the same document are correlated. For example, if an electronic document contains a risk of missing images, the corresponding context usually contains a risk of mismatch between text and images; similarly, if multiple typos exist in the contract clauses of a contract document, there is usually a risk of inconsistencies in the contract clauses. Therefore, by leveraging the identified document risks (including general and specific risks), potential risks within the same document can be identified and determined accordingly. Potential risks may be those that are difficult to identify or determine using risk identification models, but based on the identified document risks and the correlations between them, it can be determined that the probability of these potential risks existing in the document is relatively high. Subsequently, other methods or means can be used to determine whether these potential risks actually exist in the document, thereby improving the accuracy and comprehensiveness of document risk identification.

[0045] Specifically, potential risks can be those risks that are not identified by the risk identification model but whose correlation with the identified general risks is greater than a threshold, and / or whose correlation with the identified specific risks is greater than a threshold.

[0046] The degree of correlation between different risks can be determined based on the probability or frequency of different risks existing in the same document, or it can be determined based on predictions from a pre-trained model, or it can be determined based on a risk knowledge graph.

[0047] The above method can be combined with a risk knowledge graph to determine the relationships and degrees of association between different risks. A risk knowledge graph can include: risk nodes representing risks, and edges connecting different risk nodes and representing the relationships between risks. It can be understood that, based on the edges connecting different risk nodes, the degree of risk association can be determined using the represented risk relationships. Specifically, the degree of risk association can be determined based on the tightness of the risk association, combined with information such as the probability or frequency of different risks existing in the same document. Correspondingly, for any two risk nodes, the degree of association between the risks represented by these two risk nodes can be determined based on the path connecting them.

[0048] Furthermore, considering specific risks, the risk knowledge graph can include risk nodes to represent specific risks. For any given document type, specific risks for other document types are generally not used during document risk identification. Therefore, different risk knowledge graphs can be constructed for different document types. Each risk knowledge graph can contain risk nodes to represent specific risks for the corresponding document type, as well as risk nodes to represent general risks, but not risk nodes to represent specific risks for other document types.

[0049] Understandably, the degree of risk correlation can be predetermined. Once general and specific risks are identified, potential risks can be identified and determined based on the predetermined degree of risk correlation. Alternatively, the degree of risk correlation can be determined in real time. This involves continuously updating the risk knowledge graph and the probability or frequency of different risks existing in the same document. Based on the current risk knowledge graph or other information, the degree of risk correlation can be determined in real time, and potential risks can be identified accordingly, improving the accuracy and timeliness of identifying potential risks.

[0050] In the above methods, for the general risk identification model and the specific risk identification model, since there are many types of risks in electronic documents, each risk identification model can contain multiple risk discrimination sub-models. The risk discrimination sub-models can be used to identify one or more risks, and together they can serve as the output layer of the risk identification model or form the output layer of the risk identification model. The document risks are identified based on the features extracted from the risk identification model.

[0051] For example, general document risks can include 10 general risk types, allowing the deployment of 10 risk discrimination sub-models, each corresponding to one of the 10 general risk types. Each risk discrimination sub-model can be used to identify its corresponding general risk type. The output layer of the general risk identification model can include these 10 risk discrimination sub-models. The output layer can use the features extracted from the hidden layer to identify risks through these 10 risk discrimination sub-models, determining whether the electronic document contains any of the 10 general risk types. The explanation of specific risks can be found in this example.

[0052] By using multiple risk discrimination sub-models, the risks identified by the risk identification model can be easily and flexibly adjusted, and individual risk discrimination sub-models can be easily trained, reducing the training cost of the risk identification model.

[0053] When a new risk type is added, a corresponding risk discrimination sub-model can be trained to identify the new risk and added to the risk identification model. Conversely, when a risk type is deleted, the risk discrimination sub-model used to identify that risk type in the risk identification model can be deleted; similarly, when a risk type is modified, the risk discrimination sub-model used to identify that risk type in the risk identification model can be modified or replaced. Therefore, based on the architecture of the risk discrimination sub-model, updating risk types does not require retraining the entire risk identification model, improving the flexibility and convenience of risk identification model updates. Both general-purpose and specific risk identification models can be updated based on the architecture of the risk discrimination sub-model.

[0054] It should be noted that the document risk identification method and apparatus provided in the embodiments of this application can be applied to the fields of artificial intelligence technology and fintech. For example, for electronic documents in financial institutions such as banks, the document risk identification method provided in the embodiments of this application can be used for risk identification. The document risk identification method and apparatus provided in the embodiments of this application can also be applied to any field other than fintech, and the application field of the document risk identification method and apparatus provided in the embodiments of this application is not limited.

[0055] In the technical solution of this application, the user information (including but not limited to the user's historical operation records, the user's document editing records, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure and application of related data all comply with relevant laws, regulations and standards, necessary confidentiality measures have been taken, they do not violate public order and good morals, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0056] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points for users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process enters the expert decision-making process.

[0057] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0058] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0059] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0060] In one or more embodiments described herein, the term "large model" can refer to a deep learning model with a large number of model parameters, which can include hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. Large models can also be called foundational models or basic models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability, such as large language models and multimodal pre-trained models. It should be understood that in practical applications, large models only require a small number of samples to fine-tune the pre-trained model before being applied to different tasks. Large models can be widely used in natural language processing, computer vision, and other fields. Specifically, they can be applied to computer vision tasks such as visual question answering, image captioning, and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. Major application scenarios for large models can include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0061] Figure 1 The illustration shows an application scenario diagram of a document risk identification method according to an embodiment of this application.

[0062] like Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send information, etc.

[0063] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal devices. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, and it can include, but is not limited to, applications, mini-programs, etc., installed on the terminal devices, and can also be in the form of web pages.

[0064] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. Server 105 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data. Server 105 can be the backend server for the aforementioned target application, used to provide backend services to the clients of the target application.

[0065] It should be noted that the document risk identification method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-103. Correspondingly, the document risk identification device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-103.

[0066] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0067] Figure 2 A flowchart illustrating a document risk identification method according to an embodiment of this application is shown.

[0068] like Figure 2 As shown, a document risk identification method 200 according to an embodiment of this application may include steps S210 to S240.

[0069] The embodiments of this application do not limit the implementing entity of a document risk identification method. Optionally, it can be applied to any electronic device or any software application, specifically to a server, client, or risk detection application.

[0070] In step S210, the target document type of the target electronic document is determined from N preset document types; each of the N preset document types corresponds to one of the N exclusive risk identification models; any exclusive risk identification model is used to identify the exclusive risk in the document of the corresponding preset document type; N is a positive integer.

[0071] In step S220, based on the specific risk identification model corresponding to the target document type, the target specific risk of the target electronic document is identified according to the target document characteristics.

[0072] In step S230, based on the general risk identification model, the target general risks of the target electronic document are identified according to the target document characteristics.

[0073] In step S240, the risk identification result of the target electronic document is determined based on the identified target general risks and target specific risks.

[0074] This method can automatically identify risks in electronic documents based on a risk identification model, thereby improving the efficiency of document risk identification.

[0075] Furthermore, by identifying general risks and specific risks based on general risk identification models and specific risk identification models respectively, the accuracy and comprehensiveness of document risk identification can be improved.

[0076] The embodiments of this application are not limited to electronic documents. Optionally, the target electronic document can be any electronic document, and any electronic document that requires document risk identification can be identified as the target electronic document. The form of the target electronic document is not limited; specifically, it can be an image-based electronic document or a document-based electronic document. For example, a paper document can be photographed or scanned to obtain an image-based electronic document.

[0077] It is understood that electronic documents can include both document content and document layout. The embodiments of this application do not limit the document content; specifically, it can include image content and text content. In an electronic document in image form, the text content can be extracted through image text recognition. The embodiments of this application do not limit the document layout; specifically, it can include the typesetting structure and spatial organization of various content elements in the document, such as headers, footers, columns, titles, body text, paragraphs, tables, formulas, signature areas, and seal areas. Different document types can have different document layouts. For example, the layout of a document of the form "invoice" can differ from the layout of a document of the form "contract".

[0078] The embodiments of this application do not limit the document type. Optionally, the document type may include: contract type, invoice type, and report type, etc. The target document type may be the document type of the target electronic document. The document type can usually be determined by information such as the document title, the document file name, or the document layout.

[0079] Regarding step S210 above, the embodiments of this application do not limit the specific method of determining the target document type of the target electronic document. Optionally, the target electronic document may contain information for characterizing the target document type, and may store a pre-determined target document type, thereby directly determining the target document type; or the target document type may be determined based on the name or title of the target electronic document, specifically based on keywords in the name or title of the target electronic document. For example, if the title of the target electronic document contains the keyword "contract," the target document type can be determined to be a contract type; or the target document type may be determined based on the document features of the target electronic document. Specific explanations of document features can be found in other embodiments. Specifically, for example, the target document type can be predicted and determined based on the document content or document layout in the target electronic document, using a pre-trained type prediction model.

[0080] Understandably, this can be done by selecting one of N preset document types as the target document type using the methods described above. For example, keywords can be set corresponding to each of the N preset document types, allowing the target document type to be determined based on the keywords contained in the name or title of the target document. Alternatively, document type labels can be assigned to N preset document types for training samples, which are then used to train a type prediction model.

[0081] The embodiments of this application do not limit the distinction between general risks and specific risks. Optionally, general risks may include risks present in documents of different document types, such as the risk of typos, missing information, mismatched information, or skewed horizontal or vertical lines. Specific risks may include risks present only in documents of a single document type, or risks that have a probability higher than an upper probability threshold in a single document type but a probability lower than a lower probability threshold in other document types. The upper probability threshold may be greater than the lower probability threshold. For example, inconsistencies in clauses in contract documents or missing signature images in bill documents. By distinguishing between general and specific risks, it is convenient to identify general and specific risks separately for target electronic documents. After determining the document type of the target electronic document, the scope of specific risks can be narrowed down to the specific risks present in that document type, and general risks and specific risks can be identified separately, thereby improving the accuracy, precision, and comprehensiveness of document risk identification.

[0082] The embodiments of this application do not limit the specific content and form of general risks and specific risks. Optionally, general risks may include at least one of the following: general risk type, general risk location, and general risk severity (which may be in the form of risk levels); specific risks may include at least one of the following: specific risk type, specific risk location, and specific risk severity. Specific examples of risk types include typo risk, information missing risk, and information mismatch risk. Specific examples of risk locations include the document content or document location where the risk exists. It is understood that document risk identification may include identifying at least one of the following in the document: risk type, risk location, and risk severity.

[0083] The embodiments of this application do not limit the risk identification model. Optionally, for the risk identification model, a general risk identification model for identifying common risks and a specific risk identification model for identifying specific risks can be set separately. Further, for different document types, specific risk identification models for identifying corresponding specific risks can be set separately, thereby setting different specific risk identification models corresponding to different document types. Therefore, N preset document types can each correspond to N specific risk identification models; the specific risk identification model can be used to identify specific risks in documents of the corresponding preset document type. Of course, a single risk identification model can also be set to identify specific risks for documents of each preset document type. By using different risk identification models to identify common and specific risks of electronic documents respectively, the general risk identification model can focus on identifying common risks, and the specific risk identification model can focus on identifying specific risks, or specifically, it can focus on identifying specific risks of a single document type, which can improve the predictive accuracy of the risk identification model, and correspondingly improve the accuracy and comprehensiveness of document risk identification.

[0084] Optionally, for the general risk identification model, considering the need to identify general risks across different document types, the features extracted from the general risk identification model can be trained to reduce the influence of different document types on the extracted features. This allows for the extraction of general features, making it difficult to determine the corresponding document type based solely on general document features. This facilitates general risk identification based on common features across different document types. Training the general risk identification model can be done using training samples from different document types, facilitating the learning and extraction of common features across different document types. Specific training methods can be found in the explanations of other embodiments. For example, text content features can be extracted from the document.

[0085] For a dedicated risk identification model, the features extracted from the model can be used to highlight the specific characteristics of the corresponding document type during model training. This facilitates dedicated risk identification based on the specific features of that document type. Training the dedicated risk identification model can be done using training samples of documents of the corresponding document type, making it easier to learn and extract the specific features of that document type. For example, extracting clause features from contract documents or table features from invoice documents.

[0086] The embodiments of this application do not limit the architecture of the risk identification model. Optionally, the general risk identification model and the specific risk identification model can be any deep learning model architecture, specifically a neural network model or a large model architecture. The general risk identification model and the specific risk identification model can use the same model architecture or different model architectures. The number of model parameters in the general risk identification model and the specific risk identification model can be the same or different.

[0087] In an optional embodiment, considering that the general risk identification model needs to identify risks for documents of various document types, the number of model parameters of the general risk identification model can be set to be greater than the threshold of the number of model parameters, so as to improve the general risk identification accuracy of the general risk identification model.

[0088] The embodiments of this application do not limit the training method of the risk identification model. Optionally, a general risk label can be determined based on the document features in the document training samples to construct a general risk document training sample set for training the general risk identification model; alternatively, a specific risk label can be determined based on the document features in the document training samples according to the document type of the document training samples to construct a specific risk document training sample set for that document type for training the specific risk identification model corresponding to that document type.

[0089] In one optional embodiment, the general risk document training sample set may contain document training samples of different document types. To reduce the impact of different document types on general risk identification, the hidden layer in the general risk identification model can be defined as a generator of general document features, and the already trained document type identification model can be defined as a discriminator for document types. Based on the hidden layer's input document features, general document features are extracted, and these features are input into the discriminator to identify document types. Accordingly, the loss can be determined based on the discriminator's accuracy, and the parameters of the hidden layer are updated based on the loss. The loss can be positively correlated with the discriminator's accuracy; the higher the accuracy, the greater the loss. This allows the parameters of the hidden layer to be updated in a direction that reduces both the loss and the accuracy, further reducing the impact of different document types on the general document features extracted by the hidden layer. Specifically, the document type identification accuracy can be used as one of the factors to measure the loss, combined with the general risk identification accuracy, to jointly determine the loss of the general risk identification model. The parameters of the general risk identification model are then updated based on the loss to reduce the document type identification accuracy and improve the general risk identification accuracy. Of course, you can also train the hidden layer using a gradient inversion layer, and update the parameters of the hidden layer in the direction of reducing the accuracy of document type discrimination through the gradient inversion layer.

[0090] Therefore, optionally, during the training process of the general risk identification model, the model loss of the general risk identification model can be negatively correlated with the accuracy of general risk identification, and the model loss of the general risk identification model can be positively correlated with the accuracy of document type discrimination based on the general risk features extracted from the hidden layer of the general risk identification model. The model parameters of the general risk identification model can be updated in the direction of improving the accuracy of general risk identification and reducing the accuracy of document type discrimination based on general risk features.

[0091] Furthermore, the embodiments of this application do not limit the relationship between the general risk identification model and the specific risk identification model. Optionally, the general risk identification model and the specific risk identification model can be two independent models, which are built and trained independently. The general risk identification model and the specific risk identification model can also share a feature extraction layer or a hidden layer, and further include different output layers to output different risk identification results. Based on the shared feature extraction layer or the shared hidden layer between the general risk identification model and the specific risk identification model, overall model training can be facilitated, reducing model training costs.

[0092] The embodiments of this application do not limit document features. Optionally, the target document features can be the document features of the target electronic document, such as document content features and document layout features. Document features can be used as input features of a risk identification model, so that the risk identification model can identify risks in the document. Optionally, the document features may include at least one of the following: document content features, document layout features, document fusion features, and document content editing features.

[0093] Document content features can be features extracted from image or text content within the document, such as semantic or image features of the document's image content, or semantic or text features of the document's text content. Document layout features can be features extracted from the typesetting structure and spatial organization of content elements within the document, such as the size and position of heading areas, the position of paragraphs, and the size and position of signature areas. Furthermore, document content features can also include style features, such as the size, font, and color of text content, and the color of a seal image.

[0094] Based on the aforementioned document content and layout features, risks within documents can be easily identified using risk identification models. For example, textual features of the document's text content can identify the risk of typos; semantic features of the contract document's text content can identify the risk of inconsistencies in clauses; and document layout features can identify the risk of inaccurate placement of the signature area within the document.

[0095] Document fusion features can be obtained by fusing other features in a document according to their relationships. The embodiments of this application do not limit document fusion features. Optionally, document content features and document layout features can be fused. Specifically, the document content features of the document content located in the document layout area can be fused with the layout features of that layout area to obtain document fusion features. The fused layout features can specifically be layout semantic features, such as the functional semantics or role semantics represented by the layout area. For example, in a document layout, the layout semantic features of the title area can be used to represent "title," and the layout semantic features of the signature area can be used to represent "signature." By fusing document layout features and document content features, the relationship between document layout and document content can be reflected, which can be used for document risk identification and improve the accuracy and comprehensiveness of document risk identification. For example, based on document fusion features, the risk of a mismatch between the semantics of a document layout area and the content in the layout area can be identified. Specifically, for example, the signature area contains a seal image instead of a signature image. Of course, based on document fusion features, other risks can also be identified, such as the risk that the content in the layout area does not conform to the specifications of that layout area.

[0096] Furthermore, other document features can be fused to obtain document fusion features. For example, document content features and document content editing features can be fused together to obtain document fusion features. For a more detailed explanation, please refer to the explanations in other embodiments.

[0097] Therefore, optionally, the target document features may include at least one of the following: document content features, document layout features, and document fusion features. The method for determining the document fusion features may include: determining a document layout region for the target electronic document; determining the layout semantic features of any document layout region, and determining the document content features corresponding to the document content located within the document layout region; fusing the determined layout semantic features and the determined document content features to obtain the document fusion features of the document layout region. This embodiment can fuse layout semantic features and document content features to obtain document fusion features for document risk identification, which can improve the amount and dimension of information used in document risk identification, and thus improve the accuracy of document risk identification.

[0098] Of course, the target document features can also include other document features, such as document content editing features, specifically features extracted from document content editing records or features extracted from document content editing users. Document fusion features can also integrate multiple different document features, such as integrating document content editing features and document content features, or integrating document content editing features, document layout features, and document content features. Optionally, for any document content, the content features and document content editing features of the target document content can be integrated to obtain the document fusion features for that document content. The document fusion features obtained by integrating content features and editing features can be used as input features for a risk identification model, increasing the amount and dimension of information used in document risk identification and improving its accuracy.

[0099] The embodiments of this application are not limited to document content features. Optionally, document content features may include at least one of the following: document image content features and document text content features. Specifically, these may include features in the form of document content semantic features, feature maps of document image content, text vectors of document text content, etc.

[0100] The embodiments of this application do not limit the specific method of determining document content features. Optionally, the document content can be divided into one or more document contents for the target electronic document, and then document content features can be extracted for each document content. Specifically, a pre-trained machine learning model can be used to extract document content features. For example, for image-based document content, an image feature extraction model can be used to extract document image content features, and for text-based document content, a text feature extraction model can be used to extract document text content features. Optionally, semantic features of document content can be extracted based on a pre-trained machine learning model, such as a semantic feature extraction model.

[0101] In an optional embodiment, the extraction of semantic content can be combined with contextual information of the document content, which can improve the accuracy of semantic features. Therefore, optionally, document content features may include semantic features; the determination of semantic features may include: for any document content, extracting semantic features based on the document content and its contextual information. This embodiment can combine document content and contextual information to extract semantic features, which can improve the accuracy of semantic features and consequently improve the accuracy of document risk identification.

[0102] In a specific example, for image content in document form, the target electronic document usually includes an explanation or introduction to the image content in the context, so that the semantic features of the document content can be extracted by combining the contextual information and the image content itself.

[0103] This embodiment does not limit the specific method for extracting semantic features from document content. Optionally, it can be based on a pre-trained machine learning model, such as a semantic feature extraction model, to extract semantic features from the document content and its contextual information. Of course, other models or model layers, or other methods, can also be used to extract semantic features.

[0104] The embodiments of this application do not limit the way the document content is divided. Optionally, it can be divided according to the form of the document content in the target electronic document, or it can be divided according to paragraphs, or it can be divided by a pre-trained segmentation model. Different division methods can be combined with each other.

[0105] The embodiments of this application do not limit the document content editing features. Optionally, document content editing features can be features extracted based on editing operations performed on the document content. Specifically, these include features characterizing whether the document content is copied, features characterizing the number of edits to the document content, features characterizing the duration of editing the document content, features characterizing historical editing operation records for the document content, features characterizing the user editing the document content, features characterizing the historical edited content of the document content, features characterizing the historical editing error rate of the user editing the document content, etc. It is understood that document content editing features can be features extracted based on information related to document content editing operations. Furthermore, document content editing features can also include fused features obtained by combining document content features and editing operation features.

[0106] The embodiments of this application do not limit the method for determining document content editing features. Optionally, document content editing features can be extracted based on the editing operation information of the document content; document content editing features can also be extracted based on the information of the editing user who performed the editing operation on the document content; or the content features of any document content and the editing operation information features for that document content can be fused to obtain document fusion features as document content editing features. Specifically, document content editing features can be extracted and determined based on a pre-trained machine learning model, such as an editing feature extraction model.

[0107] Optionally, the document content features may include document content editing features; the method for determining document content editing features may include: for any document content, extracting document content editing features based on the document content and the editing operation information of the document content. This embodiment can combine document content and editing operation information to extract document content editing features, which can improve the accuracy of document content editing features and correspondingly improve the accuracy of document risk identification.

[0108] This embodiment does not limit the information on editing operations of the document content. Specifically, it may include information about the editing user who performs the editing operation on the document content, historical editing operation records of the document content, and historical editing version information of the document content.

[0109] The embodiments of this application do not limit the specific method of extracting document content editing features. Optionally, based on a pre-trained machine learning model, such as an editing feature extraction model, document content editing features can be extracted for any document content according to the document content and the editing operation information of the document content; alternatively, for any document content, the content features of the document content and the editing features of the editing operation information of the document content can be fused to obtain document content editing features. Specifically, it can also be based on a pre-trained machine learning model that fuses content features and editing features.

[0110] In a specific example, document content editing features can be used as target document features and input into a risk identification model for document risk identification. This adds the dimension of document content editing operations to the document risk identification process, thereby improving the accuracy of risk identification. For instance, when identifying typos in a document, the historical typo probability of the editing user performing the editing operation can be incorporated for identification.

[0111] In one optional embodiment, document content with a risk probability higher than a threshold can be identified based on document content editing characteristics. A risk identification weight is then added to the identified document content to focus on identifying whether any risk exists within it. This embodiment does not limit the method for determining the risk probability of document content. Specifically, based on document content editing characteristics, at least one of the following can be identified as document content with a risk probability higher than the threshold: directly copied document content, document content with more than a threshold number of edits, and document content edited by users with a historical error rate higher than the threshold. Alternatively, the risk probability of document content can be predicted based on a pre-trained risk probability prediction model using the document content editing characteristics.

[0112] The above embodiments explain the target document features, which can be input into the dedicated risk identification model and the general risk identification model for document risk identification.

[0113] Regarding step S220 above, the embodiments of this application do not limit the specific process for identifying target-specific risks. Optionally, based on the target document type, from N preset document type-specific risk identification models, a target document type-specific risk identification model is determined, and the target document features are input into the determined target document type-specific risk identification model to obtain the target-specific risk identification result output by the target document type-specific risk identification model, which is the target-specific risk. Optionally, the target-specific risk identification model can identify document risks based on the input document features and output the target-specific risk identification result.

[0114] The embodiments of this application do not limit the content of the target-specific risk. Optionally, the identified target-specific risk may include at least one of the following: the risk type of the target-specific risk, the location of the target-specific risk, and the severity of the target-specific risk. Specifically, the location of the target-specific risk may be a document location where the target-specific risk exists, such as the document content or document layout area where the target-specific risk exists.

[0115] In a specific example, the target document type can be a contract type, and the target-specific risk can be identified based on the exclusive risk identification model corresponding to the contract type. This can include: (1) clause contradiction type risk, which is located in the first and second clauses, and the risk severity is mild; (2) signature missing type risk, which is located in the signature area, and the risk severity is severe.

[0116] Optionally, the specific risk identification model can be a large-scale model architecture, allowing the identification of target-specific risks based on this architecture. In a specific example, the target electronic document and its features can be input into the specific risk identification model within the large-scale model architecture. Specific risk identification can then be performed using prompts to obtain the target-specific risk.

[0117] Regarding step S230 above, the embodiments of this application do not limit the specific process for identifying target general risks. Optionally, the target document features can be input into a general risk identification model to obtain the general risk identification result output by the general risk identification model, which is the target general risk. Optionally, the general risk identification model can identify document risks based on the input document features and output a general risk identification result.

[0118] The embodiments of this application do not limit the content of the target general risk. Optionally, the identified target general risk may include at least one of the following: the risk type of the target general risk, the location of the target general risk, and the severity of the target general risk. Specifically, the location of the target general risk may be a document location where the target general risk exists, such as the document content or document layout area where the target general risk exists.

[0119] In a specific example, based on the general risk identification model, the target general risk can be identified for the target electronic document, which may include: (1) typo type risk, located at page 3, line 4, with a risk severity of mild; (2) information missing type risk, located at page 5, line 2, with a blank space, with a risk severity of severe.

[0120] Optionally, the general risk identification model can be a large model architecture, allowing the identification of target general risks based on this architecture. In a specific example, the target electronic document and its features can be input into the general risk identification model within the large model architecture, and the target general risk can be obtained through prompt words.

[0121] The embodiments of this application do not limit the order of steps in the above method flow. Optionally, steps S210 to S240 can be executed sequentially, or steps S220 and S230 can be executed in parallel. Since step S230 does not need to determine the target document type, steps S210 and S220 can also be regarded as a whole step of identifying specific risks and executed in parallel with step S230. Step S240 can be executed after steps S220 and S230 have been executed.

[0122] In one optional embodiment, the risk identification model architecture can also deploy multiple sub-models, each designed to identify different types of risks. This allows for convenient adjustments to the sub-models as risk identification needs evolve. For example, if identifying a certain type of risk is no longer required, the sub-model identifying that risk can be deleted; if identifying a new type of risk is needed, a new sub-model can be added; and if updating a certain type of risk is required, the sub-model identifying that risk can be updated. This model architecture enhances the flexibility of adjusting the risk identification model.

[0123] In a specific example, document general risks can include 10 general risk types, thus allowing the deployment of 10 risk discrimination sub-models, each corresponding to one of the 10 general risk types. Each risk discrimination sub-model can be used to identify its corresponding general risk type. In the general risk identification model, the risk discrimination sub-models can be deployed at the output layer, and the features extracted from the hidden layer can be input into the aforementioned 10 risk discrimination sub-models to identify risks and determine whether the electronic document contains the aforementioned 10 general risk types. Alternatively, the general risk identification model can also involve inputting the target document features into the aforementioned 10 risk discrimination sub-models to identify risks and determine whether the electronic document contains the aforementioned 10 general risk types. The explanation of specific risks can be found in this example.

[0124] Therefore, optionally, any specific risk identification model may include: P specific risk discrimination sub-models; different specific risk discrimination sub-models are used to identify different specific risks in documents of the corresponding preset document type; P is a positive integer; the general risk identification model may include: Q general risk discrimination sub-models; different general risk discrimination sub-models are used to identify different general risks in documents; Q is a positive integer. The above method flow may also include at least one of the following: (1) when a new specific risk is added for any preset document type, a new specific risk discrimination sub-model for identifying the new specific risk is determined, and the determined new specific risk discrimination sub-model is added to the specific risk identification model corresponding to the preset document type; (2) when a new general risk is added, a new general risk discrimination sub-model for identifying the new general risk is determined, and the determined new general risk discrimination sub-model is added to the general risk identification model; (3) when a specified specific risk is deleted for any preset document type, the specific risk discrimination sub-model used to identify the deleted specified specific risk is deleted from the specific risk identification model corresponding to the preset document type. (4) In the case of deleting a specified general risk, in the general risk identification model, delete the general risk identification sub-model used to identify the deleted specified general risk; (5) In the case of modifying a specified specific risk for any preset document type, in the specific risk identification model corresponding to the preset document type, replace the specific risk identification sub-model used to identify the specified specific risk before modification with the specific risk identification sub-model used to identify the specified specific risk after modification; (6) In the case of modifying a specified general risk, in the general risk identification model, replace the general risk identification sub-model used to identify the specified general risk before modification with the general risk identification sub-model used to identify the specified general risk after modification.

[0125] This embodiment can adapt to changes in risk based on the architecture of the risk discrimination sub-model, and adjust the risk discrimination sub-model accordingly, thereby improving the flexibility of the risk identification model. Of course, the risk identification model can also adopt other architectures, as detailed in the explanations of other embodiments. The risk discrimination sub-model can include a dedicated risk discrimination sub-model and a general risk discrimination sub-model.

[0126] The embodiments of this application do not limit the number of risks identified by a single risk discrimination sub-model. Optionally, a single risk discrimination sub-model can be used to identify one or more risks, and different risk discrimination sub-models can be used to identify different risks. The embodiments of this application do not limit the architecture of the risk discrimination sub-model. Optionally, the risk discrimination sub-model can be a lightweight model, a model with fewer than a threshold of parameters, thereby facilitating model training and improving the model training efficiency of the risk discrimination sub-model. It is understood that the new specific risk discrimination sub-model, the new general risk discrimination sub-model, the specific risk discrimination sub-model used to identify the modified specified specific risk, and the general risk discrimination sub-model used to identify the modified specified general risk can be obtained by training the model according to the corresponding risk type.

[0127] The embodiments of this application do not limit the input features of the risk discrimination sub-model. Optionally, the target document features can be input into a dedicated risk discrimination sub-model or a general risk discrimination sub-model for risk identification; alternatively, the target document features can be input into a risk identification model, hidden features can be extracted based on the hidden layer in the risk identification model, and the hidden features can be input into a dedicated risk discrimination sub-model or a general risk discrimination sub-model for risk identification.

[0128] Regarding step S240, the embodiments of this application do not limit the specific method for determining the risk identification result of the target electronic document. The risk identification result of the target electronic document can be referred to as the comprehensive risk identification result. Optionally, the identified general risks and specific risks can be determined as the comprehensive risk identification result of the target electronic document; alternatively, a risk report can be generated by combining the identified general risks and specific risks, serving as the comprehensive risk identification result of the target electronic document; or the overall risk severity can be determined by combining the identified general risks and specific risks, and the determined risk severity can be used as the comprehensive risk identification result.

[0129] In an optional embodiment, to improve the comprehensiveness of document risk identification, potential risks can be identified among those not identified by the general risk identification model and the specific risk identification model. For the identified potential risks, their actual existence can be determined through other methods. Accordingly, the risk identification result of the target electronic document is determined based on the identified target general risks and target specific risks. Specifically, this can include determining the risk identification result of the target electronic document based on the identified target general risks, target specific risks, and potential risks. This can be achieved by using the identified target general risks, target specific risks, and potential risks to determine the risk identification result of the target electronic document; alternatively, it can be by comprehensively considering the identified target general risks, target specific risks, and potential risks to generate a risk assessment report as the risk identification result of the target electronic document.

[0130] The embodiments of this application do not limit the specific method of identifying potential risks. Optionally, potential risks can be identified based on a pre-trained machine learning model, according to the identified target-specific and general risks, or more specifically, potential risks can be identified based on the correlation between risks.

[0131] For example, if the probability or frequency of typo risk (a general risk) and clause contradiction risk (a specific risk) appearing in the same document exceeds a threshold, then even if the typo risk is identified but the clause contradiction risk is not identified based on the general risk identification model and the specific risk identification model, the clause contradiction risk can be identified as a potential risk based on this correlation. This potential risk can then be verified and confirmed through other methods. Alternatively, training samples can be constructed based on the correlation between risks to train the potential risk identification model. The identified target general risks and target specific risks can then be input into the potential risk identification model to predict and determine potential risks.

[0132] Therefore, based on the degree of correlation between risks, and based on the identified target general risks and target specific risks, the target potential risks of the target electronic document can be identified from among the risks not identified by the general risk identification model and the specific risk identification model. That is, the target potential risks of the target electronic document are identified from among the risks other than the identified target general risks and target specific risks. The target potential risk can be any potential risk of the target electronic document. The embodiments of this application do not limit the specific content of the target potential risk. Optionally, the target potential risk may include at least one of the following: the risk type of the target potential risk, the location of the target potential risk, and the severity of the target potential risk, etc.

[0133] Optionally, the above method may further include: identifying potential risks of the target electronic document based on the identified target general risks and target specific risks; the potential risks of the target satisfying preset potential conditions. The preset potential conditions may include at least one of the following: (1) the degree of risk correlation between the potential risks of the target and the identified target general risks is greater than a first correlation threshold, (2) the degree of risk correlation between the potential risks of the target and the identified target specific risks is greater than a second correlation threshold. Determining the risk identification result of the target electronic document based on the identified target general risks and target specific risks may specifically include: determining the risk identification result of the target electronic document based on the identified target general risks, target specific risks, and target potential risks. This embodiment can further identify potential risks based on target general risks and target specific risks, which can improve the accuracy and comprehensiveness of document risk identification. Among them, the target potential risks may be risks other than the identified target general risks and target specific risks.

[0134] The embodiments of this application do not limit the first correlation threshold and the second correlation threshold. Optionally, the first correlation threshold and the second correlation threshold can be the same value or different values.

[0135] The embodiments of this application do not limit the specific method of identifying potential risks of a target. Optionally, potential risks of a target can be determined based on a pre-trained potential risk identification model, according to the identified general risks and specific risks of the target; alternatively, for risks other than the identified general risks and specific risks of the target, the degree of risk correlation with the general risks of the target and the degree of risk correlation with the specific risks of the target can be determined, thereby screening and determining potential risks of the target according to preset potential conditions.

[0136] The embodiments of this application do not limit the method for determining the degree of risk association between different risks. Optionally, the degree of risk association between different risks can be determined based on the probability, frequency, or frequency of different risks appearing in the same document. The probability, frequency, or frequency of any two risks appearing in the same document can be positively correlated with the degree of risk association between the two risks. Alternatively, the degree of risk association can be determined based on a pre-trained association prediction model for different risks.

[0137] Alternatively, knowledge graphs can be used to determine the degree of risk correlation between different risks.

[0138] In one alternative embodiment, a risk knowledge graph can be constructed. The risk knowledge graph may include: risk nodes for representing risks, and edges connecting different risk nodes and representing the relationships between risks.

[0139] Understandably, based on the edges connecting different risk nodes, the degree of risk correlation can be determined using the represented risk relationships. Specifically, the degree of risk correlation can be determined based on the tightness of the risk relationships, or by combining the tightness of the risk relationships with information such as the probability, frequency, or other information about the presence of different risks in the same document. The degree of risk correlation can be positively correlated with the tightness of the risk relationships.

[0140] Accordingly, for any two risk nodes, the degree of association between the risks represented by the two risk nodes can be determined based on the edges contained in the path connecting the two risk nodes.

[0141] In a specific example, in a risk knowledge graph, edges connecting different risk nodes can have edge weights. These weights characterize the strength of the relationship between the risks represented by the connected risk nodes, and can be positively correlated with the strength of the relationship. Accordingly, for any two risk nodes, the degree of correlation between the risks represented by those two nodes can be determined based on the edge weights of the edges in the path connecting them. It is understandable that the number of edges in the path can be negatively correlated with the degree of risk correlation.

[0142] The embodiments of this application are not limited to a risk knowledge graph. Optionally, the risk knowledge graph may also include other nodes, such as rule nodes for representing rules and business nodes for representing business operations. Accordingly, the edges between risk nodes and rule nodes can be used to represent the relationship between risks and rules, specifically, for example, the risk represented by a risk node violates the rule represented by a rule node. The edges between risk nodes and business nodes can be used to represent the relationship between risks and business operations, specifically, the risk represented by a risk node belongs to the business represented by a business node. It is understood that different risk nodes connected to the same business node may have a certain relationship, such as belonging to the same business; different risk nodes connected to the same rule node may have a certain relationship, such as violating the same rule.

[0143] Therefore, for different risks, different risk nodes representing different risks can be identified in the risk knowledge graph. When there are connecting paths between different risk nodes, the degree of risk correlation between different risks can be determined based on these paths. The path can include other risk nodes, other types of nodes, or only a single edge. Combining the information from the path helps determine the degree of risk correlation between different risks, and integrating this information with the risk knowledge graph can improve the accuracy of the risk correlation determination.

[0144] Optionally, based on the identified target general risks and target specific risks, the target potential risks of the target electronic document are identified. Specifically, this may include: determining a target risk knowledge graph; the target risk knowledge graph contains risk nodes representing document risks, and the edges connecting different risk nodes in the target risk knowledge graph represent the association relationships between the connected risk nodes; for any risk node in the target risk knowledge graph, based on the path between the targeted risk node and the risk node representing the identified target general risks, determining the degree of risk association between the risk represented by the targeted risk node and the identified target general risks, and based on the path between the targeted risk node and the risk node representing the identified target specific risks, determining the degree of risk association between the risk represented by the targeted risk node and the identified target specific risks; the risks represented by risk nodes whose determined degree of risk association satisfies preset potential conditions are identified as target potential risks of the target electronic document. This embodiment can combine a risk knowledge graph to determine the degree of risk association, which can improve the accuracy of the degree of risk association and, correspondingly, improve the accuracy of potential risk identification.

[0145] Among these measures, considering that risk knowledge graphs can be updated frequently, it is possible to determine the current target risk knowledge graph and improve the real-time performance of the risk knowledge graph.

[0146] The embodiments of this application do not limit the specific method of determining the degree of risk association based on the path between risk nodes. Optionally, the degree of association between risks can be determined based on the path between risk nodes, and the degree of risk association can be determined based on the degree of association between risks. The degree of association between risks can be positively correlated with the degree of risk association; the number of edges contained in the path between risk nodes can be negatively correlated with the degree of risk association.

[0147] In a specific example, for two directly connected risk nodes, the degree of risk association can be determined based on the closeness of the association represented by the direct connection edge; for two risk nodes connected by a path composed of multiple edges, the degree of risk association can be determined by dividing the average or minimum value of the closeness of the association represented by each edge by the number of edges.

[0148] Alternatively, potential risks can be identified by filtering among the directly connected neighboring risk nodes of risk nodes representing the identified target general risks, and among the directly connected neighboring risk nodes of risk nodes representing the identified target specific risks. Optionally, for any specified risk node in the target risk knowledge graph, the degree of risk association between the risk represented by the specified risk node and the identified target general risks can be determined, as well as the degree of risk association between the risk represented by the specified risk node and the identified target specific risks. The specified risk node can be a neighboring risk node directly connected to a general risk node or a neighboring risk node directly connected to a specific risk node. A general risk node can be a risk node representing any identified target general risk. A specific risk node can be a risk node representing any identified target specific risk. It is understood that the degree of risk association can be determined based on the closeness of the association represented by the direct edges; alternatively, it can be determined by combining information such as the probability, frequency, or frequency of different risks appearing in the same document.

[0149] Furthermore, considering specific risks, the risk knowledge graph can include risk nodes to represent specific risks. For any given document type, specific risks for other document types are generally not used during document risk identification. Therefore, different risk knowledge graphs can be constructed for different document types. Each risk knowledge graph can contain risk nodes to represent specific risks for the corresponding document type, as well as risk nodes to represent general risks, but not risk nodes to represent specific risks for other document types.

[0150] Therefore, optionally, the N preset document types can each correspond one-to-one with N different risk knowledge graphs; any risk knowledge graph can include risk nodes representing document-specific risks of the corresponding preset document type, and risk nodes representing general document risks; determining the target risk knowledge graph can specifically include: determining the risk knowledge graph corresponding to the target document type as the target risk knowledge graph. This embodiment can deploy different risk knowledge graphs corresponding to different preset document types, and based on the risk knowledge graph corresponding to the target document type, determine the degree of risk correlation to identify potential risks, which can reduce the impact of other document types besides the target document type and improve the efficiency of identifying potential risks.

[0151] Different preset document types can correspond to different risk knowledge graphs. However, any risk knowledge graph may not include risk nodes that represent document-specific risks other than those corresponding preset document types.

[0152] Of course, the target risk knowledge graph can also contain risk nodes that represent specific risks for different preset document types.

[0153] The embodiments of this application do not limit the specific method of determining the risk identification result of the target electronic document based on the identified target general risks, target specific risks, and target potential risks. Optionally, the identified target general risks, target specific risks, and target potential risks can be determined as the risk identification result of the target electronic document; it can also be that the target potential risks are verified to see if there are any target potential risks in the target electronic document, and the target potential risks found to exist in the target electronic document are determined as the target actual risks, and the identified target general risks, target specific risks, and target actual risks can be determined as the comprehensive risk identification result of the target electronic document; alternatively, a risk assessment report can be generated based on the identified target general risks, target specific risks, and target potential risks as the risk identification result of the target electronic document. For specific explanations, please refer to other embodiments. Different embodiments can be combined with each other, and the embodiments of this application do not limit the specific combination method.

[0154] The embodiments of this application do not limit the method of verifying whether there is a potential risk in the target electronic document. Specifically, it can be done by using other large models to verify whether there is a potential risk in the target electronic document.

[0155] After obtaining the comprehensive risk identification results of the target electronic document, the embodiments of this application do not limit subsequent operations. Optionally, a risk assessment report or remediation suggestions can be generated based on the comprehensive risk identification results and fed back to the user; alternatively, it can be determined whether a prompt is needed based on the severity of the risk. If the severity of the risk is determined to be greater than a preset severity threshold based on the comprehensive risk identification results, a risk prompt message can be generated; if the severity of the risk is determined to be less than or equal to the preset severity threshold, no prompt may be given or automatic remediation may be performed, and the remediation results may be displayed to the user.

[0156] For ease of understanding, this application also provides an application embodiment.

[0157] Spatial topological coding can map the coordinates and structural relationships of text, seals, tables, signatures, etc. in a document into vectors, thereby achieving unified modeling of physical layout and semantic information.

[0158] Risk feature decoupling: Separate general document features from specific risk features to improve the generalization ability of risk identification across document types and enhance the ability to identify various risk features.

[0159] Strategy distillation: Compress new rules into a lightweight discriminator to enable incremental model updates and continuously improve risk assessment and identification capabilities.

[0160] Traditional manual document review methods face core pain points such as efficiency bottlenecks, error risks, and collaborative lags, making them unsuitable for enterprises' needs for handling massive amounts of documents and real-time compliance management. However, critical business documents (such as contracts, agreements, invoices, reports, and statements) are the core carriers of compliant business operations across various industries, and their review efficiency and risk identification accuracy directly impact compliance management and customer experience. Traditional document review relies on manual page-by-page review, manual verification of clauses and data, and manual verification of the authenticity of seals and signatures. This manual review model can no longer meet the business needs of high-frequency, high-volume, and rigorous review.

[0161] This embodiment proposes an automatic risk point identification system for business documents based on deep learning algorithms and computer vision technology. Through multimodal parsing, intelligent risk matching, and dynamic compliance verification, it realizes automated risk control of the entire financial document process, solving the pain points of low efficiency, high risk, and lagging compliance of manual review.

[0162] The "Intelligent Financial Document" risk point automatic identification system proposed in this embodiment, based on deep learning algorithms and computer vision technology, achieves automatic identification, accurate positioning and dynamic compliance verification of financial document risk points by integrating large-scale model multimodal fusion, graph neural network knowledge graph reasoning and strategy distillation incremental update technology. It innovatively introduces spatial topology coding, risk feature decoupling and cross-document risk transmission analysis, providing an integrated intelligent risk control solution for document review in various scenarios.

[0163] This solution effectively addresses the pain points of traditional manual review, including low efficiency, high false negative rates, high costs, lagging compliance, and inconsistent standards. Compared to manual review, it offers more accurate identification, higher efficiency, more real-time compliance, and greater generalization capabilities. Through intelligent processing of financial documents, this system can significantly improve risk control capabilities and business processing efficiency, comprehensively enhance compliance levels and customer experience, and create significant business value and security benefits.

[0164] This system is an "intelligent financial document" risk point automatic identification system based on large models, deep learning algorithms and computer vision technology. Its core features are a multimodal hierarchical parsing architecture, domain-adaptive risk identification, dynamic knowledge graph reasoning and lightweight incremental updates. It covers various types of financial documents such as contracts, reports, bills and documents, and realizes automatic risk point identification, location, rating and report output.

[0165] The system comprises four core modules: 1. Multi-type document collection and data preprocessing module; 2. Multi-type document parsing module based on deep learning algorithms; 3. Domain-adaptive risk identification and knowledge graph reasoning module; 4. Dynamic compliance and lightweight incremental update module.

[0166] 1. Multi-type document collection and data preprocessing module.

[0167] By using image recognition technology and electronic document parsing technology, raw data from various documents is collected and processed into standardized computer data, providing high-quality data input for subsequent document parsing.

[0168] For electronic documents in image format (scanned copies or images), image recognition can be used to identify the text within them, such as clause text, numbers, dates, entity names, seals, signatures, and other data.

[0169] For electronic documents in file format, the data can be obtained through parsing and used for semantic risk identification.

[0170] Preprocessing steps: Image enhancement: noise reduction, sharpening, contrast adjustment, and resolution of blur, shadows, and tilt issues. Layout standardization: uniform size and resolution, splitting documents into multiple pages, and locating areas for seals, signatures, and tables. Text cleaning: content correction, removal of redundant symbols, and standardization of formatting. Tagging and construction: tagging based on risk type (term risk, data risk, seal risk, compliance risk).

[0171] 2. A multi-type document parsing module based on deep learning algorithms.

[0172] By employing a variety of model architectures, including multimodal large models, we can achieve in-depth analysis of the physical structure and semantic information of documents, laying a data foundation for subsequent risk identification.

[0173] Step 1: Multimodal Feature Extraction. This step extracts joint features from document layout, images, and text to achieve spatial semantic alignment. Through a multimodal fusion module, text, layout, and image features are jointly modeled, and an attention gating mechanism is introduced to dynamically weight visual and semantic features.

[0174] Step 2: Spatial Topological Encoding. The coordinates, hierarchical relationships, and adjacency relationships of text, seals, tables, and signatures are encoded into vectors to establish topological associations and achieve a two-dimensional understanding of "location + semantics".

[0175] Among these, the document type can be determined based on the extracted document features.

[0176] Step 3: Key Element Extraction. Automatically extract core elements. For example, in a contract, this includes the party A, party B, amount, interest rate, term, default clauses, and disclaimers; in a negotiable instrument, it includes the drawer, payee, amount, date, seal, and endorsement information.

[0177] 3. Domain-Adaptive Risk Identification and Knowledge Graph Reasoning Module.

[0178] It can identify general risk features of documents and specific risk features specific to document types. The separation process employs adversarial training between general feature branches and specific feature branches, which effectively separates common features from specific features.

[0179] Common risks include typos, date conflicts, and missing key data. Specific risks include conflicting terms, missing signatures, inconsistent amounts in words, missing seals, and missing key elements on the invoice.

[0180] (1) General feature branch (shared layer): In order to realize general document feature recognition, the document general features (such as text content and signature position) are extracted through the shared encoder to identify common risks (data missing, format error, basic logical contradiction).

[0181] Data Input: Input document data, such as text content, layout position (signature coordinates), and visual styles (e.g., font, stamp color). Multimodal Feature Concatenation: Text embedding + layout coordinates + visual features.

[0182] Data processing: The document data is analyzed step-by-step using an encoder. For example: Level 1: Check basic syntax. Level 2: Check signature locations. Level 12: Comprehensive risk assessment.

[0183] Output: A general risk feature vector for the document, such as "date conflict" or "missing data".

[0184] (2) Adversarial training mechanism: A gradient inversion layer is introduced and placed at the end of the general feature branch to force the features extracted by the general feature branch to be unrelated to the document type, ensuring that the exclusive feature branch focuses on the exclusive type risk.

[0185] The gradient inversion layer can be located at the end of the general feature branch, performing an identity mapping during forward propagation and inverting the gradient during backward propagation.

[0186] Adversarial Discriminator Design: Input is general features, output is document type probability. During training, the general branch parameters are frozen, and only the discriminator weights are updated. Joint Training Flow: Stage 1: Train the general branch and the adversarial branch normally to minimize risk identification loss. Stage 2: Fix the general branch and train the adversarial discriminator (maximizing type classification accuracy). Stage 3: Enable the gradient reversal layer to train the general branch to "deceive" the discriminator (making it unable to distinguish document types).

[0187] (3) Specific Feature Branch (Dynamic Adaptation Layer): To achieve specific risk feature identification, a lightweight adapter based on document type is used to learn specific features for each scenario (such as clauses in contracts and the continuity of endorsement of bills) on the basis of general features. Specific features can be extracted by using the corresponding specific feature branch according to different document types.

[0188] For example, for each of the N preset document types, N corresponding exclusive feature branches can be deployed to train the model accordingly.

[0189] Specifically, for contract documents, clause features can be enhanced for subsequent risk identification based on corresponding dedicated feature branches. Regional semantic embeddings (such as "signature area" and "stamp area" labels) can also be integrated to improve the spatial sensitivity of dedicated features by encoding coordinate relationships. Furthermore, cross-modal attention gating can be implemented: a cross-attention layer of visual and textual features can be added within the dedicated feature branches, dynamically adjusting the weights between visual and textual features according to different document types.

[0190] Based on the general features extracted from the aforementioned general feature branches, general risks can be identified; based on the specific features extracted from the aforementioned specific feature branches, specific risks can be identified. General and specific risks can be labeled according to risk level (high / medium / low), the risk location (page number, paragraph, area) can be located, and risk descriptions and modification suggestions can be output.

[0191] 4. Dynamic compliance and lightweight incremental update module.

[0192] Step 1: Dynamic rule capture. The system automatically captures the latest rules and updates the risk rule database in real time, ensuring that review standards are synchronized in real time.

[0193] Step 2: Policy distillation for incremental updates. When updating rules, policy distillation is used to compress the new rules into a lightweight discriminator, which is then quickly embedded into the model.

[0194] Distillation-style rule embedding process: Rule compression: Transform new rules into structured rule triples (trigger condition, risk type, threshold). Discriminator generation: Train a lightweight logical discriminator (<100 parameters) based on the rule triples, decoupling it from the original model. Dynamic embedding: Connect the discriminator to the corresponding risk branch (such as a compliance risk branch) through a gating mechanism, activating it only when a new rule is triggered, avoiding global retraining.

[0195] Step 3: Output and closed-loop feedback.

[0196] Output: Risk identification report (including risk points, locations, levels, and recommendations), element extraction list, and compliance verification results. Feedback: The verification results are fed back to the model for incremental training to continuously improve accuracy.

[0197] Within the general risk layer, lightweight modules such as data integrity discriminators and basic logic verifiers can be deployed in parallel. In the dedicated risk layer, corresponding discriminator groups can be dynamically loaded based on document type.

[0198] Strategy distillation linkage: Automatically generate discriminators for new rules and insert them into the corresponding levels. For example, when adding a new general risk, a new discriminator generated for the new general risk can be inserted into the general risk layer.

[0199] In addition, a potential risk mining mechanism can be set up. A risk knowledge graph can be constructed, and potential risk transmission paths can be mined through graph neural networks to achieve cross-document joint prevention and control.

[0200] Constructing a risk association graph: Node types: entities, risks, events, rules, etc. in documents. Association edge definitions: causal relationships, rule constraints, etc.

[0201] Graph Neural Network Inference: Graph neural networks are used to calculate the correlation between risk nodes, locate the risk transmission path across documents, and also determine the potential risk transmission path within the same document.

[0202] This embodiment can use the document risk identification results for document review, thereby improving document review efficiency, increasing the accuracy of document risk identification, reducing the document missed detection rate, lowering document review costs, facilitating rapid model iteration, and improving user experience.

[0203] Based on the above method embodiments, embodiments of this application also provide a document risk identification device. The following will be combined with... Figure 3 The device is described in detail. Figure 3 The diagram illustrates a structural block diagram of a document risk identification device according to an embodiment of this application.

[0204] like Figure 3 As shown, the document risk identification device 300 provided in this embodiment includes: a type determination module 310, a specific risk module 320, a general risk module 330, and a result determination module 340.

[0205] The type determination module 310 is used to determine the target document type of the target electronic document from N preset document types; each of the N preset document types corresponds to one of the N exclusive risk identification models; each exclusive risk identification model is used to identify the exclusive risk in the document of the corresponding preset document type; N is a positive integer. In one embodiment, the type determination module 310 can be used to execute the steps S210 and related steps described above, which will not be repeated here.

[0206] The dedicated risk module 320 is used to identify the target dedicated risk of the target electronic document based on the dedicated risk identification model corresponding to the target document type and according to the target document characteristics. In one embodiment, the dedicated risk module 320 can be used to perform the steps S220 and related steps described above, which will not be repeated here.

[0207] The general risk module 330 is used to identify target general risks of the target electronic document based on the target document characteristics and a general risk identification model. In one embodiment, the general risk module 330 can be used to perform the steps S230 and related steps described above, which will not be repeated here.

[0208] The result determination module 340 is used to determine the risk identification result of the target electronic document based on the identified target general risks and target specific risks. In one embodiment, the result determination module 340 can be used to perform the steps S240 and related steps described above, which will not be repeated here.

[0209] According to embodiments of this application, the target document features include at least one of the following: document content features, document layout features, and document fusion features; the method for determining the document fusion features includes: for the target electronic document, determining a document layout region; for any document layout region, determining the layout semantic features of the targeted document layout region, and determining the document content features corresponding to the document content located in the targeted document layout region, fusing the determined layout semantic features and the determined document content features to obtain the document fusion features of the targeted document layout region.

[0210] According to an embodiment of this application, the document content features include document content semantic features; the method for determining the document content semantic features includes: for any document content, extracting document content semantic features based on the document content and the context information of the document content.

[0211] According to an embodiment of this application, the document content features include document content editing features; the method for determining the document content editing features includes: for any document content, extracting document content editing features based on the document content and the editing operation information of the document content.

[0212] According to an embodiment of this application, the result determination module 340 can also be used to: identify target potential risks of the target electronic document based on the identified target general risks and target specific risks; the target potential risks meet preset potential conditions; the preset potential conditions include at least one of the following: the degree of risk correlation between the target potential risks and the identified target general risks is greater than a first correlation degree threshold, and the degree of risk correlation between the target potential risks and the identified target specific risks is greater than a second correlation degree threshold; and determine the risk identification result of the target electronic document based on the identified target general risks, target specific risks, and target potential risks.

[0213] According to an embodiment of this application, the result determination module 340 is specifically used for: determining a target risk knowledge graph; the target risk knowledge graph contains risk nodes for characterizing document risks, and the edges connecting different risk nodes in the target risk knowledge graph are used to characterize the association relationship between the connected risk nodes; for any risk node in the target risk knowledge graph, based on the path between the targeted risk node and the risk node for characterizing the identified target general risk, determining the degree of risk association between the risk represented by the targeted risk node and the identified target general risk, and based on the path between the targeted risk node and the risk node for characterizing the identified target specific risk, determining the degree of risk association between the risk represented by the targeted risk node and the identified target specific risk; and determining the risk represented by the risk node whose determined degree of risk association satisfies the preset potential conditions as the target potential risk of the target electronic document.

[0214] According to an embodiment of this application, N preset document types correspond one-to-one with N different risk knowledge graphs; each risk knowledge graph includes risk nodes for representing document-specific risks of the corresponding preset document type, and risk nodes for representing general document risks; the result determination module 340 is specifically used to: determine the risk knowledge graph corresponding to the target document type as the target risk knowledge graph.

[0215] According to embodiments of this application, any specific risk identification model includes: P specific risk discrimination sub-models; different specific risk discrimination sub-models are used to identify different specific risks in documents of corresponding preset document types; P is a positive integer; the general risk identification model includes: Q general risk discrimination sub-models; different general risk discrimination sub-models are used to identify different general risks in documents; Q is a positive integer; the above device may further include: a adding module, used to: when a new specific risk is added for any preset document type, determine a new specific risk discrimination sub-model for identifying the new specific risk, and add the determined new specific risk discrimination sub-model to the specific risk identification model corresponding to the preset document type; when a new general risk is added, determine a new general risk discrimination sub-model for identifying the new general risk, and add the determined new general risk discrimination sub-model to the general risk identification model.

[0216] According to embodiments of this application, any multiple modules among the type determination module 310, dedicated risk module 320, general risk module 330, result determination module 340, and new module can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the type determination module 310, dedicated risk module 320, general risk module 330, result determination module 340, and new module can be at least partially implemented as hardware circuitry, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the type determination module 310, the specific risk module 320, the general risk module 330, the result determination module 340, and the new module can be implemented at least partially as a computer program module that can perform corresponding functions when the computer program module is run.

[0217] The explanation of the above device embodiments can be found in the explanation of other embodiments, and any step in the above method embodiments can be executed by the corresponding module.

[0218] Figure 4 A block diagram schematically illustrates an electronic device suitable for implementing a document risk identification method according to an embodiment of this application. For example... Figure 4 As shown, an electronic device 1200 according to an embodiment of this application includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory 1202 or a program loaded from a storage portion 1208 into a random access memory 1203. The processor 1201 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 1201 may also include onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0219] Random access memory 1203 stores various programs and data required for the operation of electronic device 1200. Processor 1201, read-only memory 1202, and random access memory 1203 are interconnected via bus 1204. Processor 1201 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 1202 and / or random access memory 1203. It should be noted that the programs may also be stored in one or more memories other than read-only memory 1202 and random access memory 1203. Processor 1201 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0220] According to embodiments of this application, the electronic device 1200 may further include an input / output interface 1205, which is also connected to the bus 1204. The electronic device 1200 may also include one or more of the following components connected to the input / output interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card, such as a local area network card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the input / output interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1210 as needed so that computer programs read from it can be installed into the storage section 1208 as needed.

[0221] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0222] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 1202, and / or random access memory 1203, and / or one or more memories other than read-only memory 1202 and random access memory 1203 described above.

[0223] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0224] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1209, and / or installed from the removable medium 1211. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0225] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by processor 1201, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0226] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0227] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0228] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A document risk identification method, characterized in that, The method includes: Among N preset document types, the target document type of the target electronic document is determined; each of the N preset document types corresponds to one of the N exclusive risk identification models; any exclusive risk identification model is used to identify the exclusive risk in the document of the corresponding preset document type; N is a positive integer. Based on the specific risk identification model corresponding to the target document type, the target specific risk of the target electronic document is identified according to the target document characteristics; Based on a general risk identification model, the target general risks of the target electronic document are identified according to the target document characteristics. Based on the identified general and specific risks to the target, the risk identification result of the target electronic document is determined.

2. The method according to claim 1, characterized in that, The target document features include at least one of the following: document content features, document layout features, and document fusion features; The method for determining the document fusion features includes: for the target electronic document, determining a document layout area; for any document layout area, determining the layout semantic features of the target document layout area, and determining the document content features corresponding to the document content located in the target document layout area; fusing the determined layout semantic features and the determined document content features to obtain the document fusion features of the target document layout area.

3. The method according to claim 2, characterized in that, The document content features include document content semantic features; The method for determining the semantic features of the document content includes: for any document content, extracting the semantic features of the document content based on the document content and the context information of the document content.

4. The method according to claim 2, characterized in that, The document content features include document content editing features; The method for determining the document content editing features includes: for any document content, extracting document content editing features based on the document content and the editing operation information of the document content.

5. The method according to claim 1, characterized in that, The method further includes: identifying target potential risks of the target electronic document based on the identified target general risks and target specific risks; the target potential risks satisfy preset potential conditions; the preset potential conditions include at least one of the following: the risk correlation between the target potential risks and the identified target general risks is greater than a first correlation threshold, and the risk correlation between the target potential risks and the identified target specific risks is greater than a second correlation threshold; The step of determining the risk identification result of the target electronic document based on the identified target general risks and target specific risks includes: Based on the identified general risks, specific risks, and potential risks of the target, the risk identification results of the target electronic document are determined.

6. The method according to claim 5, characterized in that, The process of identifying potential risks of the target electronic document based on the identified target-specific and target-general risks includes: A target risk knowledge graph is determined; the target risk knowledge graph contains risk nodes for representing document risks, and the edges connecting different risk nodes in the target risk knowledge graph are used to represent the association between the connected risk nodes; For any risk node in the target risk knowledge graph, the degree of risk correlation between the risk represented by the target risk node and the identified target general risk is determined based on the path between the target risk node and the risk node used to represent the identified target specific risk. The degree of risk correlation between the risk represented by the target risk node and the identified target specific risk is also determined based on the path between the target risk node and the risk node used to represent the identified target specific risk. The risk represented by the risk node whose risk correlation degree meets the preset potential conditions is identified as the target potential risk of the target electronic document.

7. The method according to claim 6, characterized in that, The N preset document types each correspond to one of the N different risk knowledge graphs; each risk knowledge graph includes risk nodes that represent document-specific risks of the corresponding preset document type, and risk nodes that represent general document risks. The determination of the target risk knowledge graph includes: The risk knowledge graph corresponding to the target document type is identified as the target risk knowledge graph.

8. The method according to claim 1, characterized in that, Any specific risk identification model includes: P specific risk discrimination sub-models; different specific risk discrimination sub-models are used to identify different specific risks in documents of corresponding preset document types; where P is a positive integer; The general risk identification model includes: Q general risk discrimination sub-models; different general risk discrimination sub-models are used to identify different general risks of a document; where Q is a positive integer; The method further includes: When a new specific risk is added for any preset document type, a new specific risk discrimination sub-model is determined to identify the new specific risk, and the determined new specific risk discrimination sub-model is added to the specific risk identification model corresponding to the preset document type. In the case of a new general risk, a new general risk discrimination sub-model is determined to identify the new general risk, and the determined new general risk discrimination sub-model is added to the general risk identification model.

9. A document risk identification device, characterized in that, The device includes: The type determination module is used to determine the target document type of the target electronic document from N preset document types; each of the N preset document types corresponds to one of the N exclusive risk identification models; any exclusive risk identification model is used to identify the exclusive risk in the document of the corresponding preset document type; N is a positive integer; The dedicated risk module is used to identify the target dedicated risk of the target electronic document based on the dedicated risk identification model corresponding to the target document type and according to the target document characteristics of the target electronic document; A general risk module is used to identify the target general risks of the target electronic document based on the target document characteristics and a general risk identification model. The result determination module is used to determine the risk identification result of the target electronic document based on the identified target general risks and target specific risks.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.

12. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 8.