Contract auditing method and device, equipment, storage medium and program product

Abstract

The invention relates to a contract auditing method and device, equipment, a storage medium and a program product. Comprising the steps of obtaining text data of a to-be-audited contract; identifying risk-free terms in the text data by adopting a lightweight classification model, and screening to obtain a plurality of suspected risk terms; performing deep semantic analysis on the plurality of suspected risk terms in sequence by using a preset large model, and determining corresponding key elements and a first risk auditing result; constructing a logic association graph through a graph neural network model; and determining a second risk auditing result between terms according to the logical association graph. According to the method, a hierarchical cascade type large model architecture is adopted, preliminary screening is completed through a lightweight classification model, and the overall calculation cost and response time are remarkably reduced. And the large model centralizes calculation power to carry out deep reasoning on high-risk terms. And revealing association risks and potential conflicts among terms through a logic association graph to realize global logic verification. The auditing efficiency and the auditing accuracy can be balanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a contract auditing method and device, computer equipment, computer readable storage medium and computer program product. BACKGROUND

[0002] In the current field of intelligent contract auditing, in addition to traditional manual auditing, rule engine, and statistical model methods, there is also a solution relying on large models. This solution directly applies a general large language model to the contract auditing task, and outputs an auditing conclusion after inputting the contract text. The model relies on its general semantic understanding ability to analyze the content of the clauses, identify potential risks and compliance issues.

[0003] However, in this way, the large model needs to process a large amount of clause content, and the auditing efficiency is low, which is difficult to meet the timeliness requirements of large-scale contract auditing. The large model tends to give a general risk assessment, lacks a clause-by-clause analysis of key clauses, and the correlation analysis capability of the large model is limited, which cannot effectively distinguish the associated risks and potential conflicts scattered among different clauses. SUMMARY

[0004] Therefore, it is necessary to provide a contract auditing method, device, computer equipment, computer readable storage medium and computer program product to balance the auditing efficiency and accuracy by rapid preliminary screening, deep semantic analysis and logical correlation graph analysis, so as to meet the timeliness requirements of large-scale contract auditing.

[0005] In a first aspect, the present application provides a contract auditing method, comprising:

[0006] obtaining text data of a contract to be audited;

[0007] identifying risk-free clauses in the text data using a lightweight classification model, and screening a plurality of suspected risk clauses according to the identification result;

[0008] performing deep semantic analysis on the plurality of suspected risk clauses in sequence using a preset large model, to determine key elements and a first risk auditing result corresponding to each suspected risk clause;

[0009] constructing a logical correlation graph according to the key elements corresponding to the plurality of suspected risk clauses through a graph neural network model;

[0010] determining a second risk auditing result between clauses according to the logical correlation graph;

[0011] obtaining an auditing result of the contract to be audited according to the first risk auditing result and the second risk auditing result.

[0012] In one of the embodiments, the preset large model is used to sequentially perform deep semantic analysis on the plurality of suspected risk clauses to determine the key elements and the first risk audit result corresponding to each suspected risk clause, including:

[0013] Retrieving audit points corresponding to the contract to be audited from a preset knowledge graph;

[0014] Generating prompt words for prompting the preset large model according to the audit points;

[0015] Under the indication of the prompt words, the preset large model is used to sequentially perform deep semantic analysis on the plurality of suspected risk clauses to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0016] In one of the embodiments, the audit points include historical high-frequency risk points and policy-related points, and the method further includes:

[0017] Generating historical high-frequency risk points in each business scenario according to historical contract audit records to obtain first graph data;

[0018] Generating policy-related points of each contract clause according to policy and regulation texts to obtain second graph data;

[0019] According to the first graph data and the second graph data, a preset knowledge graph is constructed.

[0020] In one of the embodiments, after obtaining the audit result of the contract to be audited according to the first risk audit result and the second risk audit result, the method further includes:

[0021] According to the risk type corresponding to the audit result, retrieving the corresponding negotiation basis from a pre-constructed strategy knowledge base.

[0022] In one of the embodiments, after obtaining the text data of the contract to be audited, the method further includes:

[0023] Evaluating the risk score of the text data;

[0024] In the case where the risk score is less than a set threshold, performing the step of identifying risk-free clauses in the text data using a lightweight classification model, and screening to obtain the plurality of suspected risk clauses according to the identification result;

[0025] In the case where the risk score is greater than or equal to the set threshold, determining all clauses in the text data as suspected risk clauses, and performing the step of using the preset large model to sequentially perform deep semantic analysis on the plurality of suspected risk clauses to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0026] In one of the embodiments, the method further includes:

[0027] According to the contract-related corpus, the large model is unsupervised pre-trained to obtain a domain basic model;

[0028] The domain basic model is subjected to knowledge distillation processing to obtain a preset large model.

[0029] In a second aspect, the present application further provides a contract auditing device, comprising:

[0030] An acquisition module is configured to acquire text data of a contract to be audited;

[0031] A screening module is configured to identify risk-free clauses in the text data by using a lightweight classification model, and to screen a plurality of suspected risk clauses according to the identification result;

[0032] A first auditing module is configured to perform deep semantic analysis on the plurality of suspected risk clauses in sequence by using a preset large model, to determine key elements and a first risk auditing result corresponding to each suspected risk clause;

[0033] A second auditing module is configured to construct a logical association graph according to the key elements corresponding to the plurality of suspected risk clauses by using a graph neural network model, to determine a second risk auditing result between clauses according to the logical association graph, and to obtain an auditing result of the contract to be audited according to the first risk auditing result and the second risk auditing result.

[0034] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the method of the first aspect when executing the computer program.

[0035] In a fourth aspect, the present application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the steps of the method of the first aspect.

[0036] In a fifth aspect, the present application further provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the steps of the method of the first aspect.

[0037] The contract auditing method, device, computer equipment, computer readable storage medium and computer program product obtain text data of a contract to be audited. A lightweight classification model is used to identify non-risk clauses in the text data, and a plurality of suspected risk clauses are obtained by screening according to the identification result. A preset large model is used to sequentially perform deep semantic analysis on the plurality of suspected risk clauses, to determine a key element and a first risk auditing result corresponding to each suspected risk clause. A graph neural network model is used to construct a logical association graph according to the key elements corresponding to the plurality of suspected risk clauses. A second risk auditing result between clauses is determined according to the logical association graph. The first risk auditing result and the second risk auditing result are used to obtain an auditing result of the contract to be audited. In this way, a hierarchical cascading large model architecture is used, and the lightweight classification model is used to complete preliminary screening, thereby avoiding the need to send a large number of non-risk clauses to the large model for redundant calculation, significantly reducing the overall calculation cost and response time. The large model can concentrate computing power to perform deep reasoning on high-risk clauses, thereby improving the identification accuracy and analysis depth of core risks. The logical association graph is further used to reveal the associated risks and potential conflicts between clauses, thereby achieving global logical verification. Through rapid preliminary screening, deep semantic analysis and logical association graph analysis, the auditing efficiency and accuracy can be balanced to meet the timeliness requirements of large-scale contract auditing. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the description of the embodiments of the present application or the related art will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other related drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 An application environment diagram of a contract auditing method in an embodiment;

[0040] Figure 2 A flowchart of a contract auditing method in an embodiment;

[0041] Figure 3 A structural block diagram of a contract auditing device in an embodiment;

[0042] Figure 4 An internal structure diagram of a computer equipment in an embodiment. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0044] It should be noted that the terms "first", "second", etc. used in the present application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "include" and "have" and any variations thereof used in the present application are intended to cover non-exclusive inclusion. The term "multiple" used in the present application refers to two or more. The term "and / or" used in the present application refers to one of the options or any combination of multiple options.

[0045] The contract auditing method provided by the embodiments of the present application can be applied to the application environment as shown in Figure 1 The terminal 102 communicates with the server 104 through the network. The data storage system can store the data required to be processed by the server 104. The data storage system can be integrated on the server 104, or placed on the cloud or other network servers. The server 104 is deployed with a hierarchical cascading large model architecture, i.e. a lightweight classification model, a preset large model and a graph neural network model. The terminal 102 sends a contract to be audited to the server, and the server 104 audits the contract to be audited uploaded by the terminal 102 based on the contract auditing method provided by the embodiments of the present application, and returns the audit result to the terminal 102. The terminal 102 can be, but is not limited to, various personal computers, notebook computers, smart phones, tablet computers, unmanned aerial vehicles, low-altitude flying vehicles, Internet of Things devices and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle device, a projection device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The head-mounted device can be a virtual reality (VR) device, an augmented reality (AR) device, smart glasses, etc. The server 104 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0046] In an exemplary embodiment, as shown in Figure 2 A contract auditing method is provided, which is applied to the server 104 in Figure 1 for example, including:

[0047] Step 202, obtaining text data of a contract to be audited.

[0048] The text data is read from the contract to be audited. For scanned copies or picture type contracts, an OCR (Optical Character Recognition) tool is used to identify the contract to be audited to obtain the text data.

[0049] In step 204, a lightweight classification model is used to identify the risk-free clauses in the text data, and the suspected risk clauses are obtained by screening according to the identification result.

[0050] The lightweight classification model is a relatively simple model for dividing data into preset categories, which consumes less computing resources, runs faster, and is easy to deploy. The lightweight classification model is used to analyze the input text data, predict the category label (e.g., risk-free clause or non-risk-free clause) of each clause, and obtain the identification result. All clauses that are not identified as risk-free clauses by the lightweight classification model are retained as suspected risk clauses, and the marked risk-free clauses are filtered to achieve preliminary efficient screening.

[0051] In step 206, a preset large model is used to perform deep semantic analysis on the suspected risk clauses one by one to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0052] The preset large model refers to a deep learning model that is prepared in advance, configured or fine-tuned. In specific implementation, the preset large model is called to perform deep semantic analysis on each suspected risk clause, such as precise reasoning on the details of the liability definition and compensation calculation logic of the default clause, extracting structured core information points, i.e., key elements, from each suspected risk clause, and determining the risk situation of each suspected risk clause to obtain the first risk audit result.

[0053] The key elements refer to the core information points extracted from the risk clauses, which constitute the basis for understanding and evaluating the risk of the clauses, and can be structured data such as responsible party, trigger condition, and amount.

[0054] For example, a prompt template for prompting the preset large model to perform key element screening and risk audit is pre-configured, the suspected risk clauses and the prompt template are combined to generate an input prompt, and the input prompt is input into the preset large model to instruct the preset large model to perform deep semantic analysis on the suspected risk clauses one by one to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0055] In step 208, a graph neural network model is used to construct a logical association graph based on the key elements corresponding to the suspected risk clauses.

[0056] The graph neural network model refers to a neural network for processing graph structure data, which can not only learn the features of nodes, but also learn the topological structure information of the entire network through the connection relationship between nodes. The key elements are input into the graph neural network model to determine the nodes in the graph and the edges between the nodes, thereby establishing the logical association graph between the clauses.

[0057] Step 210, determining the second risk audit result between clauses according to the logical association graph.

[0058] Wherein, the right and obligation conflicts across clauses are detected according to the logical association graph, such as the contradiction between the payment time and the delivery time, the inconsistent responsibility subject of different clauses, etc., to obtain the second risk audit result between clauses.

[0059] Step 212, obtaining the audit result of the contract to be audited according to the first risk audit result and the second risk audit result.

[0060] Wherein, the first risk audit result and the second risk audit result are integrated to obtain the audit result of the contract to be audited.

[0061] It can be understood that the embodiment of the application adopts a hierarchical cascading large model architecture: the first level is a lightweight classification model, the second level is a preset large model, and the third level is a graph neural network model. Both single-point audit and global logical verification are considered to improve the accuracy of contract audit.

[0062] In the above contract audit method, the text data of the contract to be audited is obtained; the lightweight classification model is used to identify the risk-free clauses in the text data, and the multiple suspected risk clauses are obtained by screening according to the identification result; the preset large model is used to perform deep semantic analysis on the multiple suspected risk clauses in sequence to determine the key elements and the first risk audit result corresponding to each suspected risk clause; the graph neural network model is used to construct a logical association graph according to the key elements corresponding to the multiple suspected risk clauses; the second risk audit result between clauses is determined according to the logical association graph; and the audit result of the contract to be audited is obtained according to the first risk audit result and the second risk audit result. Through the above-mentioned manner, a hierarchical cascading large model architecture is adopted, the lightweight classification model is used for preliminary screening, which avoids sending a large number of risk-free clauses into the large model for redundant calculation, significantly reduces the overall calculation cost and response time. The large model can concentrate computing power to perform deep reasoning on high-risk clauses, improving the identification accuracy and analysis depth of core risks. Further, the logical association graph is used to reveal the associated risks and potential conflicts between clauses, realizing global logical verification. Through rapid preliminary screening, deep semantic analysis and logical association graph analysis, the audit efficiency and accuracy can be balanced to meet the time efficiency requirements of large-scale contract audit.

[0063] In one exemplary embodiment, step 206 includes: retrieving the audit points corresponding to the contract to be audited from the preset knowledge graph; generating prompt words for prompting the preset large model according to the audit points; under the indication of the prompt words, using the preset large model to perform deep semantic analysis on the multiple suspected risk clauses in sequence to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0064] It can be understood that the traditional general large model has deviation in understanding the professional terms of the contract field, and the audit points are not fully identified. In view of this problem, the embodiments of the present application recommend audit points based on a preset knowledge graph, so that the large model can more accurately cover the key risks in the complex contract scenario.

[0065] The prompt template for prompting the preset large model to perform key element screening and risk audit according to the audit points is preconfigured. According to the industry attribute, subject type and other characteristics of the current contract to be audited, the corresponding audit points are retrieved from the preset knowledge graph, and the input prompt is generated by combining multiple suspected risk clauses, audit points and prompt templates. The input prompt is input into the preset large model to instruct the preset large model to sequentially perform deep semantic analysis on the multiple suspected risk clauses to determine the key elements and the first risk audit result corresponding to each suspected risk clause.

[0066] In one exemplary embodiment, the audit points include historical high-frequency risk points and policy-related points, and the method further includes: generating historical high-frequency risk points in each business scenario according to historical contract audit records to obtain first type of graph data; generating policy-related points of each contract clause according to policy and regulation texts to obtain second type of graph data; and constructing the preset knowledge graph according to the first type of graph data and the second type of graph data.

[0067] It can be understood that in the conventional technology, the fixed audit points are generally summarized by artificial, which cannot combine the dynamic changes of historical dispute data, and the influence of the latest policies and regulations on the contract clauses is not fused, resulting in that the audit points lag behind the business and compliance requirements, and the risk omission or excessive audit is easy to occur. Based on this, in the embodiments of the present application, the knowledge graph including the graph nodes of the historical dimension and the policy dimension is constructed, the high-frequency risk point data of the historical contract audit and the semantic analysis result of the latest policies and regulations are fused, the intelligent recommendation of the combined audit points of the historical experience and the new rule requirements for the contract audit task is realized, and the problems of lagging of the point recommendation and insufficient precision are solved.

[0068] For example, in the historical dimension, multi-dimensional analysis is performed on historical contract review records, including industry classification, contract type tags, and risk point distribution statistics, to extract historical high-frequency risk points under various business scenarios and form historical experience knowledge. In the policy dimension, semantic analysis technology of policy and regulations texts (including regulatory revision identification algorithms and effectiveness level judgment models) is used to analyze policy and regulatory texts, capturing in real time the impact of new regulations on contract terms, such as the compliance requirements of tax policy changes on payment terms and the adjustments to the definition of breach of contract liability by newly promulgated regulations, to obtain policy-related key points. When a new contract review task is received, based on the industry attributes, type of subject matter, and other characteristics of the contract, a combined review list of historical high-frequency risk points and policy-related key points is intelligently retrieved from a preset knowledge graph and pushed. At the same time, manual adjustment and feedback of recommended key points are supported to continuously optimize the accuracy of the recommendation model.

[0069] In an exemplary embodiment, after step 212, the method further includes: retrieving the corresponding negotiation basis from a pre-built strategy knowledge base according to the risk type corresponding to the audit result.

[0070] Understandably, traditional technology-based review processes only reach the risk identification stage, failing to integrate with contract negotiation and clause optimization. This necessitates manual conversion of review conclusions into negotiation strategies, resulting in fragmented processes and low efficiency in contract optimization. To address this issue, this application's embodiment utilizes a closed-loop linkage mechanism between contract review and intelligent negotiation, directly transforming risk points into negotiation grounds. This achieves seamless process integration from review to optimization, improving the overall efficiency of contract lifecycle processing.

[0071] For example, a risk-negotiation strategy mapping knowledge base is constructed, which integrates resources such as interpretations of legal provisions, industry practice data, and historical successful negotiation cases. When a contract risk is identified, the corresponding negotiation basis is automatically retrieved from the knowledge base, generating a suggested text that can be directly used in business negotiations, including legal support and strategic suggestions. This achieves a closed loop from risk identification to solution output, improving the efficiency of the entire process from contract review to optimization.

[0072] In an exemplary embodiment, after step 202, the method further includes: assessing the risk score of the text data; if the risk score is less than a set threshold, performing step 204; if the risk score is greater than or equal to the set threshold, identifying all terms in the text data as suspected risk terms, and performing step 206.

[0073] Understandably, traditional review processes lack dynamic priority scheduling, with high-risk, high-value contracts sharing resources with ordinary contracts, leading to delays in the review of core contracts. Furthermore, the general-purpose large-scale model uses indiscriminate reasoning for all contracts, resulting in significant waste of computational resources and an inability to meet the demands of batch contract processing. To address this issue, a dynamic priority scheduling mechanism for contract review tasks is implemented, dynamically allocating large-scale model computational resources based on contract attributes. Compared to the traditional method of indiscriminate resource allocation, this approach ensures timely review of high-priority contracts while avoiding resource waste for low-priority tasks, thus improving overall resource utilization efficiency.

[0074] For example, using a risk-resource intelligent matching algorithm, when a contract review request is received, the algorithm analyzes the text data of the contract to be reviewed based on dimensions such as contract type, amount, and risk level (e.g., semantic recognition of clause sensitivity, historical dispute data mining, etc.) to automatically assess the contract's risk score. Based on this risk score, a large model inference thread is scheduled, prioritizing the allocation of computing resources to high-risk, high-value contracts. When the risk score is greater than or equal to a set threshold, a deep inference is performed on all clauses of the contract using a pre-set large model, shortening the response time for high-priority contract reviews. When the risk score is less than the set threshold, a rapid pre-review is performed through a branch of a lightweight classification model, achieving dynamic optimization of review resource allocation and improving the resource utilization efficiency of contract reviews.

[0075] In an exemplary embodiment, the method further includes: performing unsupervised pre-training on a large model based on contract-related corpus to obtain a domain-based basic model; and performing knowledge distillation on the domain-based basic model to obtain a pre-defined large model.

[0076] Understandably, traditional general-purpose large-scale models lack sufficient understanding of industry-specific rules (such as credit rules in financial contracts and ownership rules in real estate contracts), limiting their accuracy in review. Unoptimized general-purpose large-scale model architectures suffer from insufficient reasoning ability and high resource consumption when handling complex contract logic (such as cross-clause conflicts of rights and obligations). To address this issue, this application's embodiments utilize unsupervised pre-training with contract-related corpora to enhance the model's understanding of legal concepts and clause logic. Knowledge distillation simplifies model parameters while preserving professional semantic understanding capabilities. Compared to general-purpose large-scale models, this improves review speed and reduces deployment costs. Furthermore, the hierarchical cascaded contract review large-scale model architecture balances initial screening efficiency with deep verification accuracy, resulting in higher review accuracy and more reasonable resource consumption compared to a single-model architecture.

[0077] For example, legal provisions and court judgments are used as contract-related corpora to perform unsupervised pre-training on the basic model within the domain, resulting in a domain-based basic model capable of understanding legal concepts and contractual logic. A specialized training task set is constructed, including tasks for predicting legal terminology masks and logical reasoning of contractual clauses. A general-purpose language model is used as the teacher model, and the domain-based basic model as the student model. Knowledge distillation is performed through the specialized training task set, compressing the legal knowledge from the teacher model into the student model. The lightweight student model, after knowledge distillation, is deployed in the contract review system to perform semantic understanding and risk identification tasks related to contractual clauses. This approach preserves the ability to understand legal semantics while significantly reducing the number of model parameters, thereby improving the speed of subsequent contract review.

[0078] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0079] Based on the same inventive concept, this application also provides a contract review apparatus for implementing the contract review method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more contract review apparatus embodiments provided below can be found in the limitations of the contract review method described above, and will not be repeated here.

[0080] In one exemplary embodiment, such as Figure 3 As shown, a contract review device is provided, comprising:

[0081] Module 302 is used to obtain the text data of the contract to be reviewed;

[0082] The filtering module 304 is used to identify risk-free clauses in text data using a lightweight classification model, and to filter multiple suspected risk clauses based on the identification results.

[0083] The first review module 306 is used to perform in-depth semantic analysis on multiple suspected risk clauses in sequence using a preset large model, and to determine the key elements and first risk review results corresponding to each suspected risk clause.

[0084] The second review module 308 is used to construct a logical association graph based on the key elements corresponding to multiple suspected risk clauses using a graph neural network model; determine the second risk review results between clauses based on the logical association graph; and obtain the review results of the contract to be reviewed based on the first risk review results and the second risk review results.

[0085] The aforementioned contract review device employs a hierarchical, cascaded large-scale model architecture. Initial screening is performed using a lightweight classification model, avoiding redundant calculations by feeding a large number of risk-free clauses into the large model, significantly reducing overall computational costs and response time. This allows the large model to concentrate its computing power on in-depth reasoning of high-risk clauses, improving the accuracy and depth of core risk identification and analysis. Furthermore, a logical association graph reveals the associated risks and potential conflicts between clauses, achieving global logical verification. Through rapid initial screening, in-depth semantic analysis, and logical association graph analysis, a balance between review efficiency and accuracy can be struck to meet the timeliness requirements of large-scale contract reviews.

[0086] In an exemplary embodiment, the first review module 306 is further configured to retrieve the review points corresponding to the contract to be reviewed from a preset knowledge graph; generate prompt words for prompting the preset big model based on the review points; and, under the guidance of the prompt words, use the preset big model to perform in-depth semantic analysis on multiple suspected risk clauses in sequence to determine the key elements and first risk review results corresponding to each suspected risk clause.

[0087] In an exemplary embodiment, the key points of review include historical high-frequency risk points and policy-related key points. The contract review device also includes a knowledge graph construction module, which is used to generate historical high-frequency risk points for each business scenario based on historical contract review records to obtain a first type of graph data; generate policy-related key points for each contract clause based on policy and regulatory texts to obtain a second type of graph data; and construct a preset knowledge graph based on the first type of graph data and the second type of graph data.

[0088] In one exemplary embodiment, the contract review device further includes a recommendation module, which is used to retrieve corresponding negotiation basis from a pre-built strategy knowledge base based on the risk type corresponding to the review result.

[0089] In an exemplary embodiment, the acquisition module 302 is further configured to evaluate the risk score of the text data; the screening module 304 is further configured to perform a step of using a lightweight classification model to identify risk-free clauses in the text data when the risk score is less than a set threshold, and to screen multiple suspected risk clauses based on the identification results; the first review module 306 is further configured to determine all clauses in the text data as suspected risk clauses when the risk score is greater than or equal to a set threshold, and to perform a step of using a preset large model to sequentially perform deep semantic analysis on multiple suspected risk clauses to determine the key elements corresponding to each suspected risk clause and the first risk review result.

[0090] In an exemplary embodiment, the contract review device further includes a training module, which is used to perform unsupervised pre-training on a large model based on contract-related corpus to obtain a domain-based basic model; and to perform knowledge distillation on the domain-based basic model to obtain a preset large model.

[0091] Each module in the aforementioned contract review device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0092] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores knowledge graph data or policy knowledge base data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a contract review method.

[0093] Those skilled in the art will understand that Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0094] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to perform the following steps: acquiring text data of a contract to be reviewed; using a lightweight classification model to identify risk-free clauses in the text data, and filtering based on the identification results to obtain multiple suspected risk clauses; using a preset large model to sequentially perform deep semantic analysis on the multiple suspected risk clauses to determine the key elements and first risk review result corresponding to each suspected risk clause; constructing a logical association graph based on the key elements corresponding to the multiple suspected risk clauses using a graph neural network model; determining the second risk review result between clauses based on the logical association graph; and obtaining the review result of the contract to be reviewed based on the first risk review result and the second risk review result.

[0095] In one embodiment, when the processor executes the computer program, it further performs the following steps: retrieving the audit points corresponding to the contract to be audited from a preset knowledge graph; generating prompt words for prompting the preset big model based on the audit points; and, under the guidance of the prompt words, using the preset big model to sequentially perform deep semantic analysis on multiple suspected risk clauses to determine the key elements and first risk audit results corresponding to each suspected risk clause.

[0096] In one embodiment, when the processor executes the computer program, it further performs the following steps: generating historical high-frequency risk points for each business scenario based on historical contract review records to obtain a first type of graph data; generating policy-related key points for each contract clause based on policy and regulatory texts to obtain a second type of graph data; and constructing a preset knowledge graph based on the first type of graph data and the second type of graph data.

[0097] In one embodiment, when the processor executes the computer program, it also performs the following steps: retrieving the corresponding negotiation basis from a pre-built strategy knowledge base based on the risk type corresponding to the audit result.

[0098] In one embodiment, when the processor executes the computer program, it further performs the following steps: assessing the risk score of the text data; if the risk score is less than a set threshold, performing a step of using a lightweight classification model to identify risk-free clauses in the text data, and filtering out multiple suspected risk clauses based on the identification results; if the risk score is greater than or equal to the set threshold, identifying all clauses in the text data as suspected risk clauses, and performing a step of using a preset large model to sequentially perform deep semantic analysis on multiple suspected risk clauses, and determining the key elements and first risk review result corresponding to each suspected risk clause.

[0099] In one embodiment, when the processor executes the computer program, it further performs the following steps: performing unsupervised pre-training on a large model based on contract-related corpus to obtain a domain-based basic model; and performing knowledge distillation on the domain-based basic model to obtain a pre-defined large model.

[0100] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, it performs the following steps: acquiring text data of the contract to be reviewed; using a lightweight classification model to identify risk-free clauses in the text data, and filtering based on the identification results to obtain multiple suspected risk clauses; using a pre-defined large model to sequentially perform deep semantic analysis on the multiple suspected risk clauses to determine the key elements and first risk review results corresponding to each suspected risk clause; constructing a logical association graph based on the key elements corresponding to the multiple suspected risk clauses using a graph neural network model; determining the second risk review results among the clauses based on the logical association graph; and obtaining the review result of the contract to be reviewed based on the first risk review result and the second risk review result.

[0101] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: retrieving the audit points corresponding to the contract to be audited from a preset knowledge graph; generating prompt words for prompting the preset big model based on the audit points; and, under the guidance of the prompt words, using the preset big model to sequentially perform deep semantic analysis on multiple suspected risk clauses to determine the key elements and first risk audit results corresponding to each suspected risk clause.

[0102] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: generating historical high-frequency risk points for each business scenario based on historical contract review records to obtain a first type of graph data; generating policy-related key points for each contract clause based on policy and regulatory texts to obtain a second type of graph data; and constructing a preset knowledge graph based on the first type of graph data and the second type of graph data.

[0103] In one embodiment, when the computer program is executed by the processor, it also performs the following steps: retrieving the corresponding negotiation basis from a pre-built strategy knowledge base based on the risk type corresponding to the audit result.

[0104] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: assessing the risk score of the text data; if the risk score is less than a set threshold, performing a step of using a lightweight classification model to identify risk-free clauses in the text data, and filtering out multiple suspected risk clauses based on the identification results; if the risk score is greater than or equal to the set threshold, identifying all clauses in the text data as suspected risk clauses, and performing a step of using a preset large model to sequentially perform deep semantic analysis on multiple suspected risk clauses, and determining the key elements and first risk review result corresponding to each suspected risk clause.

[0105] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: performing unsupervised pre-training on the large model based on contract-related corpus to obtain a domain-based basic model; and performing knowledge distillation on the domain-based basic model to obtain a pre-defined large model.

[0106] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, performs the following steps: acquiring text data of a contract to be reviewed; identifying risk-free clauses in the text data using a lightweight classification model, and filtering out multiple suspected risk clauses based on the identification results; sequentially performing deep semantic analysis on the multiple suspected risk clauses using a pre-defined large model to determine the key elements and first risk review result corresponding to each suspected risk clause; constructing a logical association graph based on the key elements corresponding to the multiple suspected risk clauses using a graph neural network model; determining the second risk review result among the clauses based on the logical association graph; and obtaining the review result of the contract to be reviewed based on the first and second risk review results.

[0107] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: retrieving the audit points corresponding to the contract to be audited from a preset knowledge graph; generating prompt words for prompting the preset big model based on the audit points; and, under the guidance of the prompt words, using the preset big model to sequentially perform deep semantic analysis on multiple suspected risk clauses to determine the key elements and first risk audit results corresponding to each suspected risk clause.

[0108] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: generating historical high-frequency risk points for each business scenario based on historical contract review records to obtain a first type of graph data; generating policy-related key points for each contract clause based on policy and regulatory texts to obtain a second type of graph data; and constructing a preset knowledge graph based on the first type of graph data and the second type of graph data.

[0109] In one embodiment, when the computer program is executed by the processor, it also performs the following steps: retrieving the corresponding negotiation basis from a pre-built strategy knowledge base based on the risk type corresponding to the audit result.

[0110] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: assessing the risk score of the text data; if the risk score is less than a set threshold, performing a step of using a lightweight classification model to identify risk-free clauses in the text data, and filtering out multiple suspected risk clauses based on the identification results; if the risk score is greater than or equal to the set threshold, identifying all clauses in the text data as suspected risk clauses, and performing a step of using a preset large model to sequentially perform deep semantic analysis on multiple suspected risk clauses, and determining the key elements and first risk review result corresponding to each suspected risk clause.

[0111] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: performing unsupervised pre-training on the large model based on contract-related corpus to obtain a domain-based basic model; and performing knowledge distillation on the domain-based basic model to obtain a pre-defined large model.

[0112] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0113] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0114] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0115] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A contract auditing method, characterized by, The method comprises: acquiring text data of a contract to be audited; identifying risk-free clauses in the text data using a lightweight classification model, and screening a plurality of suspected risk clauses according to the identification result; performing deep semantic analysis on the plurality of suspected risk clauses in sequence using a preset large model to determine key elements and a first risk audit result corresponding to each suspected risk clause; constructing a logical association graph according to the key elements corresponding to the plurality of suspected risk clauses through a graph neural network model; determining a second risk audit result between clauses according to the logical association graph; obtaining an audit result of the contract to be audited according to the first risk audit result and the second risk audit result.

2. The method of claim 1, wherein, The method comprises: retrieving audit points corresponding to the contract to be audited from a preset knowledge graph; generating prompt words for prompting the preset large model according to the audit points; under the instruction of the prompt words, performing deep semantic analysis on the plurality of suspected risk clauses in sequence using a preset large model to determine key elements and a first risk audit result corresponding to each suspected risk clause.

3. The method of claim 2, wherein, The audit points include historical high-frequency risk points and policy association points, and the method further comprises: generating historical high-frequency risk points in each business scenario according to historical contract audit records to obtain first graph data; generating policy association points of each contract clause according to policy and regulation texts to obtain second graph data; constructing a preset knowledge graph according to the first graph data and the second graph data.

4. The method of claim 1, wherein, After obtaining the audit result of the contract to be audited according to the first risk audit result and the second risk audit result, the method further comprises: retrieving corresponding negotiation basis from a pre-constructed strategy knowledge base according to the risk type corresponding to the audit result.

5. The method according to any one of claims 1 to 4, characterized in that, After acquiring the text data of the contract to be audited, the method further comprises: evaluating a risk score of the text data; in the case that the risk score is less than a set threshold, performing the step of identifying risk-free clauses in the text data using a lightweight classification model, and screening a plurality of suspected risk clauses according to the identification result; in the case that the risk score is greater than or equal to the set threshold, determining all clauses in the text data as suspected risk clauses, and performing the step of performing deep semantic analysis on the plurality of suspected risk clauses in sequence using a preset large model to determine key elements and a first risk audit result corresponding to each suspected risk clause.

6. The method according to any one of claims 1 to 4, characterized in that, The method further comprises: performing unsupervised pre-training on a large model according to contract-related corpus to obtain a domain basic model; performing knowledge distillation processing on the domain basic model to obtain a preset large model.

7. A contract auditing apparatus characterized by comprising: The device comprises: an acquisition module configured to acquire text data of a contract to be audited; a screening module configured to identify risk-free clauses in the text data using a lightweight classification model, and screen a plurality of suspected risk clauses according to the identification result; The first auditing module is configured to sequentially perform deep semantic analysis on the plurality of suspected risk clauses by using a preset large model, determine key elements corresponding to each suspected risk clause, and determine a first risk auditing result of each suspected risk clause. The second auditing module is configured to construct a logical association graph based on the key elements corresponding to the plurality of suspected risk clauses by using a graph neural network model, determine a second risk auditing result between clauses based on the logical association graph, and obtain an auditing result of the contract to be audited based on the first risk auditing result and the second risk auditing result.

8. A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images