A contract lifecycle management method and system

Abstract

This invention relates to the field of contract management technology and discloses a method and system for contract lifecycle management. The method includes: constructing a multi-level knowledge graph containing contracts, clauses, and legal entities; acquiring external regulations in real time and generating structured risk vectors; using graph convolutional networks to perform risk propagation calculations within the graph, quantifying the impact of new regulations on existing contracts; and generating a targeted early warning list sorted by risk index. The system includes modules for contract data processing, knowledge graph construction, regulatory intelligence acquisition, risk propagation calculation, and early warning presentation. This invention achieves real-time, accurate, and efficient identification and closed-loop optimization of compliance risks through semantic understanding and graph neural network technology.

Description

Technical Field

[0001] This invention belongs to the field of contract management technology, specifically relating to a contract lifecycle management method and system. Background Technology

[0002] As enterprises accelerate their digital transformation, contracts, as the core carrier of business activities, have become a crucial aspect of corporate compliance management throughout their entire lifecycle. Modern enterprises typically rely on electronic contract management systems to centrally control the drafting, approval, signing, performance, and archiving of contracts. These systems generally integrate rule engines and keyword matching mechanisms to identify potential compliance risk clauses in contract texts.

[0003] However, the existing technical architecture has significant shortcomings: laws, regulations, and industry regulatory requirements are constantly being updated dynamically, while the system's built-in rule base often fails to cover newly issued compliance points in a timely manner due to version iteration delays or manual maintenance delays; traditional matching strategies based on static keywords lack semantic understanding capabilities, making it difficult to identify variations in expression or implicit violations, resulting in the failure to detect a large number of new risk clauses. Once high-risk contracts enter the performance stage without warning, companies will face serious consequences such as regulatory penalties, litigation disputes, or reputational damage.

[0004] Dynamic perception and accurate identification of compliance risks in contract lifecycle management has become a crucial technological area requiring breakthroughs. This approach aims to achieve rapid response to regulatory changes through intelligent means, and to efficiently and accurately screen existing and new contracts for risks without disrupting business processes. The core challenge lies in ensuring comprehensive detection coverage while avoiding the need for a full rescan of massive historical contracts triggered by rule updates, which would otherwise waste computing resources and cause system response delays.

[0005] Existing technologies typically employ a crude "rule update—full database rerun" approach, re-executing risk matching logic for all contract texts whenever the regulatory database changes. This method not only consumes significant storage and computing resources but also has a long scanning cycle, making it difficult to meet enterprises' needs for real-time compliance monitoring. Furthermore, due to the lack of modeling capabilities for the correlation between old and new regulations, the system cannot distinguish which historical contracts are actually affected by the new regulations, resulting in a large number of unrelated contracts being invalidated. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides a contract lifecycle management method. This method continuously acquires newly effective legal and regulatory texts from external data sources and transforms them into structured risk feature vectors using a legal domain-specific language model. Simultaneously, it performs clause-level segmentation and semantic vectorization of all internal contract texts, constructing a unified, multi-level contract clause knowledge graph.

[0007] When new regulatory risk feature vectors are generated, this invention utilizes graph neural networks to perform risk propagation calculations within a knowledge graph, quantifying the potential impact of new regulations on each existing contract clause, thereby generating a list of contracts to be reviewed, sorted by risk index. This method abandons the traditional full-scan model, achieving rapid, accurate, and targeted identification and early warning of new compliance risks, significantly improving the efficiency and accuracy of risk response in enterprise contract management.

[0008] This invention provides a contract lifecycle management method, which includes the following steps:

[0009] Constructing a knowledge graph of contract terms, the steps of which include:

[0010] Obtain the full volume of contract documents within the enterprise; preprocess the full volume of contract documents by splitting each contract's unstructured text into independent clause-level text units;

[0011] A pre-trained language model for the legal domain based on a transformer architecture is used to semantically encode each clause-level text unit, generating a corresponding clause semantic vector.

[0012] Each contract, its constituent clause-level text units, and legal entities extracted from the text are treated as nodes. The attribution, reference, and association relationships between them are treated as edges to construct the contract clause knowledge graph. The clause semantic vectors are stored as attributes of the corresponding clause nodes.

[0013] Real-time acquisition and processing of external regulatory data, the steps of which include:

[0014] By using web crawler modules or application programming interfaces, the system continuously monitors and obtains the original text data of newly effective or newly released regulations from pre-set official legislative information release platforms, industry regulatory agency websites, and commercial databases; the regulations include: laws, departmental rules, and judicial interpretations.

[0015] The acquired raw text data is parsed to extract mandatory provisions, prohibitions, obligations, and penalties, forming a structured text of key legal points. A pre-trained language model in the legal domain is used to semantically encode the structured text of key legal points, generating one or more legal risk vectors.

[0016] Conducting a risk communication and impact assessment, the steps of which include:

[0017] The newly generated regulatory risk vector is injected into the contract terms knowledge graph as a risk source node; a graph convolutional network algorithm is applied to perform multiple rounds of iterative calculations on the contract terms knowledge graph.

[0018] In each iteration, each clause node updates its own risk score based on the cosine similarity between its own clause semantic vector and the regulatory risk vector, as well as the risk status of its neighboring nodes in the graph; the calculation is iterated until the risk scores of all nodes converge, thus obtaining the final risk score of each contract clause node for the new regulations.

[0019] Generating a targeted risk warning list, the steps of generating the targeted risk warning list include:

[0020] Set a risk score threshold; filter out all clause nodes in the contract clause knowledge graph whose final risk score is higher than the risk score threshold;

[0021] Based on the attribution relationship of the aforementioned clause nodes in the knowledge graph, trace back to the contract subject to which they belong;

[0022] All contracts containing high-risk clauses are aggregated and sorted in descending order according to their cumulative risk scores at the contract level to generate a targeted risk warning list. This targeted risk warning list is then pushed to the user interface of the contract management platform, highlighting the specific risk clauses, related legal points, and quantified risk scores.

[0023] As one embodiment of the present invention, the legal domain pre-trained language model is based on a general transformer model architecture and is obtained through secondary pre-training using a legal professional corpus of no less than 100 gigabytes, including legal documents, judgment cases, academic papers, and contract templates. Specifically, the semantic encoding process involves taking clause-level text units or key legal points as input, and taking the output vector of the corresponding classification label from the last hidden layer of the model as a 768-dimensional floating-point vector, which is the clause semantic vector or legal risk vector.

[0024] Furthermore, in the step of constructing the contract clause knowledge graph, the extracted legal entity nodes include the contract signatories, the subject matter of the contract, the place of performance, the dispute resolution method, the applicable law, and the specific amount and date. The edges not only include the parent-child relationship between contracts and clauses, but also the cross-reference relationship between clauses, as well as the attribute relationship between clauses and legal entities, thus forming a heterogeneous graph structure.

[0025] As one embodiment of the present invention, the iterative update rule of the graph convolutional network algorithm is as follows: the next state feature of each node is calculated by a nonlinear activation function based on the weighted average of its own current state feature and the current state features of all its neighboring nodes. The weights of the weighted average are determined by the structural connections and semantic similarity between nodes, ensuring that risk propagation occurs not only along the contractual structure path but also along the semantic association path. The convergence condition is that, in two consecutive iterations, the change in the L2 norm of the risk score vectors of all nodes in the graph is less than a preset convergence accuracy value.

[0026] Furthermore, after generating the targeted risk warning list, the method also includes a closed-loop feedback and model optimization step. This step specifically includes: human legal experts reviewing the warning items in the risk warning list and marking them as true positives or false positives; the system collecting all labeled data and using it as training samples to fine-tune the weight parameters in the graph convolutional network algorithm through supervised learning, or dynamically and adaptively adjusting the risk score threshold, thereby continuously improving the accuracy of the risk assessment model.

[0027] This invention provides a contract lifecycle management system, which includes:

[0028] The contract data processing module is configured to acquire all contract document data within the enterprise, perform clause-level text unit splitting, and use a pre-trained language model based on a transformer architecture to semantically encode each clause-level text unit, generating clause semantic vectors.

[0029] The knowledge graph construction and storage module is configured to receive the semantic vectors and structured data of the contract terms generated by the contract data processing module, and construct and persistently store a knowledge graph of contract terms, using contracts, terms, and legal entities as nodes and their inherent relationships as edges. The knowledge graph construction and storage module is implemented by a graph database management system, which provides efficient query and traversal functions for graph-structured data.

[0030] The external regulatory intelligence acquisition module is configured to acquire newly effective regulatory texts in real time from multiple preset online legal information sources through its built-in network interface and data parser, extract key regulatory points, and call the pre-trained language model in the legal field to encode them into regulatory risk vectors.

[0031] The risk propagation computation engine, whose core is a graph convolutional network processor, is configured to perform iterative computation on the contract terms knowledge graph when a new regulatory risk vector is received, and to calculate a quantified risk score for each term node in the graph.

[0032] The warning and presentation module is configured to filter out all contracts that exceed a preset risk threshold based on the risk score output by the risk propagation calculation engine, generate a warning list sorted by risk level, and present the list and associated risk details, including specific terms, corresponding regulations and risk scores, to the user through a graphical user interface.

[0033] In one embodiment of the present invention, the contract data processing module includes an optical character recognition unit and a layout analysis unit. The optical character recognition unit is used to convert scanned contract images into editable text files. The layout analysis unit uses a document layout recognition algorithm to automatically identify chapter titles, clause numbers, and body text in the contract text, thereby achieving accurate clause-level text unit segmentation.

[0034] Furthermore, the knowledge graph construction and storage module employs a graph database management system whose underlying storage architecture is optimized for sparse adjacency matrices and supports large-scale parallel graph computation. The system establishes an index for each node and edge to ensure high-speed access to adjacent nodes during risk propagation computation.

[0035] In one embodiment of the present invention, the hardware implementation of the risk propagation computation engine includes one or more graphics processing units (GPUs). The computational task of the graph convolutional network is compiled into computational kernels that are executed in parallel on the GPUs, thereby accelerating the iterative convergence process of the risk score using its massively parallel computing architecture.

[0036] Furthermore, the warning and presentation module also integrates a workflow engine. After a risk warning list is generated, the workflow engine automatically creates review tasks according to preset rules and assigns the tasks to the corresponding legal personnel or business department heads. At the same time, it records all nodes and operation logs of the entire risk handling process, forming a traceable compliance audit record.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. It achieves immediacy and foresight in risk identification. By monitoring external regulatory sources in real time and performing automated semantic analysis, this invention can initiate impact assessments as soon as new regulations take effect or even when drafts are released, completely eliminating the lag of several days or even weeks caused by relying on manual updates to the rule base, thus buying valuable response time for enterprises.

[0039] 2. Improved accuracy and depth of risk discovery. This invention employs semantic vector representation and knowledge graph technology based on deep learning, enabling it to understand the deeper meaning of legal texts and identify compliance risks hidden in complex clause logic that cannot be discovered through traditional keyword matching, significantly reducing the risk underreporting rate.

[0040] 3. Significantly improves the efficiency of reviewing existing contracts. This invention, through a risk propagation algorithm, precisely targets a small number of high-risk contracts affected by new regulations, transforming full-scale scanning into targeted review. This reduces the required computing resources and manpower costs by several orders of magnitude, making routine compliance monitoring of massive amounts of existing contracts possible.

[0041] 4. A dynamically evolving risk management system has been constructed. By introducing a closed-loop feedback mechanism for manual review, the system can learn from each review result and continuously optimize its risk assessment model, thereby improving the accuracy of risk identification over time and forming an adaptive and self-optimizing intelligent contract risk management system. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the overall technical architecture of a contract lifecycle management method and system proposed in this invention;

[0043] Figure 2 This is a schematic diagram of the core principle framework of risk propagation calculation based on legal language models and knowledge graphs in this invention;

[0044] Figure 3 This is a logical flowchart of the knowledge graph construction and semantic vectorization processing of contract terms in this invention;

[0045] Figure 4 This is a flowchart illustrating the logical process framework for real-time acquisition, parsing, and risk vector generation of external regulatory data in this invention.

[0046] Figure 5 This is a logical flowchart of the targeted risk warning list generation and closed-loop feedback optimization in this invention;

[0047] Figure 6 This is a schematic diagram of the multi-level interaction relationships and data flow between the functional modules of the contract management system in this invention. Detailed Implementation

[0048] Please refer to Figures 1 to 6 This invention provides a contract lifecycle management method and system, aiming to solve the problems of missed compliance risks and delayed response caused by lagging updates to the compliance rule base, reliance on keyword matching for risk identification, and low efficiency of full rescanning of existing contracts. The method constructs a multi-level contract clause knowledge graph, combining a pre-trained language model in the legal field with graph neural network technology to achieve semantic-level understanding and targeted risk propagation of newly released regulations. This allows for the accurate identification of high-risk clauses affected by new regulations within a massive amount of existing contracts, and the generation of an actionable risk warning list.

[0049] The contract lifecycle management method includes the following steps: S1, constructing a knowledge graph of contract terms; S2, acquiring and processing external regulatory data in real time; S3, performing risk propagation and impact assessment; S4, generating a targeted risk warning list.

[0050] In step S1, the specific process of constructing the contract terms knowledge graph is as follows: First, acquire all contract document data in batches from the enterprise's internal contract management system, document storage server, or electronic archive. This document data covers all historically signed valid contracts, archived contracts, and contracts currently being performed, in formats including but not limited to editable text files, portable document format files, and scanned image files.

[0051] For scanned contract documents, the system invokes the Optical Character Recognition (OCR) unit for text reconstruction. This unit employs a character recognition model based on a deep convolutional neural network, supporting high-precision text extraction under multiple languages, fonts, and complex layouts, with an accuracy rate of no less than 98.5%. Subsequently, the system activates the layout analysis unit. Based on a document structure recognition algorithm, this unit automatically detects chapter titles, clause numbers, paragraph start positions, and nested hierarchical relationships in the contract text, accurately splitting the entire unstructured contract text into independent clause-level text units. Each clause-level text unit is bounded by complete semantic logic, such as independent clauses like "liability for breach of contract," "confidentiality obligations," and "dispute resolution," ensuring the completeness and accuracy of subsequent semantic encoding.

[0052] After completing the clause-level segmentation, the system calls a pre-trained legal language model based on a transformer architecture to semantically encode each clause-level text unit. This language model is obtained through secondary pre-training on a general transformer model using a legal professional corpus of no less than 100 gigabytes in size. The corpus includes judgments issued by the Supreme People's Court, normative documents issued by administrative organs at all levels, contract templates published by authoritative legal publishers, articles in core legal journals, and Chinese translations of international treaties.

[0053] The semantic encoding process is as follows: individual clause-level text units are fed into the model as input sequences. During the forward propagation process, the model calculates the contextual representation of each word. Finally, the output of the last hidden layer corresponding to the classification label is taken to form a 768-dimensional floating-point vector, which is the clause semantic vector. This vector represents the legal meaning, obligatory attributes, binding strength, and relationship with other legal concepts of the clause in the high-dimensional semantic space.

[0054] After generating semantic vectors for all clauses, the system performs a knowledge graph construction operation. The graph nodes fall into three categories: contract nodes, clause nodes, and legal entity nodes. A contract node represents a complete contract document, with attributes including contract number, signing date, effective period, contract type, and signatory information. A clause node corresponds to each clause-level text unit extracted earlier; its core attribute is the generated clause semantic vector, and it also includes clause number, chapter / section, original text content, and its positional offset within the contract.

[0055] Legal entity nodes are key legal elements automatically extracted from the clause text using named entity recognition technology, including but not limited to the names of the contracting parties, description of the subject matter of the contract, place of performance, dispute resolution institution, applicable legal provisions, monetary amount, and date and time. There are three types of edges: the first is attribution edges, which connect a contract node to all its subordinate clause nodes, indicating an inclusion relationship; the second is reference edges, which connect clause nodes with cross-reference relationships, such as "as stated in Clause 5.2," which establishes a directed reference edge between the Clause 5.2 node and the current clause node; the third is attribute edges, which connect a clause node to the legal entity node it mentions, such as a clause mentioning "Beijing Arbitration Commission," which establishes an attribute edge between that clause node and the "Beijing Arbitration Commission" entity node.

[0056] The resulting contract terms knowledge graph is a heterogeneous graph structure, with nodes possessing rich semantic attributes and edges exhibiting clear semantic and structural relationships. This graph is persistently stored in a graph database management system. This system employs a columnar storage engine optimized for sparse adjacency matrix access patterns and establishes composite indexes for all node and edge types, supporting millisecond-level adjacency queries and subgraph traversal operations.

[0057] In step S2, the specific process of acquiring and processing external regulatory data in real time is as follows: The system continuously monitors and acquires the original text data of newly effective or newly released regulations from preset official legislative information release platforms, industry regulatory agency websites, and commercial databases through a built-in web crawler module or standard application programming interface; the regulations include: laws, departmental rules, and judicial interpretations.

[0058] When a new or updated regulatory text is detected, the system immediately downloads its original text data and initiates a parsing process. The regulations include laws, departmental rules, and judicial interpretations. The parsing process first performs structured analysis on the text, identifying mandatory provisions, prohibitions, obligations, and penalties. For example, the system identifies the statement "Operators shall not collect personal information unrelated to the services they provide" as a prohibition; and the statement "Network operators shall take technical measures and other necessary measures to ensure the security of the personal information they collect" as an obligation. All identified regulatory points are extracted individually to form a structured set of regulatory point texts.

[0059] Subsequently, the system invokes the same pre-trained language model for the legal domain as in step S1 to semantically encode each regulatory point text. The encoding method is completely consistent with the generation of clause semantic vectors, i.e., inputting the regulatory point text, taking the output of the last hidden layer at the classification marker position, and generating a 768-dimensional regulatory risk vector. Each regulatory risk vector accurately represents the legal effect, binding object, behavioral boundaries, and potential consequences of violation of the regulatory point in the semantic space. If a new regulation contains multiple independent compliance requirements, multiple corresponding regulatory risk vectors are generated, each representing a different risk dimension.

[0060] In step S3, the specific process of performing risk propagation and impact assessment is as follows: The system injects one or more newly generated regulatory risk vectors into the constructed contract clause knowledge graph as temporary risk source nodes. These risk source nodes do not establish permanent connections with existing nodes in the graph; they only serve as initial activation signals in this risk assessment calculation. Subsequently, the system starts the graph convolutional network algorithm to perform multiple rounds of iterative calculations across the entire knowledge graph. In each iteration, each clause node updates its risk score according to the following rules:

[0061] Let the first In rounds of iteration, clause nodes The risk score is Its set of adjacent nodes is Then the first Risk score of the round Calculated by the following formula:

[0062] ;

[0063] in, To modify the activation function of the linear unit; For clause nodes The semantic vector of the terms; The current regulatory risk vector; The cosine similarity function is used to measure the degree of semantic matching between the clause's semantics and the regulatory risks. The self-loop weight is fixed at 0.6; For nodes With adjacent nodes The propagation weight between them is calculated using the following formula:

[0064] ;

[0065] in For nodes and Indicator variables indicating whether there is a structural connection between them. The cosine similarity is the semantic similarity between the two nodes. The structure-semantic balance coefficient is set to 0.7. This weighting design ensures that risk propagation occurs both along the structural path within the contract (such as parent-child clauses or reference clauses) and along the semantically similar clause path (such as clauses in different contracts that both involve "data security"), thereby achieving global risk diffusion across contracts and structures.

[0066] Iterative calculations continue until the convergence condition is met: in two consecutive iterations, the change in the L2 norm of the risk score vectors of all clause nodes is less than the preset convergence precision value of 1×10⁻⁶. -5 At this point, each clause node receives a stable final risk score, which quantifies the extent to which the clause is affected by the new regulations. The higher the score, the greater the compliance risk.

[0067] In step S4, the specific process of generating the targeted risk warning list is as follows: The system sets a risk score threshold, with an initial value of 0.75, which can be dynamically adjusted through a closed-loop feedback mechanism. The system traverses all clause nodes in the contract clause knowledge graph and filters out high-risk clause nodes whose final risk scores are higher than the threshold. For each high-risk clause node, the system traces back to its parent contract node through its parent edge to obtain the complete metadata of the contract. Subsequently, the system calculates the cumulative risk score at the contract level for each contract node, which is the sum of the risk scores of all its subordinate high-risk clause nodes. All contracts containing at least high-risk clauses are aggregated and sorted in descending order of cumulative risk scores to form a targeted risk warning list. This list includes the contract number, contract name, signatories, a list of high-risk clauses, corresponding legal points, the risk score of each clause, and the cumulative risk score of the contract.

[0068] After generating the list, the system pushes it to the contract management personnel's workbench via a graphical user interface. High-risk clauses are highlighted in red; clicking on any clause displays its original text, related legal provisions, semantic similarity score, and a visualization of the risk propagation path. Simultaneously, the system's integrated workflow engine automatically creates review tasks, assigning them to the corresponding legal specialists or business managers based on contract type and risk level, and setting processing deadlines. All operation logs, modification records, and review comments generated during task processing are recorded by the system, forming a complete compliance audit trail.

[0069] After generating the targeted risk warning list, the method also includes closed-loop feedback and model optimization steps. Legal experts manually review each item on the warning list to determine whether it is a genuine compliance risk and mark it as a true positive or false positive in the system. The system periodically collects all labeled data to construct a supervised learning sample set. For the propagation weight parameters w_self and α in the graph convolutional network, the system uses gradient descent for fine-tuning, aiming to minimize the weighted sum of the false positive and false negative rates. Simultaneously, the system dynamically adjusts the risk score threshold based on historical review results: if the false positive rate exceeds 30% in 10 consecutive warnings, the threshold is automatically increased by 0.05; if the false negative rate exceeds 5%, it is automatically decreased by 0.03. This closed-loop mechanism ensures that the risk assessment model continuously evolves over time, with a steady improvement in accuracy.

[0070] The contract lifecycle management system includes a contract data processing module, a knowledge graph construction and storage module, an external legal intelligence acquisition module, a risk propagation calculation engine, and an early warning and presentation module. The contract data processing module is responsible for acquiring the text of all contracts, optical character recognition, layout analysis, and clause segmentation, and uses a pre-trained language model in the legal domain to generate clause semantic vectors. The knowledge graph construction and storage module receives structured contract data and semantic vectors, constructs a heterogeneous knowledge graph, and stores it in a graph database, supporting efficient graph queries. The external legal intelligence acquisition module captures legal text in real time through a network interface, parses key clauses, and generates legal risk vectors. The risk propagation calculation engine is deployed on a server equipped with multiple graphics processing units, compiling graph convolutional network computation tasks into parallel computing kernels to accelerate the iterative convergence of risk scores at the hardware level. The early warning and presentation module is responsible for risk list generation, user interface rendering, task assignment, and audit log recording, ensuring that the risk handling process is traceable and verifiable.

[0071] Through the collaborative work of the aforementioned modules, the system achieves full-process automation from regulatory awareness, semantic understanding, risk propagation to targeted early warning, completely changing the inefficient traditional model that relies on manual rule maintenance and full keyword scanning, and providing enterprises with high-precision, high-efficiency, and highly adaptable intelligent contract compliance management capabilities.

Claims

1. A contract lifecycle management method, characterized in that, include: Obtain the full volume of contract documents within the enterprise; The full volume of contract document data is preprocessed to break down each contract's unstructured text into independent clause-level text units; A pre-trained language model for the legal domain based on a transformer architecture is used to semantically encode each clause-level text unit, generating a corresponding clause semantic vector; a contract clause knowledge graph is constructed, and the clause semantic vectors are stored as attributes of the corresponding clause nodes. Obtain the original text data of newly effective or newly issued regulations; the regulations include: laws, departmental rules, and judicial interpretations; The system parses the acquired raw text data to form a structured text of key legal points. Using a pre-trained language model in the legal domain, the system semantically encodes the structured text of key legal points, generating one or more legal risk vectors. The system injects the newly generated or multiple legal risk vectors into the constructed contract clause knowledge graph as temporary risk source nodes. These risk source nodes do not establish permanent connections with existing nodes in the graph; they only serve as initial incentive signals in this risk assessment calculation. Subsequently, the system initiates a graph convolutional network algorithm to perform multiple rounds of iterative calculations across the entire knowledge graph. In each iteration, each clause node updates its risk score according to the following rules: Let the first In round iteration, clause nodes The risk score is Its set of adjacent nodes is Then the first Risk score of the round Calculated by the following formula: ； in, To modify the activation function of the linear unit; For clause nodes The semantic vector of the terms; The current regulatory risk vector; The cosine similarity function is used to measure the degree of semantic matching between the clause's semantics and the regulatory risks. The self-loop weight is fixed at 0.6; For nodes With adjacent nodes The propagation weight between them is calculated using the following formula: ； in For nodes and Indicator variables indicating whether there is a structural connection between them. The cosine similarity is the semantic similarity between the two nodes. This is the structure-semantic balance coefficient, with a value of 0.

7. In each iteration, each clause node updates its own risk score based on the cosine similarity between its own clause semantic vector and the regulatory risk vector, as well as the risk status of its neighboring nodes in the graph; the calculation is iterated until the risk scores of all nodes converge, thus obtaining the final risk score of each contract clause node for the new regulations. Set a risk score threshold; filter out all clause nodes in the contract clause knowledge graph whose final risk score is higher than the risk score threshold; Based on the attribution relationship of the clause nodes in the knowledge graph, trace back to the contract subject to which they belong; summarize all contracts containing high-risk clause nodes, and sort them in descending order according to the cumulative risk score at the contract level to generate a targeted risk warning list. The targeted risk warning list is pushed to the user interface of the contract management platform, and the specific risk clauses, related regulatory points, and quantified risk scores are highlighted.

2. The contract lifecycle management method according to claim 1, characterized in that, The full volume of contract document data is preprocessed, and each contract's unstructured text is split into independent clause-level text units, including: For contract documents in scanned image format, the optical character recognition unit is invoked to convert them into editable text; By using layout analysis units to identify chapter titles, clause numbers, and body text in the contract text, precise clause-level text unit segmentation can be achieved.

3. The contract lifecycle management method according to claim 2, characterized in that, A pre-trained language model for the legal domain based on a transformer architecture is used to semantically encode each clause-level text unit, generating a corresponding clause semantic vector, including: The clause-level text unit is fed into the pre-trained language model of the legal domain as input; the output vector of the corresponding classification label in the last hidden layer of the model is taken as a 768-dimensional floating-point vector, which is used as the clause semantic vector.

4. The contract lifecycle management method according to claim 3, characterized in that, Constructing a knowledge graph of contract terms, including: Contract nodes, clause nodes, and legal entity nodes are used as graph nodes; the legal entity nodes include the contract signatories, the subject matter of the contract, the place of performance, the dispute resolution method, the applicable law, and the specific amount and date; the parent-child relationship between contracts and clauses, the cross-reference relationship between clauses, and the attribute relationship between clauses and legal entities are used as edges to construct a knowledge graph with a heterogeneous graph structure.

5. The contract lifecycle management method according to claim 4, characterized in that, The acquired raw text data is parsed to extract mandatory provisions, prohibitions, obligations, and penalties, forming a structured text of key legal points. This includes: identifying the types of behavioral constraints in the legal text; and extracting each type of constraint as an independent text unit of key legal points.

6. The contract lifecycle management method according to claim 5, characterized in that, Using the pre-trained language model in the legal domain, the structured text of key legal points is semantically encoded to generate one or more legal risk vectors, including: Each key text unit of the regulations is fed into the pre-trained language model of the legal domain as input; the output vector of the corresponding classification label in the last hidden layer of the model is taken as a 768-dimensional floating-point vector, which is used as the regulatory risk vector.

7. The contract lifecycle management method according to claim 6, characterized in that, The graph convolutional network algorithm is applied to perform multiple rounds of iterative computation on the contract terms knowledge graph, including: The next state risk score of each clause node is calculated by a nonlinear activation function through a weighted average of its own current state features and the current state features of all its neighboring nodes; wherein the weight of the weighted average is determined by the structural connection relationship and semantic similarity between nodes.

8. The contract lifecycle management method according to claim 7, characterized in that, After generating the targeted risk warning list, the process also includes closed-loop feedback and model optimization steps: Human legal experts review the warning items in the risk warning list and mark them as true positives or false positives. The system collects all labeled data and uses it as training samples. It then fine-tunes the weight parameters in the graph convolutional network algorithm through supervised learning, or dynamically and adaptively adjusts the risk score threshold.

9. A contract lifecycle management system, characterized in that, Contract lifecycle management is implemented using the contract lifecycle management method according to any one of claims 1 to 8, including: The contract data processing module is configured to acquire all contract document data within the enterprise, perform clause-level text unit splitting, and use a pre-trained language model for the legal domain based on a transformer architecture to perform semantic encoding on each clause-level text unit, generating clause semantic vectors. The knowledge graph construction and storage module is configured to receive the semantic vectors and structured data of the contract data processing module, and construct and persistently store the contract clause knowledge graph with contracts, clauses and legal entities as nodes and internal relationships as edges. The knowledge graph construction and storage module is implemented by a graph database management system, which provides efficient query and traversal functions for graph structure data. The external regulatory intelligence acquisition module is configured to acquire newly effective regulatory texts in real time from multiple preset online legal information sources through its built-in network interface and data parser, extract key regulatory points, and call the pre-trained language model in the legal field to encode them into regulatory risk vectors. The risk propagation calculation engine, with a graph convolutional network processor at its core, is configured to perform iterative calculations on the contract terms knowledge graph when a new regulatory risk vector is received, and to calculate a quantified risk score for each term node in the graph. The warning and presentation module is configured to filter out all contracts that exceed a preset risk threshold based on the risk score output by the risk propagation calculation engine, generate a warning list sorted by risk level, and present the list and associated risk details, including specific terms, corresponding regulations and risk scores, to the user through a graphical user interface.

Images