Artificial Intelligence-Based Intelligent Contract Review Method and System

Abstract

This invention discloses an intelligent contract review method and system based on artificial intelligence, relating to the fields of computer and artificial intelligence technologies. The method includes: receiving and parsing contract documents to obtain structured text data and contract type; automatically extracting multiple key clauses based on the contract type; performing compliance review and difference comparison on the key clauses to generate intermediate review results containing risk points and modification suggestions; dynamically analyzing the clause content in the intermediate review results, determining the responsible party for review according to preset rules, and automatically generating and assigning independent review sub-tasks to corresponding review terminals; completing the review process based on feedback from the review terminals, and structurally storing the results in a contract knowledge base. The system includes modules for document parsing, intelligent analysis, task assignment, and knowledge base management. This invention achieves automation, intelligence, and collaboration in contract review, significantly improving review efficiency, risk identification accuracy, and process traceability, and promoting the accumulation and reuse of enterprise contract knowledge.

Description

Technical Field

[0001] This invention relates to the fields of computer technology and artificial intelligence application technology, and in particular to an intelligent contract review method and system based on artificial intelligence. Background Technology

[0002] In the daily operations of large enterprises, especially multinational corporations, contract review is a high-frequency, high-load core business process for legal, business, and related functional departments. Currently, this process mainly exists in two models, both of which have significant technical shortcomings: 1. Traditional purely manual review model: This model relies entirely on the reviewer's personal experience and subjective judgment, and has four inherent flaws: Inefficiency: Reviewers must read through lengthy contract texts word by word, which is time-consuming and laborious, making it difficult to cope with the increasing number and complexity of contracts and becoming a bottleneck in business processes.

[0003] Inconsistent and easily overlooked risk controls: Limited by individual knowledge, experience level, and work status, it is difficult to systematically and comprehensively identify all potential legal risks, compliance points, and business pitfalls in contracts. Inconsistent judgment standards among different auditors lead to significant subjective differences in audit results, making it impossible to establish a unified risk control standard.

[0004] The process is opaque and difficult to manage: the review process usually relies on offline paper circulation, email or instant messaging tools for communication, which makes it impossible to track the review progress in real time, the key responsibility nodes are unclear, and the historical modification traces and decision-making basis are difficult to fully trace, forming a process "black box".

[0005] Knowledge assets cannot be retained and reused: Valuable auditing experience and judgment logic are retained by individual employees and are lost with staff turnover. They cannot be transformed into structured knowledge assets that can be shared by the team and used for machine learning, resulting in repeated investigations of similar risks and hindering the continuous evolution of auditing capabilities.

[0006] 2. Existing electronic or basic automation tools: Some contract management software or tools are available on the market, but their technical limitations are obvious. Most products focus on the electronic storage of documents and static process approvals, but they do not fundamentally change the core of human reading and judgment, and cannot deeply understand and analyze the content of contracts.

[0007] The few tools that attempt to introduce automation mostly employ rule-based matching techniques based on simple keywords or fixed templates. This approach suffers from low accuracy, high false positive rates, and poor flexibility and adaptability. It cannot understand complex and ever-changing natural language expressions, contextual logic, and diverse business scenarios, making it difficult to meet the accuracy requirements of enterprise-level contract review.

[0008] In summary, existing technical solutions cannot effectively address the multiple challenges of "efficiency, quality, collaboration, and knowledge management" in contract review. Therefore, the industry urgently needs a comprehensive intelligent solution that can deeply understand the semantic content of contracts, intelligently identify and assess multi-dimensional risks, automatically perform differentiated comparisons, and drive efficient, transparent, and traceable multi-party collaborative processes. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide an intelligent contract review method and system based on artificial intelligence. This method uses an AI model to automatically analyze contracts, intelligently review risks and compare them with historical contracts. It achieves multi-department collaboration through dynamic task assignment based on the content of the terms, solving the problems of low efficiency, easy risk omission, opaque process and difficulty in knowledge reuse in traditional review.

[0010] To achieve the above objectives, the present invention provides the following technical solution: The intelligent contract review method based on artificial intelligence provided by this invention includes the following steps: S1: Receive contract documents pending review; S2: Parse the contract document to obtain structured text data and contract type; S3: Based on the contract type, automatically extract multiple key clauses from the structured text data; S4: Conduct compliance review and difference comparison of the key terms, and generate intermediate review results including risk points and modification suggestions; S5: Based on the clause content in the intermediate audit results, perform dynamic analysis, determine at least one auditing party according to the preset configurable business keyword-responsible department mapping rules, and automatically generate an independent auditing subtask for each auditing party based on the clause content and AI auditing suggestions related to the auditing party. The content of the sub-auditing task only includes clause content, risk points and AI auditing suggestions related to the responsibilities of the responsible party, and is accompanied by contextual information related to the assignment reason. Then, each independent auditing subtask is assigned to the corresponding auditing terminal. S6: Based on feedback from the review terminal, complete the contract review process and store the review results in a structured manner in the contract knowledge base.

[0011] Furthermore, in step S2, the parsing of the contract document includes: The pre-trained large language model API service is invoked to perform OCR recognition and text parsing on the contract document, converting it into structured JSON format text data containing chapter titles, paragraphs, tables, and list tags.

[0012] Furthermore, in step S3, the automatic extraction of multiple key terms includes: Based on the contract type, a finely tuned large language model is selected from the model pool as the clause extraction model. The structured JSON text data is input into the model, and named entity recognition and sequence labeling technologies are used to automatically identify and extract a set of predefined key clause entities and their contents.

[0013] Furthermore, in step S4, the difference comparison includes: Based on the contract type and contracting party information, the most relevant standard template contracts or signed historical contracts are automatically selected from the historical contract database as a comparison benchmark.

[0014] Furthermore, in step S5, the dynamic analysis based on the content of the clauses in the intermediate review results includes: From the intermediate audit results, identify specific clauses and their contents that present risks or discrepancies; Based on the preset "clause keyword responsible department" mapping rules, the keywords in the specific clause content are analyzed to dynamically determine the responsible department corresponding to the at least one review terminal; The step of automatically assigning the generated review tasks to at least one corresponding review terminal includes: For each identified responsible department, an independent sub-review task is generated. The content of the sub-review task only includes clauses, risk points, and AI review suggestions related to the department's responsibilities. Furthermore, in step S5, the review task also includes contextual information related to the assignment reason in the intermediate review results; In step S5, the preset rules also include priority rules for audit tasks, wherein the priority is dynamically set according to the risk level identified in the audit intermediate results.

[0015] Furthermore, in step S6, the process of completing the contract review based on feedback from the review terminal includes: Receive confirmation, modification, or rejection instructions and additional annotations from the review terminal for the review task; Based on the operation instructions from all audit terminals, generate the final contract audit report or revised version; In step S6, the contract knowledge base is a hybrid database, including a relational database for storing structured review logs and a vector database for supporting clause-based semantic retrieval.

[0016] The present invention provides an artificial intelligence-based intelligent contract review system, comprising: The file receiving and parsing module is used to receive and parse contract documents to obtain structured text data and contract type; An intelligent analysis engine is used to extract key clauses from the structured text data and to perform compliance review and difference comparison on the key clauses to generate intermediate review results. The task assignment engine is used to automatically assign the generated review tasks to at least one corresponding review terminal based on the contract type and preset rules, and based on the clause content in the intermediate review results. The review tasks include at least the clause content and AI review suggestions related to the review terminal. The knowledge base management module is used to complete the contract review process based on feedback from the review terminal and to store the final review results in a structured manner.

[0017] Furthermore, the task dispatch engine includes: The dynamic analysis unit is used to identify specific clause content from the intermediate audit results and determine the responsible department based on the "clause keyword responsible department" mapping rule; The task generation and routing unit is used to generate independent sub-audit tasks for each responsible department and send them to the corresponding department's audit terminal.

[0018] Furthermore, it also includes a human-computer interaction module, which is used to display the intermediate review results and review tasks to the user in the form of highlights and annotations, and to receive user feedback operations on the review terminal.

[0019] The beneficial effects of this invention are as follows: This invention provides an AI-based intelligent contract review method and system, relating to the fields of computer and artificial intelligence technologies. The method includes: receiving and parsing contract documents to obtain structured text data and contract type; automatically extracting multiple key clauses based on the contract type; conducting compliance review and difference comparison on the key clauses to generate intermediate review results containing risk points and modification suggestions; dynamically analyzing the clause content in the intermediate review results, determining the responsible party for review according to preset rules, and automatically generating and assigning independent review sub-tasks to corresponding review terminals; completing the review process based on feedback from the review terminals, and structurally storing the results in a contract knowledge base. The system includes modules for document parsing, intelligent analysis, task assignment, and knowledge base management. This invention achieves automation, intelligence, and collaboration in contract review, significantly improving review efficiency, risk identification accuracy, and process traceability, and promoting the accumulation and reuse of enterprise contract knowledge. Compared with existing technologies, the AI-based intelligent contract review method and system provided by this invention, by introducing and integrating core technologies such as intelligent document parsing, deep learning extraction, multi-dimensional AI analysis, and content-based intelligent task assignment, achieves the following significant beneficial effects: 1. A revolutionary improvement in review efficiency By automatically parsing contract documents, extracting key clauses, screening for compliance risks, and comparing historical discrepancies using AI models, legal and business personnel are completely freed from tedious and repetitive basic text reading and information retrieval work, allowing them to focus on higher-value core risk assessment and business decisions. In particular, by dynamically generating and automatically assigning review tasks based on the review content, the time-consuming manual coordination, task allocation, and progress tracking of traditional offline or simple electronic workflows are eliminated, achieving seamless integration and parallel processing of cross-departmental review processes. According to actual application calculations, this technical solution can shorten the overall processing cycle of a single contract by more than 60%.

[0020] 2. Enhanced comprehensiveness and accuracy of risk identification. By employing a compliance review model that integrates a large-scale standard clause library and a legal and regulatory database for systematic scanning, and by comparing semantic differences with historical contracts, the system overcomes the oversights, experience limitations, and subjective biases inherent in manual review. The system not only achieves a high recall rate (≥90%) for high-risk clauses but also ensures, through intelligent task assignment, that risks in specific areas (such as financial and technical clauses) are accurately routed to experts in the corresponding fields for review. This combination of machine and human review significantly improves the accuracy of final risk identification (≥95%), effectively constructing a comprehensive contract risk prevention and control system.

[0021] 3. Standardization, transparency, and traceability of the review process The entire offline, fragmented review process has been transformed into an online, standardized digital workflow. Through rule-based intelligent automatic task assignment, real-time tracking of the entire process status, and the recording of all operations and feedback, the "process black box" has been completely eliminated. This makes the responsible person, processing opinion, and dwell time of each review stage clearly traceable, greatly enhancing the management transparency and internal control compliance level of contract review.

[0022] 4. Intelligent accumulation and continuous evolution of enterprise knowledge assets All intermediate results analyzed by AI, final decisions reviewed by humans, modification annotations, and version differences are stored in a structured and interconnected manner in the contract knowledge base. This not only transforms personal experience into shareable, searchable, and analyzable corporate digital assets, but also enables the system to continuously learn from historical review data through machine learning. This provides precise data support and intelligent assistance for subsequent contract review and negotiation clause preparation (such as automatically generating risk point lists and counterparty special clause lists), achieving continuous accumulation and spiral improvement of the organization's contract risk management capabilities.

[0023] 5. The system's agile scalability and ecosystem integration capabilities Adopting a modular and microservices design philosophy, the system allows for easy addition of new contract type recognition models, clause extraction rules, or audit knowledge bases, quickly adapting to business changes. Simultaneously, the system easily integrates deeply with existing enterprise office systems (such as Microsoft 365, Teams, OA, or ERP) through standardized interfaces, enabling single sign-on, task synchronization, and real-time message notifications. This seamlessly embeds intelligent auditing capabilities into daily enterprise operations, lowering the barrier to entry and integration costs.

[0024] The above and other objects, advantages, and features of the present invention will be more fully set forth and demonstrated through the following detailed description of specific embodiments in conjunction with the accompanying drawings. Those skilled in the art, upon referring to the following detailed description and the accompanying drawings, will be able to better understand and realize the above advantages of the present invention. Other objects, features, and advantages of the present invention will become clearer after being described in detail in the detailed description section in conjunction with the accompanying drawings. Attached Figure Description

[0025] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0026] Figure 1 A flowchart for AI-based intelligent contract review process; Figure 2 This is an architecture diagram of an AI-based intelligent contract review system. Figure 3 This is a core business process diagram; Figure 4 This is a diagram illustrating the intelligent extraction of key terms. Figure 5 This is intended to illustrate a comparison of the accuracy of clause extraction. Figure 6 This is intended to represent the purpose of statistics on compliance audit response time. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention. Example 1

[0028] like Figure 1 As shown, Figure 1 This embodiment provides a flowchart for an AI-based intelligent contract review process, characterized by the following steps: S1: Receive contract documents pending review; S2: Parse the contract document to obtain structured text data and contract type; In this embodiment, TextIn intelligent document processing technology is used to perform OCR recognition and text parsing on contract documents. Combined with its powerful layout analysis capabilities, it converts contract files in mainstream formats such as PDF, DOC / DOCX, and MD into structured JSON format text data containing chapter titles, paragraphs, tables, and list tags.

[0029] S3: Based on the contract type, automatically extract multiple key clauses from the structured text data; S4: Conduct compliance review and difference comparison of the key terms, and generate intermediate review results including risk points and modification suggestions; S5: Based on the clause content in the intermediate audit results, perform dynamic analysis, determine at least one auditing party according to the preset configurable business keyword-responsible department mapping rules, and automatically generate an independent auditing subtask for each auditing party based on the clause content and AI auditing suggestions related to the auditing party. The content of the sub-auditing task only includes clause content, risk points and AI auditing suggestions related to the responsibilities of the responsible party, and is accompanied by contextual information related to the assignment reason. Then, each independent auditing subtask is assigned to the corresponding auditing terminal. The mapping rules described in this embodiment are a configurable rule base that records the associations between keywords such as "payment", "amount", "invoice", and "payment period" and "finance department", the associations between "confidentiality", "intellectual property", and "dispute resolution" and "legal department", the associations between "delivery", "logistics", and "transportation" and "logistics department", and the associations between "quality", "acceptance", and "technical standards" and "quality department".

[0030] S6: Based on feedback from the review terminal, complete the contract review process and store the review results in a structured manner in the contract knowledge base.

[0031] In this embodiment, step S2, parsing the contract document, includes: The pre-trained large language model API service is invoked to perform OCR recognition and text parsing on the contract document, converting it into structured JSON format text data containing chapter titles, paragraphs, tables, and list tags.

[0032] In this embodiment, step S3, the automatic extraction of multiple key terms includes: Based on the contract type, a finely tuned large language model is selected from the model pool as the clause extraction model. The model pool described in this embodiment includes clause extraction models finely tuned for different contract types, such as procurement agreements, quality agreements, technical agreements, security agreements, confidentiality agreements, framework agreements, after-sales agreements, logistics agreements, leasing agreements, transparency agreements, and supplier codes of conduct. These models are based on BERT+CRF or RoBERTa+BiLSTM+CRF architectures, trained on historical labeled contract data, and continuously updated. The structured JSON text data is input into the model, and named entity recognition and sequence labeling technologies are used to automatically identify and extract a set of predefined key clause entities and their contents.

[0033] In this embodiment, step S5, the dynamic analysis based on the content of the clauses in the intermediate review results, includes: From the intermediate audit results, identify specific clauses and their contents that present risks or discrepancies; Based on the preset "clause keyword responsible department" mapping rules, the keywords in the specific clause content are analyzed to dynamically determine the responsible department corresponding to the at least one review terminal.

[0034] In this embodiment, automatically assigning the generated review task to at least one corresponding review terminal includes: An independent sub-review task is generated for each identified responsible department. The content of the sub-review task includes only the clauses, risk points and AI review suggestions related to the responsibilities of that department.

[0035] In this embodiment, in step S5, the review task also includes contextual information related to the assignment reason in the intermediate review results.

[0036] In this embodiment, in step S5, the preset rule also includes a priority rule for the review task, and the priority is dynamically set according to the risk level identified in the intermediate review result.

[0037] This embodiment uses a software procurement contract as an example to propose a quantitative comprehensive assessment method to determine the risk level. First, the compliance risk score (Rc) is calculated: the cosine similarity between the clause vector and the "high-risk" reference vector in the compliance knowledge base is calculated, such as "50% prepayment" having a similarity of 0.91. Second, the difference risk score (Rd) is calculated: the semantic difference is obtained by subtracting the cosine similarity with historical contracts / templates from 1, such as a difference of 0.4 for changes in confidentiality period. Third, the basic weight of the clause (Wt) is obtained, preset according to clause type: payment / breach of contract liability 0.9, confidentiality 0.8, performance period 0.5, general clauses 0.3. Fourth, a weighted comprehensive calculation is performed: S = 0.4 × Rc + 0.3 × Rd + 0.3 × Wt (coefficients are configurable). Finally, the risk level is mapped: S ≥ 0.75 is high risk (urgent), 0.50 ≤ S < 0.75 is medium risk (normal), and S < 0.50 is low risk (low). This method ensures that the risk level is objective and repeatable.

[0038] In this embodiment, a preliminary review task set is generated based on the intermediate review results and assigned to the corresponding first review terminal. The specific steps are as follows: S5a: Receive feedback information from the first review terminal, wherein the feedback information includes review comments on the terms and instructions regarding other responsible departments; S5b: Based on the feedback information, dynamically analyze whether the review task of the clause or related clause needs to be routed to a new second review terminal; if so, generate a secondary review task and assign it to the second review terminal. S5c: If feedback from different review terminals results in conflicting review opinions on the same contract clause, a conflict resolution process is automatically triggered. The conflict resolution process includes: retrieving historical resolution records of similar conflict cases from the contract knowledge base, generating a summary report containing the records and the focus of the conflict, and assigning it to a preset arbitration terminal.

[0039] In this embodiment, step S6, which involves completing the contract review process based on feedback from the review terminal, includes: Receive confirmation, modification, or rejection instructions and additional annotations from the review terminal for the review task; Based on the operation instructions from all audit terminals, generate the final contract audit report or revised version.

[0040] In this embodiment, step S4, the difference comparison includes: Based on the contract type and contracting party information, the most relevant standard template contracts or signed historical contracts are automatically selected from the historical contract database as a comparison benchmark.

[0041] In this embodiment, in step S6, the contract knowledge base is a hybrid database, including a relational database for storing structured review logs and a vector database for supporting clause-based semantic retrieval.

[0042] like Figure 2 As shown, Figure 2 This is an architecture diagram of an AI-based intelligent contract review system. The AI-based intelligent contract review system provided in this embodiment is characterized by including a file receiving and parsing module, an intelligent analysis engine, a task assignment engine, and a knowledge base management module. The file receiving and parsing module is used to receive and parse contract documents to obtain structured text data and contract type; The intelligent analysis engine is used to extract key clauses from the structured text data and to perform compliance review and difference comparison on the key clauses to generate intermediate review results. The task assignment engine is used to automatically assign the generated review tasks to at least one corresponding review terminal based on the contract type and preset rules, and based on the clause content in the intermediate review results. The review tasks include at least the clause content and AI review suggestions related to the review terminal. The knowledge base management module is used to complete the contract review process based on feedback from the review terminal and to store the final review results in a structured manner.

[0043] In this embodiment, the task dispatch engine includes: The dynamic analysis unit is used to identify specific clause content from the intermediate audit results and determine the responsible department based on the "clause keyword responsible department" mapping rule; The task generation and routing unit is used to generate independent sub-audit tasks for each responsible department and send them to the corresponding department's audit terminal.

[0044] In this embodiment, a human-computer interaction module is also included, which is used to display the intermediate review results and review tasks to the user in the form of highlights and annotations, and to receive user feedback operations on the review terminal. Example 2

[0045] like Figure 3 As shown, this embodiment details the implementation process of an AI-based intelligent contract review method, using the review process of a software product procurement contract from a multinational technology company as an example. Specific implementation steps: S1: Receive contract documents pending review Legal staff can upload a PDF document (provided by a supplier) for review after single sign-on via the web interface of the system integrated with the company's Microsoft Teams.

[0046] S2: Parse the contract document to obtain structured text data and contract type. The system calls a file parsing service integrated with an OCR engine to convert the PDF document into plain text. The text is then input into a pre-trained contract classification model (e.g., a classifier based on the BERT architecture and fine-tuned on massive amounts of contract data). The model identifies the contract as a "software procurement" contract. Simultaneously, the system further calls a large language model interface to parse the plain text into structured JSON data containing hierarchical tags such as chapters, clauses, and tables.

[0047] S3: Automatically extracts multiple key clauses based on contract type; Based on the identified "software procurement" tag, the system retrieves the corresponding clause extraction model (e.g., using a BERT+CRF architecture) from the model pool, tailored to the procurement contract. This model performs sequence labeling on the structured JSON data, successfully identifying and extracting key clauses and their specific contents, including the contracting parties, the software license subject matter, the total contract amount, payment terms in three installments, the 90-day performance period, breach of contract clauses, confidentiality obligations, and arbitration settlement clauses.

[0048] S4: Conduct compliance reviews and discrepancy comparisons of key clauses, and generate intermediate review results. Compliance Review: The system vectorizes the extracted "Payment Terms" text and performs a similarity search in the enterprise compliance vector knowledge base. It finds that the vector for the "50% prepayment" clause has a similarity of 0.91 (exceeding the threshold of 0.85) to the reference vector for "excessive prepayment ratio" marked as "high risk" in the knowledge base. The system identifies this as a financial risk and generates a modification suggestion including specific legal basis and a recommendation to "modify to no more than 30%".

[0049] Difference Comparison: Based on contract type and signatory information, the system automatically selects the most recent procurement contract signed with the same supplier from the historical contract database as the comparison benchmark. Through semantic similarity calculation, it was found that the "confidentiality period" of the current contract has changed from "permanent" in the historical contract to "3 years after contract termination", with a semantic difference of 0.4 (exceeding the threshold of 0.25), and is therefore identified as a significant difference.

[0050] The automatic filtering logic in this embodiment is performed as follows: When a user uploads a "Production Equipment Procurement Contract" signed with "XX Equipment Manufacturing Co., Ltd.", the system will automatically execute the following filtering steps: Tag matching: Extract the core tags of the current contract: Contract type = "Purchase Contract", Contracting supplier = "XX Equipment Manufacturing Co., Ltd."; In-database search: In the historical contract database, accurately match all historical contracts that simultaneously meet both the conditions of "contract type is purchase contract" and "contracting entity is XX Equipment Manufacturing Co., Ltd."; Time sorting: The matched historical contracts are sorted in descending order by "contract signing date"; Benchmark Lock: The contract with the latest signing date after sorting is automatically selected as the benchmark contract for this difference comparison.

[0051] The above analysis results were summarized and structured to generate an intermediate review result, which clearly listed each risk point / difference, the relevant clause, the risk level, the basis for judgment, and AI modification suggestions.

[0052] S5: Based on the contract type and preset rules, the audit tasks generated from the intermediate audit results are automatically assigned to the corresponding audit terminals. Dynamic Analysis and Task Generation: The system reads intermediate audit results. Regarding the risk of "excessive prepayment ratio," the analysis reveals keywords containing "payment" and "amount." Based on the preset "term keyword responsibility department" mapping rules, it determines that the task should be assigned to the Finance Department. Regarding the difference in "shortened confidentiality period," the keyword contains "confidentiality," determining that the task should be assigned to the Legal Department.

[0053] Automatic task assignment: The system generates separate sub-review tasks for the finance and legal departments. The task sent to the finance department only includes the original text of the "Payment Terms" clause, AI-identified risk details, and modification suggestions. The task sent to the legal department includes the current content of the "Confidentiality Clause," historical comparison content, and discrepancy analysis. Tasks are automatically pushed to the to-do lists of designated reviewers in both departments via system messages and integrated emails.

[0054] In this embodiment, the assignment technique employs a dynamic routing system based on multi-round feedback and an intelligent conflict resolution mechanism, as detailed below: 1. Multi-round dynamic routing: The system not only assigns tasks initially, but also dynamically generates new or adjusts subsequent review sub-tasks based on the feedback from the first review department (such as "This clause involves technical specifications and needs to be confirmed by the technical department"), and routes them to the newly determined department.

[0055] 2. Conflict Detection and Resolution: When conflicting opinions arise between different departments (e.g., the legal department deems it compliant, while the finance department considers it risky), the system can automatically identify the conflict type and initiate the corresponding resolution process. For example: The system automatically retrieves historical review decisions and final adoption plans for similar clauses from the "Contract Knowledge Base" and pushes them to all relevant parties as reference suggestions.

[0056] The system summarizes the points of conflict and the reasons given by both parties, automatically generates a summary of the key issues, and prioritizes them for routing to higher-level decision-makers (such as the general counsel) or initiating an online collaborative review meeting.

[0057] S6: Based on feedback from the review terminal, complete the contract review process and store the review results in a structured format in the contract knowledge base. The finance department auditor viewed the task on their terminal, adopted the AI ​​suggestion, and provided online feedback on the proposed modification, "adjust the prepayment to 30%".

[0058] After evaluation, the legal department reviewer deemed the 3-year confidentiality period reasonable and confirmed "no objection" online.

[0059] Once all departments have completed their reviews, the system automatically compiles and generates a final review report. The report, original contract, all intermediate analysis results (JSON structured clauses, intermediate review results), manual review comments, and operation logs are all stored in the company's contract knowledge base. The full text of the clauses and other textual data are stored in a vector database to support semantic retrieval, while process data is stored in a relational database to support efficient querying. Example 3

[0060] This embodiment further illustrates the method with specific illustrations and implementation processes, and specifically includes the following steps: S1: Contract Document Upload Users log in to the system through a web application interface that supports Single Sign-On (SSO) authentication and upload contract documents awaiting review. The system supports multiple file formats, including `doc`, `docx`, `pdf`, `txt`, and `xlsx`.

[0061] S2: File parsing Call the API service of a pre-trained large language model (such as a hybrid model combining OCR and NLP, such as LayoutLMv3 or PaddleOCR+ERNIE) to parse the uploaded contract document: If it is a scanned PDF document, perform OCR recognition first; Convert the document into structured JSON format data, including chapter titles, paragraphs, tables, lists, and other tags, to facilitate subsequent processing.

[0062] In this embodiment, the PDF contract to be reviewed is first subjected to OCR recognition to extract plain text. Then, a request message is constructed to call the Large Language Model API, instructing that the text be parsed into structured JSON, distinguishing between chapter titles, paragraphs, tables, and lists, while maintaining the original text order. The API parameters are set to output format JSON, with a temperature of 0.1 to ensure stability, and return a JSON string. An example of the parsed JSON data pattern is as follows: The root node contains the document type (e.g., "Software Procurement Contract") and metadata (title, total number of pages). Each element under the "Chapter" array represents a contract clause, containing a chapter title, a paragraph array, a table array (table header + data rows), and a list array (ordered / unordered and item content). For example, the payment clause: title "Article 5 Payment Conditions", paragraph "50% Prepayment", payment schedule table (node, percentage, date), and invoice requirement list (VAT invoice, designated account). This pattern preserves the document hierarchy, facilitating the automatic extraction of key clauses later.

[0063] S3: Terms Extraction like Figure 4 As shown, Figure 4 This diagram illustrates the intelligent extraction of key clauses. Based on the contract type (e.g., purchase contract, lease contract), a corresponding fine-tuned large language model (e.g., BERTCRF or RoBERTa+BiLSTMCRF) is selected from the model pool for clause extraction. Named entity recognition and sequence labeling technologies are used to automatically extract the following key clauses: The contract should include the following terms: contract amount, payment terms, performance period, liability for breach of contract, confidentiality clauses, and dispute resolution methods. All models in the pool are tagged according to two dimensions: "contract type + business scenario," for example: A dedicated model for procurement contracts: tagged with contract type = procurement contract, scenario = procurement of production materials / equipment, adapted to the clause structure and terminology of manufacturing procurement contracts; In this embodiment, the model pool centrally manages large language models fine-tuned for different contract types. Its management mechanism includes: model registration (binding contract type, version number, and performance metrics), version control (retaining multiple versions to support rollback and A / B testing), dynamic updates (incremental fine-tuning after accumulating new data), and health monitoring (periodic testing, automatic rollback when performance degrades). The contract type selection model employs a three-level matching algorithm: First, exact matching, directly searching for the corresponding model using the contract type label; second, hierarchical matching, if no exact match is found, extracting the parent category (e.g., "procurement contract") and selecting the model of that category; third, semantic similarity matching, if the first two steps fail, vectorizing the contract text and calculating the cosine similarity with the distribution center vectors of each model's training data, selecting the model with the highest score exceeding a threshold (e.g., 0.7); if no match is still found, a general basic model is called and recorded for future expansion. The matching process is completed in milliseconds, ensuring real-time performance.

[0064] S4: Compliance Audit Construct a vector knowledge base, including internal company standard clauses and external legal and regulatory provisions. Use a text embedding model (such as SentenceBERT) to vectorize the extracted clauses and perform semantic similarity retrieval within the knowledge base. If a clause's similarity to clauses marked as "high-risk" or "non-compliant" exceeds a preset threshold, the system automatically generates a risk warning and associates it with legal basis and modification suggestions, specifically in the following manner: 1. Data preprocessing: External regulations (such as the Contract Law section of the Civil Code and industry regulatory regulations) and internal standard clauses / prohibitory clauses of enterprises are broken down by clause granularity and labeled with metadata such as risk level (high / medium / low / compliant), applicable contract type, and legal basis.

[0065] 2. Vectorization encoding: The SentenceBERT model is used to convert each clause text into a 768-dimensional semantic vector, preserving the mapping relationship between the original clause text, metadata and vector.

[0066] 3. Index Building: The FAISS vector indexing engine is used to build an IVF index on the vector library, enabling millisecond-level semantic retrieval. At the same time, tag-based classification indexes are created according to contract type and risk level to improve retrieval efficiency.

[0067] Example: Encode the internal prohibition clause "no permanent confidentiality obligation shall be agreed upon" into a vector, mark it as "high risk", and store it in the index; encode the statutory clauses on liability for breach of contract in the Civil Code and mark them as "compliance basis".

[0068] The construction of the compliance vector knowledge base in this embodiment involves: collecting external laws and regulations and internal corporate standard clauses, cleaning and segmenting the text (laws are split into clauses and clauses, and internal clauses are segmented by type), converting them into vectors using an embedding model (such as Sentence-BERT), storing them in a vector database, and associating them with metadata such as source, risk label, and clause type. Encoding uses hashing or UUID to generate unique identifiers, and indexing uses near nearest neighbor (such as HNSW) to achieve efficient similarity retrieval. Similarity threshold settings: 0.85 (compliance risk) is based on the 90th percentile of the similarity distribution between historical high-risk clauses and reference vectors, ensuring recall and reducing false positives; the adjustable range is 0.80–0.90. 0.25 (difference comparison) corresponds to a cosine similarity of 0.75. Statistics show that 95% of clauses with a difference ≥ 0.25 are judged as significantly different; the adjustable range is 0.20–0.35, and for sensitive businesses, it can be reduced to 0.20.

[0069] S5: Difference Comparison Based on the contract type and signatory information, the system intelligently matches the most relevant standard templates or historical contracts from the historical contract database. A semantic similarity algorithm (such as BERTScore or edit distance algorithm) is used to compare the corresponding clauses of the current contract with those of the reference contract line by line. If the difference exceeds a set threshold, the system identifies it as a significant difference and generates a comparison report. Review tasks are dynamically assigned to the corresponding departments (such as finance and legal departments) based on the content of the clauses. The intelligent matching in this embodiment is as follows: When uploading an "Equipment Purchase Contract" signed with "XX Equipment Company", the system first extracts tags such as contract type = purchase contract and contracting party = XX Equipment Company, matches contracts of the same type and supplier in the historical database, sorts them by signing time + semantic similarity, and takes the latest and most relevant contract / template as the comparison benchmark.

[0070] Threshold determination: Based on the manufacturing contract annotation dataset, the threshold was determined by optimizing the F1 score and false alarm rate: the difference threshold was set to 0.25 (≥0.25 is considered a significant difference), which takes into account both false negatives and false alarms and is suitable for multiple scenarios such as procurement and leasing.

[0071] Dynamic assignment rules: Automatically match responsible departments according to the type of terms: Amount / Payment Terms → Finance Department, Confidentiality / Breach of Contract Liability → Legal Department, Performance / Delivery → Procurement Department. Supports custom department mapping and multi-department collaborative assignment.

[0072] S6: Knowledge Accumulation After approval, the system stores the final contract draft along with the structured clauses, review reports, annotations, etc., extracted in the preceding steps in both a relational database and a vector database, forming a queryable and comparable contract knowledge base. The vector database supports semantic-based similar contract retrieval. In this embodiment, the associated storage process is as follows: Taking a "XX Equipment Procurement Contract" as an example, the system stores the associated data according to the following logic: 1. Relational Database: Using the contract ID as the unique primary key, it stores basic contract information (contracting parties, amount, date), extracted structured clauses (payment terms, confidentiality period, etc.), review reports, departmental annotations, approval processes, and other structured data, supporting precise conditional queries. 2. Vector Database: Using SentenceBERT, the full text of the contract and key clauses are vectorized, bound to the same contract ID, and semantic vectors are stored, supporting semantic retrieval of similar contracts. 3. Bidirectional Association: The contract ID enables linkage between the two databases, allowing for one-click tracing of the original contract text from structured data, and also allowing jumps from similar contract search results to complete structured information.

[0073] like Figure 5 As shown, Figure 5 This is a schematic diagram of document parsing and structuring; the AI-based intelligent contract review system provided in this embodiment specifically includes the following modules: 1. File parsing module Employing pre-trained models based on the Transformer architecture (such as LayoutLMv3), it supports multi-format document parsing and OCR recognition, outputting structured JSON data.

[0074] 2. Clause Extraction Module By employing a multi-task learning framework and combining named entity recognition and sequence labeling technologies, high-precision clause extraction is achieved by dynamically calling the corresponding fine-tuning model (such as BERTCRF) based on the contract type.

[0075] 3. Compliance Audit Module We construct an intelligent review engine based on vector retrieval, use SentenceBERT for semantic encoding, and combine it with a rule-based reasoning engine (such as Drools) to achieve hybrid intelligent review.

[0076] 4. Difference Comparison Module Implement a semantic-based text comparison algorithm, supporting automatic screening of historical contracts and clause-level difference analysis, and support dynamic routing and priority scheduling for review tasks.

[0077] 5. Knowledge Base Module It adopts a dual storage architecture of "relational database + vector database", supports structured storage and semantic retrieval, and realizes the continuous accumulation and reuse of contract knowledge.

[0078] like Figure 5 and Figure 6 As shown, Figure 5 This is intended to illustrate a comparison of the accuracy of clause extraction. Figure 6 This is intended to illustrate the purpose of statistics on compliance audit response times. Figure 5 The text uses bar charts or line graphs to compare and illustrate the differences in clause extraction accuracy between the proposed method (based on a fine-tuned large language model, such as BERT+CRF) and traditional rule-based or keyword-matching methods. The horizontal axis typically represents different contract types (e.g., purchase contracts, confidentiality agreements, lease contracts) or different clause types (e.g., payment terms, liability for breach of contract, confidentiality clauses). The vertical axis represents the accuracy (%) of clause extraction.

[0079] Compared to traditional methods that rely on keyword or fixed template extraction, which suffer from low accuracy, especially when dealing with complex sentences or unstructured text, this method utilizes deep learning models (such as BERT+CRF) for sequence labeling and named entity recognition. This approach better understands contextual semantics, resulting in a significant improvement in accuracy.

[0080] Purpose: To demonstrate that this method has higher semantic understanding capabilities and extraction accuracy in the automatic extraction of contract terms, thereby providing a more reliable data foundation for subsequent compliance audits and discrepancy comparisons.

[0081] in, Figure 6 The time required for contract compliance review is compared between this method and purely manual review using bar charts or line graphs. The horizontal axis typically represents contract samples of different types, lengths, or complexities. The vertical axis represents the review response time (usually in minutes or seconds). Comparison: In traditional purely manual review, reviewers need to read each clause, search for regulations, and compare templates, which is time-consuming. This method automatically completes clause extraction, vector retrieval, risk identification, and difference comparison, significantly reducing response time.

[0082] Purpose: To demonstrate that this method can significantly improve efficiency in the contract review process, freeing reviewers from a large amount of repetitive and basic work, allowing them to focus on high-value judgments and decisions.

[0083] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A contract intelligent review method based on artificial intelligence, characterized in that, Includes the following steps: S1: Receive contract documents pending review; S2: Parse the contract document to obtain structured text data and contract type; S3: Based on the contract type, extract multiple key clauses from the structured text data; S4: Conduct compliance review and difference comparison of the key terms, and generate intermediate review results including risk points and modification suggestions; S5: Based on the clause content in the intermediate audit results, perform dynamic analysis, determine at least one auditing party according to the preset configurable business keyword-responsible department mapping rules, and automatically generate an independent auditing subtask for each auditing party based on the clause content related to the auditing party and AI audit suggestions. The content of the sub-audit task includes clause content, risk points and AI audit suggestions related to the responsibilities of the party, and is accompanied by contextual information related to the assignment reason. Then, each independent auditing subtask is assigned to the corresponding auditing terminal. S6: Based on feedback from the review terminal, complete the contract review process and store the review results in a structured manner in the contract knowledge base.

2. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S2, parsing the contract document includes: The pre-trained large language model API service is invoked to perform OCR recognition and text parsing on the contract document, converting it into structured JSON format text data containing chapter titles, paragraphs, tables, and list tags.

3. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S3, the automatic extraction of multiple key terms includes: Based on the contract type, a finely tuned large language model is selected from the model pool as the clause extraction model. The structured JSON text data is input into the model, and named entity recognition and sequence labeling technologies are used to automatically identify and extract a set of predefined key clause entities and their contents.

4. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S4, the difference comparison includes: Based on the contract type and contracting party information, relevant standard template contracts or signed historical contracts are automatically selected from the historical contract database as a comparison benchmark.

5. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S5, the dynamic analysis based on the content of the clauses in the intermediate audit results includes: From the intermediate audit results, identify specific clauses and their contents that present risks or discrepancies; Based on the preset keyword responsibility department mapping rules, the keywords in the specific clause content are analyzed to dynamically determine the responsible department corresponding to the at least one review terminal; The step of automatically assigning the generated review tasks to at least one corresponding review terminal includes: An independent sub-review task is generated for each identified responsible department. The content of the sub-review task includes only the clauses, risk points and AI review suggestions related to the responsibilities of that department.

6. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S5, the review task also includes contextual information related to the assignment reason in the intermediate review results; In step S5, the preset rules also include priority rules for audit tasks, wherein the priority is dynamically set according to the risk level identified in the audit intermediate results.

7. The intelligent contract review method based on artificial intelligence according to claim 1, characterized in that, In step S6, the process of completing the contract review based on feedback from the review terminal includes: Receive confirmation, modification, or rejection instructions and additional annotations from the review terminal for the review task; Based on the operation instructions from all audit terminals, generate the final contract audit report or revised version; In step S6, the contract knowledge base is a hybrid database, including a relational database for storing structured review logs and a vector database for supporting clause-based semantic retrieval.

8. A contract intelligent review system based on artificial intelligence, characterized in that, include: The file receiving and parsing module is used to receive and parse contract documents to obtain structured text data and contract type; An intelligent analysis engine is used to extract key clauses from the structured text data and to perform compliance review and difference comparison on the key clauses to generate intermediate review results. The task assignment engine is used to automatically assign the generated review tasks to at least one corresponding review terminal based on the contract type and preset rules, and based on the clause content in the intermediate review results. The review tasks include at least the clause content and AI review suggestions related to the review terminal. The knowledge base management module is used to complete the contract review process based on feedback from the review terminal and to store the final review results in a structured manner.

9. The AI-based intelligent contract review system according to claim 8, characterized in that, The task dispatch engine includes: The dynamic analysis unit is used to identify specific clause content from the intermediate audit results and determine the responsible department based on the clause keyword responsible department mapping rules; The task generation and routing unit is used to generate independent sub-audit tasks for each responsible department and send them to the corresponding department's audit terminal.

10. The AI-based intelligent contract review system according to claim 8, characterized in that, It also includes a human-computer interaction module, which is used to display the intermediate review results and review tasks to the user in the form of highlights and annotations, and to receive user feedback operations on the review terminal.

Images