Intelligent contract auditing system and contract auditing method based on large model

Through the coordinated work of the triple audit architecture and large model combined with the forward and reverse knowledge base, the problem of low contract audit accuracy in the existing technology is solved, high accuracy and adaptability are achieved, and the enterprise-specific competitive barriers are formed.

CN119090425APending Publication Date: 2024-12-06YUEZHI ARTIFICIAL INTELLIGENCE (SHENZHEN) CO LTD

Patent Information

Application Number
CN202411115423.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing artificial intelligence contract review system relies on the general capabilities of large models and prompt engineering, has low accuracy, lacks differentiated solutions, and cannot accumulate user professional experience and form a proprietary competitive barrier for enterprises.

Method used

Adopt a triple audit architecture, including risk recall based on big models, secondary optimization of forward and reverse knowledge bases, manual audit and final confirmation, through the collaborative work of the user interaction layer, service layer and model layer, using Chunk, Embedding, Prompt engineering Contract audit with the LLM model, and filter and optimize risk points in combination with forward and reverse knowledge bases.

Benefits of technology

It significantly improves the accuracy and comprehensiveness of contract audits, enhances the accuracy of AI audits, and improves the system's adaptability and user experience through manual confirmation and knowledge accumulation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119090425A_ABST
    Figure CN119090425A_ABST
Patent Text Reader

Abstract

The invention discloses an intelligent contract auditing system and contract auditing method based on a large model. The technical architecture of triple auditing is adopted, the auditing accuracy is improved, and triple auditing comprises risk recall based on a large model, secondary optimization based on a forward and reverse knowledge base and manual auditing and final confirmation. The first-layer audit uses a general large model and audit key points to audit risk points in a contract, the second-layer audit carries out secondary optimization on the risk points audited by the large model based on a forward and reverse knowledge base, and the third-layer audit is to manually confirm, supplement and modify risk contents audited by the AI. The method comprises the following steps: feeding back an AI auditing result by a legal expert, annotating and optimizing omissive and inaccurate auditing content, and then depositing corrected information to a forward knowledge base so as to enhance the accuracy and comprehensiveness of AI auditing. And meanwhile, contents which do not need to be audited are effectively filtered and are classified into a reverse knowledge base so as to clearly define an audit category.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology and machine learning algorithm technology, and in particular to an intelligent contract review system and a contract review method based on a large model. Background Art

[0002] At present, the artificial intelligence contract review system on the market only relies on the general capabilities of large models and prompt engineering to review contracts, and the accuracy rate is low. At the same time, the contract review solutions on the market only stay at the level of large models + prompt tuning, and there is no differentiated solution that is more suitable for customers. In addition, during the use of the product, the user's professional experience will not be accumulated, and it will not be able to form a proprietary competitive barrier for the enterprise in the long run.

[0003] Therefore, the prior art has defects and needs to be improved. Summary of the invention

[0004] The purpose of the present invention is to overcome the deficiencies of the prior art and provide an intelligent contract review system and a contract review method based on a large model.

[0005] The technical solution of the present invention is as follows: Provide a smart contract review system based on a large model, including: a user interaction layer, a service layer and a model layer, wherein the service layer includes a user service end and an operation service end, wherein:

[0006] User interaction layer: the device or mobile device used by the user to input and execute user interaction actions;

[0007] Service layer: includes user service end and operation service end. The user service end interacts with the user interaction layer to identify and interactively process the information input by the user interaction layer, including login and management module, contract upload module and contract review module. The operation service end analyzes and optimizes the contract according to the interactive information received from the user service end, including knowledge base module, data analysis module and default review key points module.

[0008] Model layer: Configure various large models for identification and review, which are convenient for the service layer to call for review.

[0009] Furthermore, the service layer conducts three review processes when reviewing and analyzing the contract uploaded by the user, including:

[0010] First audit: The contract will be audited based on the default audit points and the self-set audit points by calling the big model stored in the model layer;

[0011] Second audit: Based on the risk results returned by the first audit, the audit results are filtered and optimized through the positive knowledge base and reverse knowledge base in the knowledge base module;

[0012] The third audit: The risk points that have been filtered and optimized by the knowledge base module will be fed back to the user interaction layer through the user server and displayed. Users can manually optimize or ignore the risk points in the audit results through the user interaction layer. At the same time, the risk points that have not been audited in the audit results will be annotated for knowledge accumulation, and the modification records will be saved in the knowledge base.

[0013] Furthermore, the knowledge base includes a forward knowledge base and a reverse knowledge base.

[0014] Furthermore, the large model in the model layer includes: Chunk, Embedding, Prompt project and LLM.

[0015] The present invention also provides a smart contract audit method based on a big model, which uses the smart contract audit system based on a big model as described above to perform contract audit, including the following steps:

[0016] Step 1: The operator fills in the audit points, and the default audit points are saved in the database through the operation server;

[0017] Step 2: The user uploads the contract and selects the contract type in the user interaction layer, and selects and enters relevant contract information and self-set review points;

[0018] Step 3: After receiving the contract information, self-set review points and default review points from the user interaction layer, the service layer conducts the first review of the uploaded contract and returns the corresponding risk clauses;

[0019] Step 4: Based on the risk results returned by the first audit, the operation server in the service layer filters and optimizes the audit results through the knowledge base, and returns the audit results to the user server, which is displayed through the user interaction layer;

[0020] Step 5: The user manually processes the risk points through the user interaction layer according to the audit results displayed on the user interaction layer, and manually annotates the risk points that are not identified in the audit results, and saves the modification records to the knowledge base;

[0021] Step 6: After completing the review and saving of the contract based on the user's selection and processing, return and make the annotated draft and revised draft.

[0022] Furthermore, the specific steps of step 2 are:

[0023] Step 2.1: The user uploads the contract and selects the contract type. The user server queries the corresponding default review points from the database based on the contract type selected by the user;

[0024] Step 2.2: After the user enters or selects the contract information and audit optimization requirements on the user interaction layer, the user service returns the default audit points to the user interaction layer, and the service layer performs the first audit based on the contract information and audit optimization requirements selected by the user.

[0025] Furthermore, the specific steps of step 3 are:

[0026] Step 3.1: The user requests audit optimization, and the service layer calls the corresponding large model from the model layer to optimize the user's self-set audit points;

[0027] Step 3.2: After receiving the request information, the model layer optimizes the user's self-set review points;

[0028] Step 3.3: After the model layer completes the optimization of the self-set audit points, it returns the optimization results to the service layer;

[0029] Step 3.4: Call the contract review interface to identify and review the contract based on the optimized self-set review points and default review points;

[0030] Step 3.5: After completing the review of the contract, identify and analyze the risk clauses in the contract and return the risk clauses to the service layer.

[0031] Furthermore, the specific steps of step 4 are:

[0032] Step 4.1: The operation server obtains the reverse knowledge base according to the contract type selected by the user;

[0033] Step 4.2: Risk terms are matched with the reverse direct database for similarity;

[0034] Step 4.3: Filter risk terms with high similarity to the reverse knowledge base;

[0035] Step 4.4: Obtain positive knowledge base according to contract type;

[0036] Step 4.5: Assemble Prompt according to the positive knowledge base and contract content;

[0037] Step 4.6: Adjust and review the content of the contract terms according to the prompt calling the big model in the model layer;

[0038] Step 4.7: Return the final review results.

[0039] Furthermore, the specific steps of step 5 are:

[0040] Step 5.1: The service layer returns the audit results in step 4 to the user interaction layer, and the user manually reviews the returned audit results;

[0041] Step 5.2: The user chooses whether to ignore the risk points returned in the audit results;

[0042] Step 5.3: For the risk points that are not reviewed in the review results, the user makes corresponding annotations;

[0043] Step 5.4: The user modifies and saves the audit results;

[0044] Step 5.5: For the user's modification of the audit results and the processing of risk points, the service layer stores the user's modification of the audit results in the positive knowledge base, and stores the audit results ignored by the user in the negative knowledge base;

[0045] Step 5.6: Generate a note draft and a revised draft based on the user's modifications and ignored operations on the review results, and return them to the user interaction layer for display and export.

[0046] Adopting the above scheme, the present invention adopts a triple audit technical architecture to improve the audit accuracy. The triple audit is: risk recall based on a large model, secondary optimization based on a positive and reverse knowledge base, and manual review and final confirmation. The first level of audit uses a general large model and audit points to review the risk points in the contract. The second level of audit is based on the positive and reverse knowledge base to conduct secondary optimization of the risk points audited by the large model. The third level of audit is to manually confirm and supplement the risk content audited by AI. Legal experts provide feedback on the results of the AI ​​audit, annotate and optimize the omitted and inaccurate audit content, and then deposit the corrected information into the positive knowledge base to enhance the accuracy and comprehensiveness of the AI ​​audit. At the same time, effective filtering is implemented for content that does not need to be reviewed, and it is included in the reverse knowledge base to clearly define the scope of the audit. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a functional architecture diagram of the platform of the present invention.

[0048] Figure 2 This is a diagram of the entire audit process and functional breakdown of the present invention.

[0049] Figure 3 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0050] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0051] See also Figure 1 , Figure 2 The present invention provides a smart contract review system based on a large model, comprising: a user interaction layer, a service layer and a model layer, wherein the service layer comprises a user service end and an operation service end, wherein:

[0052] User interaction layer: The device or mobile device used by the user to input and execute the user's interactive actions. In some embodiments, the user interaction layer refers to the APP program on the PC or mobile terminal used by the user, which provides a platform for users to upload contracts and select and enter contract information, and transmit information with the service layer through the platform, so as to meet the needs of AI intelligent review of contracts and manual screening and modification of review results by users.

[0053] Service layer: Perform corresponding data processing and feedback based on the interactive information fed back by the user interaction layer, and review and analyze the contracts uploaded by users.

[0054] User server: interacts with the user interaction layer to identify and interactively process the information input by the user interaction layer, including login and management module, contract upload module and contract review module.

[0055] Specifically, the login and management module includes: a user information registration and login submodule, a historical contract management submodule, and a contract search submodule.

[0056] The user information registration and login submodule can facilitate users to register and log in to their accounts, thereby establishing a connection with the operation server. The historical contract management submodule can conveniently store the contracts uploaded or reviewed by users in historical operations, which is convenient for subsequent calls and viewing. The contract search submodule provides customers with search services, which facilitates quick search of historical contract documents and risk review assessments that users need to view in historical contracts.

[0057] The contract upload module includes: a contract type setting submodule, an agency relationship setting submodule, and an audit points setting submodule.

[0058] The contract type setting submodule supports users to select and set the type of the uploaded contract, so that the operation server can call the corresponding model from the model layer to conduct contract audit analysis. The agency relationship setting submodule supports users to input and set the agency relationship attached to the uploaded contract. The audit key points setting submodule supports users to input and set their own audit key points, so that the large model in the model layer can provide corresponding audit analysis results based on the user's self-set audit key points.

[0059] The contract review module includes: a new annotation sub-module, a suggestion to ignore / replace sub-module, and a contract export sub-module.

[0060] The newly added annotation submodule supports users to add and annotate risk points that were not raised by the AI ​​audit after the system provides the AI ​​audit results. The suggestion to ignore / replace submodule supports users to ignore or replace risk points in the audit results provided by the system. Through the above two submodules, users can manually modify the audit results to meet their actual audit needs and provide templates for subsequent deep learning of large models. The contract export submodule supports users to export contracts that have completed the audit.

[0061] Operation server: Analyze and optimize the contract based on the interactive information received in the user server, including knowledge base module, data analysis module and default review points module.

[0062] Specifically, the knowledge base module includes a forward knowledge base and a reverse knowledge base.

[0063] The forward knowledge base stores the modification records of the user's audit results of historical contract information, so as to facilitate in-depth learning and audit analysis of similar contract content in subsequent contract audit work. After the initial generation of risk results using AI, the manual review process of legal experts is introduced to identify and mark the risk items omitted by AI and correct the content that is incorrectly audited, and then incorporate this optimized information into the forward knowledge base. When a user uploads a new contract, the system first uses the large model's own capabilities and audit points to conduct a preliminary risk screening, and then filters out known non-critical risk points through reverse knowledge base comparison. For the remaining risk points, the system will further match them with the entries in the forward knowledge base. If the risk content and prompts identified by AI match the records in the knowledge base, the system will fine-tune and optimize the AI's audit results based on the professional knowledge in the forward knowledge base.

[0064] The reverse knowledge base stores records of users' ignored audit results of historical contract information, so as to ignore similar risk points in subsequent contract audits. Based on the risk results generated by AI, known non-critical risk points are ignored through the judgment of legal experts and stored in the reverse knowledge base. When a user uploads a new contract, the system will first use the large model's own capabilities and audit points to conduct a preliminary risk assessment, and the assessed risk points will automatically be compared with the reverse knowledge base for deep similarity. If the comparison results show a high degree of match, the corresponding risk content will be automatically filtered out; if the match is insufficient, the matching process with the forward knowledge base will be entered.

[0065] The data analysis module includes a contract analysis submodule, a user analysis submodule and a user behavior analysis submodule.

[0066] The contract analysis submodule can scan the contract files uploaded by the user, and call the big model in the model layer to review and analyze the contract files, so as to extract and mark the risk points in the contract files. The user analysis submodule can analyze the self-set review points provided by the user, so as to extract the risk points using the corresponding review points during the contract review process. The user behavior analysis submodule can analyze a series of operations such as modification, replacement, and ignoring performed by the user, so as to conduct in-depth learning of the big model, so as to provide review and analysis results that are closer to user needs in the subsequent use process.

[0067] The default audit points module includes the contract type submodule and the AI ​​optimization submodule. The contract type submodule stores the default audit points corresponding to each type of contract, which facilitates the AI ​​optimization submodule to call the corresponding type of contract and perform contract optimization and audit work.

[0068] Model layer: Configure various large models for identification and review, which are convenient for the service layer to call for review.

[0069] The large models in the model layer include: Chunk, Embedding, Prompt project and LLM.

[0070] Chunk models are used for text analysis, especially when building graph structures. Unlike traditional graph neural network research, the chunk-based approach first uses an unsupervised method to obtain chunk information in the input, and then builds a graph structure based on the relationship between words within and between chunks. This approach does not require dependency syntactic results based on automatic analysis, thus providing a new way to build graph structures.

[0071] The core of the Embedding model is to convert raw data (such as text, images, etc.) into low-dimensional vector representations that retain some key features of the original data. This conversion helps reduce the sparsity of data, improves computational efficiency, and allows processing using various machine learning algorithms. In natural language processing, the Embedding model is able to convert words or sentences into vectors, so that text data can be quantitatively analyzed and processed.

[0072] Prompt engineering refers to the art and science of designing and optimizing instructions for interacting with AI models. Prompt can be thought of as a translator between humans and AI, translating requirements into instructions that AI can understand and execute. High-quality prompts can significantly improve the quality and relevance of AI output, such as allowing AI to generate more accurate text, more creative stories, or more insightful analysis. When designing prompts, you can use the CHAT model as a framework, namely Character, History, Ambition, Task, to provide AI with clear role positioning, background information, goals, and specific tasks, thereby guiding AI to generate more targeted and valuable answers. In addition, prompt engineering is not just a skill, but also an art, and it is a basic skill that everyone who interacts with AI should master. A precise and in-depth prompt can guide the model to give the expected answer, and this skill is a common skill for major AI models.

[0073] LLM, or Large Language Model (LLM for short), refers to a deep learning model trained with a large amount of text data that can generate natural language text or understand the meaning of language text. Large language models can handle a variety of natural language tasks, such as text classification, question-answering, and dialogue, and are an important path to artificial intelligence. Currently, large language models use a Transformer architecture and pre-training objectives (such as Language Modeling) similar to small models. The difference from small models is the increase in model size, training data, and computing resources.

[0074] The service layer conducts three review processes when reviewing and analyzing the contract uploaded by the user, including:

[0075] First audit: The contract will be audited based on the default audit points and the self-set audit points by calling the big model stored in the model layer;

[0076] Second audit: Based on the risk results returned by the first audit, the audit results are filtered and optimized through the positive knowledge base and reverse knowledge base in the knowledge base module;

[0077] The third audit: The risk points filtered and optimized by the knowledge base module will be fed back to the user interaction layer through the user service end and displayed. The user can manually optimize or ignore the risk points in the audit results through the user interaction layer. At the same time, the risk points that have not been audited in the audit results are annotated to precipitate knowledge, and the modification records are saved in the knowledge base.

[0078] See also Figure 3The present invention also provides a smart contract audit method based on a big model, which uses the smart contract audit system based on a big model as described above to perform contract audit, including the following steps:

[0079] Step 1: The operator fills in the audit points, and the default audit points are saved in the database. Before the audit, a framework audit point is given to each contract type. The big model will audit the contract text based on the content of the default audit points to improve the accuracy of AI audit. In order to improve the accuracy of the default audit points, professional legal personnel are required to fill them in on the operation side to improve the efficiency of the default audit work.

[0080] Step 2: The user uploads the contract and selects the contract type in the user interaction layer, and selects and enters relevant contract information and self-set review points. The specific steps are:

[0081] Step 2.1: The user uploads the contract and selects the contract type. The user server queries the corresponding default review points from the database based on the contract type selected by the user.

[0082] Step 2.2: After the user enters or selects the contract information and audit optimization requirements on the user interaction layer, the user service returns the default audit points to the user interaction layer, and the service layer performs the first audit based on the contract information and audit optimization requirements selected by the user.

[0083] The user gives a targeted audit point based on the different contract contents, and the big model will audit based on the self-set audit points to improve the accuracy of AI audit. At the same time, by selecting the corresponding contract type after uploading the contract, the big model can call the default audit points of the corresponding contract type, and the user's self-set audit points can be stored in the default audit points of the corresponding contract type to meet the needs of learning iteration.

[0084] Step 3: After receiving the contract information, self-set review points and default review points from the user interaction layer, the service layer conducts the first review of the uploaded contract and returns the corresponding risk clauses. The risk points in the contract are extracted to facilitate subsequent filtering and review. The specific steps are:

[0085] Step 3.1: The user requests audit optimization, and the service layer calls the corresponding large model from the model layer to optimize the user's self-set audit points.

[0086] Step 3.2: After receiving the request information, the model layer optimizes the user's self-set review points.

[0087] Step 3.3: After the model layer completes the optimization of the self-set audit points, it returns the optimization results to the service layer.

[0088] Step 3.4: Call the contract review interface to identify and review the contract based on the optimized self-set review points and default review points.

[0089] Step 3.5: After completing the review of the contract, identify and analyze the risk clauses in the contract and return the risk clauses to the service layer.

[0090] Step 4: Based on the risk results returned by the first audit, the operation server in the service layer filters and optimizes the audit results through the knowledge base, and returns the audit results to the user server, which is displayed through the user interaction layer. By filtering known non-critical risk points, the user's audit cost is reduced and the accuracy is improved. The specific steps are:

[0091] Step 4.1: The operation server obtains the reverse knowledge base according to the contract type selected by the user.

[0092] Step 4.2: Risk terms are matched with the reverse direct library for similarity.

[0093] Step 4.3: Filter risk terms with high similarity to the reverse knowledge base.

[0094] Step 4.4: Obtain positive knowledge base based on contract type.

[0095] Step 4.5: Assemble Prompt based on the positive knowledge base and contract content.

[0096] Step 4.6: Adjust and review the content of the contract terms based on the big model in the model layer called by Prompt.

[0097] Step 4.7: Return the final review results.

[0098] The forward knowledge base is intended to store the knowledge accumulated by users in the process of using the system provided by the present invention. The types of accumulated knowledge include: (1) risk warnings and suggestions generated by AI, which are considered to need to be optimized by humans; (2) risk points that are not identified by AI, which are considered to need to be identified by AI later.

[0099] The reverse knowledge base is designed to store the knowledge accumulated by users in the process of using the product. The accumulated types include risk points that are manually considered to be reviewed by AI and do not need to be reviewed again. For example, some law firms will remove the subject information in the contract in advance for privacy and security. When the contract is reviewed, there will be a risk that the subject information is indeed a risk. Although the risk is correct, there is no need to review the risk point all the time in the future. Therefore, this type of risk point can be ignored and accumulated in the reverse knowledge base.

[0100] In the field of AI, especially in the context of natural language processing (NLP) and generative AI, prompt refers to the input or instruction provided to the AI ​​system to guide the AI ​​to perform specific operations or generate specific outputs. It is one of the main ways for users to interact with AI systems. By providing different prompts, users can guide AI to perform various tasks, such as answering questions, generating text, identifying objects in images, etc.

[0101] Step 5: Based on the audit results displayed on the user interaction layer, the user manually processes the risk points through the user interaction layer and saves the modification records to the knowledge base. The specific steps are:

[0102] Step 5.1: The service layer returns the audit results in step 4 to the user interaction layer, and the user manually reviews the returned audit results.

[0103] Step 5.2: The user chooses whether to ignore the risk points returned in the audit results.

[0104] Step 5.3: For risk points that are not reviewed in the review results, the user makes corresponding annotations.

[0105] Step 5.4: The user modifies and saves the audit results.

[0106] Step 5.5: For the user's modification of the audit results and the handling of risk points, the service layer stores the user's modification of the audit results in the forward knowledge base, and stores the audit results ignored by the user in the reverse knowledge base.

[0107] Step 5.6: Generate a note draft and a revised draft based on the user's modifications and ignored operations on the review results, and return them to the user interaction layer for display and export.

[0108] According to the manual modification and annotation of users, a learning template can be provided for the iteration and upgrade of the large model, and the omitted and inaccurate audit content can be annotated and optimized, and then the corrected information can be deposited into the positive knowledge base to enhance the accuracy and comprehensiveness of AI audit. At the same time, the content that does not need to be audited can be effectively filtered and classified into the reverse knowledge base to clearly define the audit scope.

[0109] Step 6: After completing the review and saving of the contract based on the user's selection and processing, return and make the annotated draft and revised draft.

[0110] In summary, the present invention adopts a triple audit technical architecture to improve the accuracy of the audit. The triple audit is: risk recall based on a large model, secondary optimization based on a positive and reverse knowledge base, and manual review and final confirmation. The first level of audit uses a general large model and audit points to review the risk points in the contract. The second level of audit is based on the positive and reverse knowledge base to conduct secondary optimization of the risk points audited by the large model. The third level of audit is to manually confirm and supplement the risk content audited by AI. Legal experts provide feedback on the results of the AI ​​audit, annotate and optimize the omitted and inaccurate audit content, and then deposit the corrected information into the positive knowledge base to enhance the accuracy and comprehensiveness of the AI ​​audit. At the same time, effective filtering is implemented for content that does not require review, and it is included in the reverse knowledge base to clearly define the scope of the audit.

[0111] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A smart contract audit system based on a big model, It is characterized in that include: User interaction layer, service layer and model layer. The service layer includes user service end and operation service end, among which: User interaction layer: the device or mobile device used by the user to input and execute user interaction actions; Service layer: includes user service end and operation service end. The user service end interacts with the user interaction layer to identify and interactively process the information input by the user interaction layer, including login and management module, contract upload module and contract review module. The operation service end analyzes and optimizes the contract according to the interactive information received from the user service end, including knowledge base module, data analysis module and default review key points module. Model layer: Configure various large models for identification and review, which are convenient for the service layer to call for review.

2. According to the big model-based smart contract audit system of claim 1, It is characterized in that The service layer conducts three review processes when reviewing and analyzing the contract uploaded by the user, including: First audit: The contract will be audited based on the default audit points and the self-set audit points by calling the big model stored in the model layer; Second audit: Based on the risk results returned by the first audit, the audit results are filtered and optimized through the positive knowledge base and reverse knowledge base in the knowledge base module; The third audit: The risk points that have been filtered and optimized by the knowledge base module will be fed back to the user interaction layer through the user server and displayed. Users can manually optimize or ignore the risk points in the audit results through the user interaction layer. At the same time, the risk points that have not been audited in the audit results will be annotated for knowledge accumulation, and the modification records will be saved in the knowledge base.

3. According to the big model-based smart contract audit system according to claim 1 or 2, It is characterized in that The knowledge base includes a forward knowledge base and a reverse knowledge base.

4. According to the big model-based smart contract audit system of claim 1, It is characterized in that The large models in the model layer include: Chunk, Embedding, Prompt project and LLM.

5. A smart contract audit method based on a big model, It is characterized in that The smart contract review system based on the big model as claimed in any one of claims 1 to 4 is used to conduct contract review, comprising the following steps: Step 1: The operator fills in the audit points, and the default audit points are saved in the database through the operation server; Step 2: The user uploads the contract and selects the contract type in the user interaction layer, and selects and enters relevant contract information and self-set review points; Step 3: After receiving the contract information, self-set review points and default review points from the user interaction layer, the service layer conducts the first review of the uploaded contract and returns the corresponding risk clauses; Step 4: Based on the risk results returned by the first audit, the operation server in the service layer filters and optimizes the audit results through the knowledge base, and returns the audit results to the user server, which is displayed through the user interaction layer; Step 5: The user manually processes the risk points through the user interaction layer according to the audit results displayed on the user interaction layer, and manually annotates the risk points that are not identified in the audit results, and saves the modification records to the knowledge base; Step 6: After completing the review and saving of the contract based on the user's selection and processing, return and make the annotated draft and revised draft.

6. According to the big model-based smart contract audit method of claim 5, It is characterized in that The specific steps of step 2 are: Step 2.1: The user uploads the contract and selects the contract type. The user server queries the corresponding default review points from the database based on the contract type selected by the user; Step 2.2: After the user enters or selects the contract information and audit optimization requirements on the user interaction layer, the user service returns the default audit points to the user interaction layer, and the service layer performs the first audit based on the contract information and audit optimization requirements selected by the user.

7. According to the big model-based smart contract audit method of claim 5, It is characterized in that The specific steps of step 3 are: Step 3.1: The user requests audit optimization, and the service layer calls the corresponding large model from the model layer to optimize the user's self-set audit points; Step 3.2: After receiving the request information, the model layer optimizes the user's self-set review points; Step 3.3: After the model layer completes the optimization of the self-set audit points, it returns the optimization results to the service layer; Step 3.4: Call the contract review interface to identify and review the contract based on the optimized self-set review points and default review points; Step 3.5: After completing the review of the contract, identify and analyze the risk clauses in the contract and return the risk clauses to the service layer.

8. According to the big model-based smart contract audit method of claim 5, It is characterized in that The specific steps of step 4 are: Step 4.1: The operation server obtains the reverse knowledge base according to the contract type selected by the user; Step 4.2: Risk terms are matched with the reverse direct database for similarity; Step 4.3: Filter risk terms with high similarity to the reverse knowledge base; Step 4.4: Obtain positive knowledge base according to contract type; Step 4.5: Assemble Prompt according to the positive knowledge base and contract content; Step 4.6: Adjust and review the content of the contract terms according to the prompt calling the big model in the model layer; Step 4.7: Return the final review results.

9. According to the big model-based smart contract audit method of claim 5, It is characterized in that The specific steps of step 5 are: Step 5.1: The service layer returns the audit results in step 4 to the user interaction layer, and the user manually reviews the returned audit results; Step 5.2: The user chooses whether to ignore the risk points returned in the audit results; Step 5.3: For the risk points that are not reviewed in the review results, the user makes corresponding annotations; Step 5.4: The user modifies and saves the audit results; Step 5.5: For the user's modification of the audit results and the processing of risk points, the service layer stores the user's modification of the audit results in the positive knowledge base, and stores the audit results ignored by the user in the negative knowledge base; Step 5.6: Generate a note draft and a revised draft based on the user's modifications and ignored operations on the review results, and return them to the user interaction layer for display and export.

Citation Information

Patent Citations

  • Contract file review analysis method based on contract knowledge base

    CN112632958A

  • Contract review method, device and system and computer readable storage medium

    CN114549241A

  • Intelligent contract review method based on natural language understanding

    CN115269874A

  • Automatic contract review method and device

    CN116384387A

  • Intelligent legal affair self-service contract review method based on large model

    CN118377707A

Cited By

  • Carbon emission intelligent auditing system based on large model

    CN120509858A

  • A large model-based carbon emission intelligent auditing system

    CN120509858B

  • Configurable legal contract intelligent review and customized annotation method based on stand identification

    CN122367404A