A contract clause risk review method based on a knowledge graph and clustering analysis

Abstract

The application discloses a kind of contract clause risk review method based on knowledge graph and clustering analysis, comprising the following steps: analysis historical contract, forms historical contract knowledge graph;Extract historical performance path, utilize CompGCN to execute forward combination propagation and reverse combination propagation;According to the performance position pairing deduction relationship message, cluster forms complete closed cluster and performance gap cluster;Analysis to be reviewed contract forms to be reviewed performance path, determine cluster attribution;Mark first non-closed position, extract protection relationship and form test combination path;Circulating propagation and deduction, freeze effective protection relationship, withdraw invalid protection relationship;Reverse sequence withdraw frozen relationship, determine necessary protection relationship, generate risk level, risk basis and modification content.The application relates to the technical field of intelligent contract review, realizes the accurate review of contract cross-clause risk, improves risk positioning efficiency and modification accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent contract review technology, and in particular to a method for reviewing the risks of contract terms based on knowledge graphs and cluster analysis. Background Technology

[0002] Enterprise contract reviews typically employ manual verification, standard clause comparison, keyword retrieval, rule base matching, text vector clustering, and large language model question answering. Enterprise risk assessment is influenced by contract type, customer type, product type, and transaction model, and long-term review experience is scattered across historical contracts, review records, and business personnel judgments. Manual rule base maintenance is costly, and rules are prone to lag as business conditions change; when large language models lack historical evidence constraints, the review basis cannot consistently correspond to the enterprise's actual risk boundaries.

[0003] Existing technologies often focus on judging based on the text or partial semantics of individual clauses, making it difficult to connect the preconditions, obligations, and results of performance distributed across different clauses. Clauses with similar wording but different rights and obligations structures are easily grouped into the same category, making it difficult to accurately locate cross-clause relationships, broken performance sequences, and incomplete responsibility chains. Conventional graph calculations emphasize node feature aggregation, lacking the pairing and deduction of positive and negative relationship messages, as well as the insertion and trial matching of protective relationships, the withdrawal of invalid relationships, and the screening of necessary relationships. The output risk level, risk basis, and modified content often lack a verifiable source chain.

[0004] Therefore, how to provide a method for risk review of contract terms based on knowledge graphs and cluster analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a risk review method for contract terms based on knowledge graphs and cluster analysis. This invention utilizes CompGCN bidirectional combination propagation, relationship message pairing deduction, protection relationship trial and reverse withdrawal to identify cross-term performance gaps and screen necessary protection relationships. It has the advantages of accurate risk positioning, traceable review basis, and highly targeted modification content.

[0006] A method for reviewing contract terms based on knowledge graphs and cluster analysis according to an embodiment of the present invention includes the following steps:

[0007] Analyze historical contracts, connect cross-clause rights and obligations, and form a knowledge graph of historical contracts;

[0008] Historical performance paths are extracted from the historical contract knowledge graph, and forward and backward combination propagation are performed through the Combinatorial Graph Convolutional Network (CompGCN).

[0009] Pairing deduction is performed on the relationship messages formed by forward and backward combination propagation according to the fulfillment location. Clustering is performed on the historical fulfillment paths according to the arrangement of the undeducted relationship messages to form complete closed clusters and fulfillment gap clusters.

[0010] The pending contract is analyzed to form a pending performance path. Forward combination propagation, backward combination propagation and pairing deduction are performed on the pending performance path. The cluster affiliation of the pending performance path is determined based on the arrangement of the undeducted relational messages.

[0011] When a pending performance path is included in a performance gap cluster, the first unclosed position is marked according to the first undeducted relationship message. The protection relationship between the nodes on both sides of the first unclosed position is extracted from the complete closed cluster and inserted into the pending performance path to form a trial path.

[0012] The trial path is repeatedly subjected to forward combination propagation, reverse combination propagation and pairing deduction, freezing protection relationships that reduce the number of undeducted relationship messages, and withdrawing protection relationships that do not reduce the number of undeducted relationship messages, until there are no undeducted relationship messages in the trial path;

[0013] Recall frozen protection relationships in the reverse order of insertion, identify protection relationships that cause the recurrence of undeducted relationship messages as necessary protection relationships, and generate risk levels, risk basis, and modification content based on necessary protection relationships.

[0014] Optionally, the formation of the historical contract knowledge graph specifically includes:

[0015] Historical contracts are broken down into clause nodes, and clause nodes corresponding to the performance prerequisites, performance obligations, and performance results distributed across different clauses are extracted according to the same performance object.

[0016] Connect the corresponding clause nodes in the order of performance premises, performance obligations, and performance results to form cross-clause rights and obligations relationships;

[0017] A knowledge graph of historical contracts is formed based on cross-clause rights and obligations.

[0018] Optionally, the execution of forward combination propagation and backward combination propagation specifically includes:

[0019] Using the node representations in the historical performance path and the relationship representations of cross-clause rights and obligations as shared inputs, the node representations and relationship representations are combined along the node order of the historical performance path to generate positive relationship messages;

[0020] Reverse the direction of cross-clause rights and obligations to generate a reverse relationship representation. Combine the node representation and the reverse relationship representation along the reverse node order of the historical performance path to generate a reverse relationship message.

[0021] Bind forward and reverse relationship messages to the same fulfillment position.

[0022] Optionally, the execution of paired deduction specifically includes:

[0023] Establish message pairs between forward and reverse relationship messages that are bound to each other, based on the condition that the relationship type is consistent and the node order is reversed.

[0024] The relationship direction of the reverse relationship message in the reverse relationship pair is calculated, and the difference between the forward relationship message and the reverse relationship message after the reverse relationship direction is calculated dimension by dimension.

[0025] Message pairs whose differences in each dimension meet the preset consistency conditions are paired and deducted; message pairs whose differences in each dimension do not meet the preset consistency conditions are retained as undeducted relationship messages.

[0026] Arrange the undeducted relationship messages according to their corresponding performance positions to form an arrangement status of the undeducted relationship messages.

[0027] Optionally, the clustering process specifically includes:

[0028] Historical fulfillment paths that do not contain undeducted relationship messages are grouped into complete closed clusters;

[0029] Historical fulfillment paths with undeducted relationship messages are grouped according to the fulfillment position corresponding to the first undeducted relationship message. Historical fulfillment paths with the same arrangement of undeducted relationship messages within the same group are grouped into the same fulfillment gap cluster.

[0030] Optionally, determining the cluster affiliation of the performance path under review specifically includes:

[0031] If there are no undeducted relationship messages in the pending performance path, the pending performance path will be classified into a complete closed cluster;

[0032] When there are undeducted relationship messages in the pending performance path and there are performance gap clusters that meet the inclusion criteria, the pending performance path will be included in the performance gap cluster with the same performance position as the first undeducted relationship message and the same arrangement of the undeducted relationship messages.

[0033] If there are undeducted relationship messages in the pending performance path and there is no performance gap cluster that meets the inclusion criteria, a new performance gap cluster will be formed using the pending performance path.

[0034] Optionally, the formation of the trial-matching path specifically includes:

[0035] Mark the fulfillment position corresponding to the first unclosed relationship message as the first unclosed position;

[0036] Extract protection relationships from the historical performance paths in the complete closed cluster that connect the nodes on both sides of the first unclosed position and whose relationship direction is consistent with the node order of the performance path to be reviewed;

[0037] The protection relationship is inserted into the pending performance path according to its corresponding historical performance position, forming a trial cooperation path.

[0038] Optionally, the step of cyclically performing forward combination propagation, backward combination propagation, and pairing subtraction on the trial path specifically includes:

[0039] For each trial path after inserting a protection relationship, perform forward combination propagation, reverse combination propagation, and pairing deduction, and compare the number of undeducted relationship messages before and after insertion;

[0040] Freeze protected relationships that reduce the number of undeducted relationship messages, and withdraw protected relationships that do not reduce the number of undeducted relationship messages;

[0041] The frozen protection relationships are retained, and the trial path is used to continue the loop until there are no undeducted relationship messages in the trial path.

[0042] Optionally, the determination of the protection relationship that causes the reproduction of the undeducted relationship message as a necessary protection relationship specifically includes:

[0043] After the trial path no longer contains messages with undeducted relationships, the frozen protection relationships are withdrawn one by one in the reverse order of insertion.

[0044] After each withdrawal, forward combination propagation, backward combination propagation, and pairing deduction are performed on the trial path;

[0045] If the undeducted relationship message does not reappear, the protection relationship remains in a withdrawn state. If the undeducted relationship message reappears, the withdrawn protection relationship is restored, and the restored protection relationship is determined as a necessary protection relationship, until all frozen protection relationships have been withdrawn and paired for deduction.

[0046] Optionally, the generation of risk level, risk basis, and modification content specifically includes:

[0047] The risk level is determined based on the number of necessary protection relationships and the corresponding performance location of those relationships.

[0048] Tracing back the source historical performance path along the necessary protection relationship to the complete closed cluster, the clause nodes and cross-clause rights and obligations relationships in the source historical performance path are identified as the basis for risk assessment;

[0049] The necessary protection relationships are mapped to the corresponding performance positions in the performance path to be reviewed, and the rights and obligations corresponding to the necessary protection relationships are written into the contract clauses of the contract to be reviewed associated with the corresponding performance positions, thus forming the revised content.

[0050] The beneficial effects of this invention are:

[0051] (1) This invention connects the performance premises, performance obligations and performance results scattered in different contract terms into a historical performance path, and uses CompGCN to perform forward combination propagation, reverse combination propagation and relationship message pairing deduction to eliminate closed rights and obligations relationships and retain the unclosed parts in the performance structure; this processing can identify contract risks with similar textual expressions but different liability assumption relationships, realize cross-term performance gap location, and the effect is different from keyword matching, text similarity calculation and single-term classification.

[0052] (2) The present invention performs clustering on historical performance paths based on the performance position and arrangement status corresponding to the undeducted relationship messages, forming complete closed clusters and performance gap clusters, and then determines the cluster affiliation of the performance path to be reviewed based on the same structural conditions; the clustering object is transformed from contract text vectors into performance gap structures, which can classify contracts with the same gap position and the same relationship break order into the same risk category, reduce the interference of differences in clause wording on the review results, and enable historical contract experience to be reused according to the rights and obligations structure.

[0053] (3) This invention extracts protection relationships from complete closed clusters and screens the protection relationships necessary to maintain the closure of the performance path under review by inserting trial contracts, recalculating relationship messages, freezing protection relationships, withdrawing and reverse withdrawing tests. The risk level, risk basis and modification content are all derived from the necessary protection relationships, which can eliminate redundant modifications that have no practical effect on the performance closure and form a traceable link from the risk location, historical source to the modification content, avoiding the direct replacement of standard clauses that would cause the scope of liability to expand or the modification content to deviate from the contract transaction structure. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of a contract clause risk review method based on knowledge graph and cluster analysis proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the relationship message pairing deduction and historical performance path clustering process proposed in this invention;

[0057] Figure 3 This is a schematic diagram of the protection relationship trial, freezing, and reverse withdrawal process proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figures 1-3 A method for reviewing contract terms based on knowledge graphs and cluster analysis includes the following steps:

[0060] Analyze historical contracts, connect cross-clause rights and obligations, and form a knowledge graph of historical contracts;

[0061] Historical performance paths are extracted from the historical contract knowledge graph, and forward and backward combination propagation are performed through the Combinatorial Graph Convolutional Network (CompGCN).

[0062] Pairing deduction is performed on the relationship messages formed by forward and backward combination propagation according to the fulfillment location. Clustering is performed on the historical fulfillment paths according to the arrangement of the undeducted relationship messages to form complete closed clusters and fulfillment gap clusters.

[0063] A complete closed cluster is a set of historical fulfillment paths that have no undeducted relationship messages after pairing deduction; a fulfillment gap cluster is a set of historical fulfillment paths that have the same fulfillment position corresponding to the first undeducted relationship message and have the same arrangement of undeducted relationship messages.

[0064] The pending contract is analyzed to form a pending performance path. Forward combination propagation, backward combination propagation and pairing deduction are performed on the pending performance path. The cluster affiliation of the pending performance path is determined based on the arrangement of the undeducted relational messages.

[0065] When a pending performance path is included in a performance gap cluster, the first unclosed position is marked according to the first undeducted relationship message. The protection relationship between the nodes on both sides of the first unclosed position is extracted from the complete closed cluster and inserted into the pending performance path to form a trial path.

[0066] The trial path is repeatedly subjected to forward combination propagation, reverse combination propagation and pairing deduction, freezing protection relationships that reduce the number of undeducted relationship messages, and withdrawing protection relationships that do not reduce the number of undeducted relationship messages, until there are no undeducted relationship messages in the trial path;

[0067] Recall frozen protection relationships in the reverse order of insertion, identify protection relationships that cause the recurrence of undeducted relationship messages as necessary protection relationships, and generate risk levels, risk basis, and modification content based on necessary protection relationships.

[0068] In this implementation method, the formation of a historical contract knowledge graph specifically includes:

[0069] Historical contracts are broken down into clause nodes, and clause nodes corresponding to the performance prerequisites, performance obligations, and performance results distributed across different clauses are extracted according to the same performance object.

[0070] Clause nodes are text fragments in historical contracts that can independently express the premises, obligations, or results of performance. Historical contracts are segmented according to clause numbers and paragraph boundaries. Text fragments that can independently express the premises, obligations, or results of performance are extracted from the segmented text fragments, and each extracted text fragment forms a clause node.

[0071] The same contractual object is determined by clauses in which the subject matter, the obligor, and the right holder are all the same; the conditions for the obligation to take effect are extracted from the clause nodes as the prerequisites for performance, the actions that the obligor needs to complete are extracted as the performance obligations, and the rights arising after the performance or the responsibilities arising if the performance is not completed are extracted as the performance results.

[0072] Connect the corresponding clause nodes in the order of performance premises, performance obligations, and performance results to form cross-clause rights and obligations relationships;

[0073] Cross-clause rights and obligations relationships are directed connections formed between different clause nodes corresponding to the same performance object; the clause node corresponding to the performance premise is connected to the clause node corresponding to the performance obligation, and the clause node corresponding to the performance obligation is connected to the clause node corresponding to the performance result. The direction of the connection is determined according to the order of performance of the performance premise, performance obligation and performance result.

[0074] A knowledge graph of historical contracts is formed based on cross-clause rights and obligations.

[0075] The historical contract knowledge graph consists of clause nodes and cross-clause rights and obligations relationships. Each historical contract's corresponding clause node is used as a node, and the cross-clause rights and obligations relationships are used as directed connections. The contract number and clause number corresponding to each clause node are recorded to form the historical contract knowledge graph.

[0076] In one implementation, an auxiliary review information layer is set up around the historical contract knowledge graph. The auxiliary review information layer records the contract type, customer type, product type, deployment method, transaction model, amount range, and performance status of each historical contract, and associates payment, acceptance, data processing, intellectual property, liability limitation, service scope, and channel cooperation as risk element categories to the corresponding clause nodes. The contract type, customer type, product type, deployment method, transaction model, amount range, and performance status are extracted from the historical contract registration information and contract text. The risk element categories are determined according to the rights and obligations content in the clause nodes. Clause nodes involving payment and settlement are associated with the payment category; clause nodes involving delivery confirmation, testing, and acceptance are associated with the acceptance category; clause nodes involving data collection, storage, use, export, and deletion are associated with the data processing category; clause nodes involving ownership and licensing of results are associated with the intellectual property category; clause nodes involving breach of contract, scope of compensation, and liability ceiling are associated with the liability limitation category; clause nodes involving service content, service period, and service standards are associated with the service scope category; and clause nodes involving channel authorization, sales area, and customer affiliation are associated with the channel cooperation category.

[0077] The auxiliary review information layer also records company preference content; the company preference content comes from historical contract terms and review conclusions that have been manually confirmed and actually adopted, including performance prerequisites that must be retained, performance results that cannot be deleted, acceptable scope of liability, and unacceptable rights and obligations; each piece of company preference content records the source contract number, source clause number, and applicable contract metadata conditions.

[0078] In this embodiment, forward combination propagation and backward combination propagation are performed, specifically including:

[0079] Using the node representations in the historical performance path and the relationship representations of cross-clause rights and obligations as shared inputs, the node representations and relationship representations are combined along the node order of the historical performance path to generate positive relationship messages;

[0080] The historical performance path is a directed path that connects the clause nodes corresponding to the performance premise, the performance obligation, and the performance result in sequence along the cross-clause rights and obligations relationship; the performance position is determined according to the order of the clause nodes and the cross-clause rights and obligations relationship in the directed path.

[0081] Node representation is a fixed-length vector obtained by vectorizing the text in the clause nodes; relationship types include the relationship between performance prerequisites and performance obligations, and the relationship between performance obligations and performance results; relationship directions include the forward direction along the historical performance path node order and the reverse direction along the historical performance path node order; relationship representation is a fixed-length vector obtained by vectorizing the relationship type and relationship direction of cross-clause rights and obligations; the vector dimension of node representation and relationship representation is the same;

[0082] Node representation and relationship representation are obtained through training. During training, the historical performance path with complete rights and obligations relationships confirmed by humans is used as a closed sample. The training objective is to minimize the norm of the dimension-wise difference vector between the forward relationship message and the reverse relationship message after the reverse relationship direction in the closed sample. This ensures that the message pairs corresponding to the closed performance path meet the preset consistency condition, and that the message pairs at the missing positions of the performance path with missing rights and obligations relationships do not meet the preset consistency condition.

[0083] The shared input consists of node representations and relation representations read jointly by forward and backward combinatorial propagation; the combination of node representations and relation representations is a dimension-wise multiplication, and a linear mapping is performed on the result of the dimension-wise multiplication to generate relation messages corresponding to the fulfillment position; forward and backward combinatorial propagation use the same dimension-wise multiplication method and linear mapping parameters;

[0084] Reverse the direction of cross-clause rights and obligations to generate a reverse relationship representation. Combine the node representation and the reverse relationship representation along the reverse node order of the historical performance path to generate a reverse relationship message.

[0085] The reverse relationship representation is obtained by reversing the relationship direction of cross-clause rights and obligations and performing vectorization processing on the cross-clause rights and obligations after reversing the relationship direction; the reverse relationship message is associated with the performance position of the corresponding cross-clause rights and obligations in the historical performance path;

[0086] Bind forward and reverse relationship messages to the same fulfillment position.

[0087] In one implementation, a graph-enhanced auxiliary representation is established for the clause node. The graph-enhanced auxiliary representation is obtained by performing a linear mapping after splicing the contract metadata representation, risk element representation, and graph location representation. The contract metadata representation is formed according to the contract type, customer type, product type, deployment method, transaction mode, amount range, and performance status encoding of the contract to which the clause node belongs. The risk element representation is formed according to the risk element category encoding corresponding to the clause node. The graph location representation is formed according to the performance prerequisite, performance obligation, or performance result category to which the clause node belongs, as well as the performance position encoding of the clause node in the historical performance path.

[0088] Graph-enhanced auxiliary representations are used to sort source historical performance paths and protection relationships when the performance positions corresponding to the first undeducted relationship message are the same and the arrangement status of the undeducted relationship messages is consistent. The graph-enhanced auxiliary representations in the historical performance path and the performance path to be reviewed are spliced ​​together according to the performance position to form the corresponding performance path auxiliary representation. The cosine similarity between the performance path auxiliary representation to be reviewed and the auxiliary representations of each source historical performance path is calculated, and the source historical performance paths are arranged from high to low according to the cosine similarity. The protection relationships are arranged according to the arrangement order of their respective source historical performance paths. Forward combination propagation, reverse combination propagation, and pairing deduction are still performed according to node representation and relationship representation.

[0089] In this embodiment, the paired deduction is performed, specifically including:

[0090] Establish message pairs between forward and reverse relationship messages that are bound to each other, based on the condition that the relationship type is consistent and the node order is reversed.

[0091] Reversed node order indicates that forward relationship messages are generated by pointing from the previous clause node to the next clause node, and reverse relationship messages are generated by pointing from the next clause node to the previous clause node.

[0092] The relationship direction of the reverse relationship message in the reverse relationship pair is calculated, and the difference between the forward relationship message and the reverse relationship message after the reverse relationship direction is calculated dimension by dimension.

[0093] The reverse relationship message is reversed by exchanging the start clause node and end clause node corresponding to the generation of the reverse relationship message, so that the reverse relationship message and the forward relationship message have the same comparison direction;

[0094] The dimension-by-dimensional difference is calculated by subtracting the corresponding dimension value from the reverse relationship message after reversing the relationship direction from the value of each dimension in the forward relationship message, according to the vector dimension number.

[0095] Message pairs whose differences in each dimension meet the preset consistency conditions are paired and deducted; message pairs whose differences in each dimension do not meet the preset consistency conditions are retained as undeducted relationship messages.

[0096] The preset consistency condition is the condition for determining whether the message pair has completed the pairing deduction. This includes the absolute value of the difference in each dimension not being greater than the preset threshold, or the norm of the difference vector in each dimension not being greater than the preset threshold. All differences in each dimension being zero is a case where the preset threshold is zero. The preset threshold is determined based on the vector dimension and numerical range of the node representation and relationship representation, as well as the statistical results of historical fulfillment path samples.

[0097] The undeducted relation message is the dimension-wise difference vector corresponding to the message pair whose differences in each dimension do not meet the preset consistency condition, and retains the corresponding performance position and relation type of the message pair. The relation direction that is consistent with the forward relation message after reversal is taken as the relation direction of the undeducted relation message.

[0098] Arrange the undeducted relationship messages according to their corresponding performance positions to form an arrangement status of the undeducted relationship messages.

[0099] The first undeducted relationship message in the fulfillment position is the one that has not yet been deducted.

[0100] In this embodiment, clustering is performed, specifically including:

[0101] Historical fulfillment paths that do not contain undeducted relationship messages are grouped into complete closed clusters;

[0102] Historical fulfillment paths with undeducted relationship messages are grouped according to the fulfillment position corresponding to the first undeducted relationship message. Historical fulfillment paths with the same arrangement of undeducted relationship messages within the same group are grouped into the same fulfillment gap cluster.

[0103] The consistent arrangement of undeducted relationship messages indicates that the number of undeducted relationship messages is the same. When comparing them item by item from front to back according to the performance position, the performance position, relationship type and relationship direction corresponding to each undeducted relationship message are the same.

[0104] In one implementation, a cluster profile is generated for complete closed clusters and performance gap clusters; the cluster profile records the contract type, customer type, product type, deployment method, transaction mode, amount range, risk element category, performance position corresponding to the first undeducted relationship message, and the arrangement status of the undeducted relationship messages corresponding to the historical performance paths within the cluster.

[0105] For two performance gap clusters with the same performance position corresponding to the first undeducted relationship message, the arrangement of undeducted relationship messages in one performance gap cluster is compared item by item with that in the other performance gap cluster. If there is only one difference in relationship type or one difference in relationship direction, the historical performance paths in the two performance gap clusters are marked as boundary samples. Performance gap clusters containing only one or two historical performance paths are marked as long-tail clusters, and the historical performance paths in the long-tail clusters are marked as long-tail samples. When the performance path under review does not have a performance gap cluster that meets the inclusion criteria and forms a new performance gap cluster, the performance path under review is marked as an outlier sample.

[0106] The labeling results of boundary samples, long-tail samples, and outlier samples are used to determine the order of manual review, without changing the determination conditions of the complete closed cluster, the performance gap cluster, and the performance path cluster to be reviewed.

[0107] In this embodiment, determining the cluster affiliation of the performance path under review specifically includes:

[0108] If there are no undeducted relationship messages in the pending performance path, the pending performance path will be classified into a complete closed cluster;

[0109] The performance path under review is a directed path formed by extracting clause nodes from the contract under review in the same way as the historical performance paths, and connecting the clause nodes in the order of performance prerequisites, performance obligations, and performance results.

[0110] When there are undeducted relationship messages in the pending performance path and there are performance gap clusters that meet the inclusion criteria, the pending performance path will be included in the performance gap cluster with the same performance position as the first undeducted relationship message and the same arrangement of the undeducted relationship messages.

[0111] If there are undeducted relationship messages in the pending performance path and there is no performance gap cluster that meets the inclusion criteria, a new performance gap cluster will be formed using the pending performance path.

[0112] In this embodiment, forming a trial-matching path specifically includes:

[0113] Mark the fulfillment position corresponding to the first unclosed relationship message as the first unclosed position;

[0114] The first unclosed position is the performance position corresponding to the first undeducted relationship message in the performance path pending review.

[0115] Extract protection relationships from the historical performance paths in the complete closed cluster that connect the nodes on both sides of the first unclosed position and whose relationship direction is consistent with the node order of the performance path to be reviewed;

[0116] The nodes on both sides of the first unclosed position are the adjacent clause nodes located before and after the first unclosed position in the performance path to be reviewed; the bridging indicates that the starting node of the protection relationship corresponds to the preceding clause node, and the ending node of the protection relationship corresponds to the following clause node;

[0117] The protected relationship is the cross-clause rights and obligations relationship extracted from the historical performance path in the complete closed cluster. During extraction, the starting clause node of the cross-clause rights and obligations relationship is matched with the node before the first unclosed position, and the ending clause node of the cross-clause rights and obligations relationship is matched with the node after the first unclosed position. Each set of matched nodes corresponds to the same performance object and belongs to the same category in terms of performance premise, performance obligation, and performance result. Cross-clause rights and obligations relationships whose relationship direction is consistent with the node order of the performance path under review are retained.

[0118] The protection relationship is inserted into the pending performance path according to its corresponding historical performance position, forming a trial cooperation path.

[0119] In this embodiment, the corresponding historical performance path is the source historical performance path for extracting the protection relationship.

[0120] The trial path is a directed path formed by retaining the original clause nodes and cross-clause rights and obligations in the pending performance path and adding protection relationships.

[0121] In this embodiment, forward combination propagation, backward combination propagation, and pairing deduction are performed iteratively on the trial path, specifically including:

[0122] For each trial path after inserting a protection relationship, perform forward combination propagation, reverse combination propagation, and pairing deduction, and compare the number of undeducted relationship messages before and after insertion;

[0123] The number of relationship messages not deducted before insertion is the number of relationship messages retained after performing forward combination propagation, reverse combination propagation, and pairing deduction on the path before the insertion of the protection relationship; the number of relationship messages not deducted after insertion is the number of relationship messages retained after performing forward combination propagation, reverse combination propagation, and pairing deduction on the formed trial path after the insertion of the protection relationship.

[0124] Freeze protected relationships that reduce the number of undeducted relationship messages, and withdraw protected relationships that do not reduce the number of undeducted relationship messages;

[0125] Freeze means to retain the currently inserted protection relationship and maintain the insertion state of the protection relationship in subsequent loops; withdraw means to remove the currently inserted protection relationship from the trial path and restore the path before the current protection relationship was inserted.

[0126] The frozen protection relationships are retained, and the trial path is used to continue the loop until there are no undeducted relationship messages in the trial path.

[0127] In this embodiment, the protection relationships that cause the reproduction of undeducted relationship messages are identified as necessary protection relationships, specifically including:

[0128] After the trial path no longer contains messages with undeducted relationships, the frozen protection relationships are withdrawn one by one in the reverse order of insertion.

[0129] The insertion order is the chronological order in which the frozen protection relationships are inserted into the trial path; the reverse order is to start from the last inserted protection relationship and withdraw it sequentially to the first inserted protection relationship.

[0130] After each withdrawal, forward combination propagation, backward combination propagation, and pairing deduction are performed on the trial path;

[0131] If the undeducted relationship message does not reappear, the protection relationship remains in a withdrawn state. If the undeducted relationship message reappears, the withdrawn protection relationship is restored, and the restored protection relationship is determined as a necessary protection relationship, until all frozen protection relationships have been withdrawn and paired for deduction.

[0132] The reappearance of undeducted relationship messages indicates that there were no undeducted relationship messages in the trial path before the protection relationship was withdrawn. After withdrawing the protection relationship and performing forward combination propagation, reverse combination propagation, and pairing deduction, at least one undeducted relationship message appears in the trial path. The failure to reproduce undeducted relationship messages indicates that there are still no undeducted relationship messages in the trial path after withdrawing the protection relationship and performing forward combination propagation, reverse combination propagation, and pairing deduction.

[0133] In this implementation method, the generation of risk level, risk basis, and modification content specifically includes:

[0134] The risk level is determined based on the number of necessary protection relationships and the corresponding performance location of those relationships.

[0135] Risk levels are categorized as high, medium, and low. One point is awarded when the performance position corresponding to a necessary protection relationship is located between the clause node corresponding to the performance prerequisite and the clause node corresponding to the performance obligation; two points are awarded when the performance position corresponding to a necessary protection relationship is located between the clause node corresponding to the performance obligation and the clause node corresponding to the performance result. The scores for each necessary protection relationship are added together. A cumulative score of 0 or 1 point indicates low risk; a cumulative score of 2 points indicates medium risk; and a cumulative score of at least 3 points indicates high risk.

[0136] Tracing back the source historical performance path along the necessary protection relationship to the complete closed cluster, the clause nodes and cross-clause rights and obligations relationships in the source historical performance path are identified as the basis for risk assessment;

[0137] When tracing back along the necessary protection relationship, the contract number and clause number recorded in the historical contract knowledge graph are used as a basis for determining the historical performance path to which the necessary protection relationship belongs, and the performance position of the necessary protection relationship in the historical performance path. All clause nodes and cross-clause rights and obligations in the historical performance path are identified as the basis for risk assessment.

[0138] The necessary protection relationships are mapped to the corresponding performance positions in the performance path to be reviewed, and the rights and obligations corresponding to the necessary protection relationships are written into the contract clauses of the contract to be reviewed associated with the corresponding performance positions, thus forming the revised content.

[0139] The rights and obligations corresponding to the necessary protection relationship originate from the starting and ending clauses that constitute the necessary protection relationship. The texts of the performance prerequisites, performance obligations, or performance results carried by the starting and ending clauses are compared with the contract clauses under review associated with the corresponding performance positions. When comparing performance prerequisites, the conditions for the obligation to take effect are compared; when comparing performance obligations, the actions that the obligor needs to complete are compared; when comparing performance results, the rights arising after performance or the liabilities arising if performance is not completed are compared. If there is no clause text in the contract clauses under review that corresponds to the same performance object, belongs to the same category as the starting or ending clause, and has the same content in terms of the conditions for the obligation to take effect, the actions that need to be completed, the rights or liabilities arising, the corresponding clause text is determined to be the rights and obligations not present in the contract clauses under review. The missing rights and obligations are written into the contract clauses under review associated with the corresponding performance positions, forming the modified content.

[0140] In one implementation, rule candidate records are generated based on necessary protection relationships. The rule candidate records include the applicable contract type, customer type, product type, deployment method and transaction mode, the performance location, risk level, risk basis, modification content, source contract number and source clause number corresponding to the necessary protection relationship. When the risk basis or modification content corresponding to the same necessary protection relationship changes, a version number is added to the rule candidate record.

[0141] The contract review agent reads the contract metadata, risk level, necessary protection relationships, risk basis, and modification content of the contract to be reviewed, and organizes the review results according to the applicable conditions in the rule candidate record. The contract review agent does not change the determination result of necessary protection relationships, but is only used to output the risk location, source historical performance path, risk basis, and modification content.

[0142] The results of manual review include acceptance, modification, and non-acceptance. When accepted, the risk basis and modified content are recorded as the version of valid evidence. When modified, the manually modified content is recorded as the new version of valid evidence, and the original version is retained. When non-accepted, the corresponding rule candidate record is marked as disabled. Subsequent contract reviews read the latest version of valid evidence and write the source contract number, clause number, and necessary protection relationship corresponding to the version of valid evidence into the auxiliary review information layer.

[0143] Example 1: To verify the feasibility of this invention in practice, it was applied to a centralized contract review scenario. The verification sample used anonymized simulated contracts covering eight business scenarios: private cloud sales, public cloud sales, hardware / software integration sales, hardware upgrades, channel cooperation, service renewals, technical services, and data services. The contract content involved seven risk elements: payment, acceptance, data processing, intellectual property, limitation of liability, scope of service, and channel cooperation. Each contract retained information such as contract type, customer type, product type, deployment method, transaction model, amount range, and performance status for verifying the business conditions under which the contract was executed. In this scenario, payment terms, delivery requirements, acceptance standards, data operation conditions, service obligations, and liability assumptions were scattered across different clauses. Viewing a single clause alone made it difficult to determine whether the preconditions, obligations, and results of performance could form a closed loop.

[0144] This embodiment sets up 30 historical contract samples, including 5 private cloud purchase and sale contracts, 4 public cloud purchase and sale contracts, 4 hardware and software integration purchase and sale contracts, 3 hardware additional purchase contracts, 4 channel cooperation contracts, 3 service renewal contracts, 4 technical service contracts, and 3 data service contracts. These 30 historical contracts generated a total of 938 clause text fragments. After review, 286 cross-clause rights and obligations relationships and 112 historical performance paths were extracted. Based on these 30 historical contracts, 30 contracts awaiting review were formed. While maintaining the contract subject categories, contract objects, and transaction structures, controlled deletions and rewritings were made to the performance prerequisites, performance obligations, or performance results in some clauses. Reviewers checked each contract awaiting review, identifying 74 cross-clause risk points, including 12 payment-related, 15 acceptance-related, 11 data-related, 9 intellectual property-related, 13 liability limitation-related, 8 service scope-related, and 6 channel cooperation-related.

[0145] When applying this invention, historical contracts are segmented according to clause numbers and paragraph boundaries. Clause nodes corresponding to the performance prerequisites, performance obligations, and performance results are extracted from the segmented text fragments. Clause nodes with the same contract subject matter, obligor, and right holder are grouped into the same performance object. Clause nodes corresponding to performance prerequisites are connected to clause nodes corresponding to performance obligations, and clause nodes corresponding to performance obligations are connected to clause nodes corresponding to performance results, forming a cross-clause rights and obligations relationship and a historical contract knowledge graph. Historical performance paths are extracted from the historical contract knowledge graph. Clause node text is converted into node representations, and relationship types and directions are converted into relationship representations. Forward and backward compositional propagation are performed through a Combinatorial Graph Convolutional Network (CompGCN).

[0146] Forward and reverse relationship messages are paired and bound at the same performance position. Message pairs are established for messages with the same relationship type and reversed node order. After reversing the relationship direction of the reverse relationship message, a dimension-by-dimensional difference is calculated on the message pairs. Message pairs whose differences in each dimension meet the preset consistency condition are paired and deducted; message pairs whose differences in each dimension do not meet the preset consistency condition form undeducted relationship messages. Historical performance paths without undeducted relationship messages are assigned to complete closed clusters. Historical performance paths with undeducted relationship messages are assigned to performance gap clusters according to the performance position corresponding to the first undeducted relationship message and the arrangement of the undeducted relationship messages. Contracts under review are processed in the same way to form pending performance paths and determine cluster affiliation.

[0147] In a pending private cloud sales contract, platform delivery, deployment startup, data migration, acceptance confirmation, payment, and post-service data export are scattered across six clauses. The contract retains the platform delivery obligation and the customer's payment obligation, but fails to specify the written startup conditions required to begin service deployment, nor does it outline the consequences if acceptance is not completed within the agreed timeframe. The data clause states that customer data belongs to the customer, while the written application conditions for special data export and administrator password reset are scattered across another clause, creating a closed data operation path. Conventional keyword rules, which only check risky words in a single clause, easily identify these scattered written application conditions as missing content.

[0148] This invention establishes three pending performance paths for the contract under review. The platform delivery and fee payment path has no undeducted relationship messages and is categorized into a complete closed cluster. The deployment initiation and service obligation path forms an undeducted relationship message between the performance prerequisite and the performance obligation. The acceptance confirmation and overdue processing path forms an undeducted relationship message between the performance obligation and the performance result. The latter two paths are categorized into their corresponding performance gap clusters. In this embodiment, the performance position corresponding to the first undeducted relationship message is marked as the first unclosed position. Four protection relationships are extracted from the historical performance paths in the complete closed cluster. These four protection relationships correspond to the written initiation conditions, deployment resource preparation conditions, the re-inspection processing result after supplementing acceptance data, and the acceptance overdue processing result, respectively.

[0149] This implementation method forms trial paths for the deployment initiation and service obligation path, and the acceptance confirmation and overdue processing path, respectively. For the deployment initiation and service obligation path, after inserting the deployment resource preparation conditions, the number of undeducted relationship messages remains at one, and the corresponding protection relationship is withdrawn; after inserting the written initiation conditions, the number of undeducted relationship messages decreases from one to zero, the corresponding protection relationship is frozen, and the cycle processing of this trial path ends. For the acceptance confirmation and overdue processing path, after inserting the re-verification processing result after supplementing the acceptance data, the number of undeducted relationship messages remains at one, and the corresponding protection relationship is withdrawn; after inserting the acceptance overdue processing result, the number of undeducted relationship messages decreases from one to zero, the corresponding protection relationship is frozen, and the cycle processing of this trial path ends. This processing only retains protection relationships that can reduce the number of undeducted relationship messages in the corresponding trial path, avoiding the inclusion of historical clauses with the same node category but unable to repair the current performance gap into the pending contract.

[0150] After both trial paths were found to have no undeducted relationship messages, the frozen protection relationships were withdrawn one by one in the reverse order of insertion within each trial path. After withdrawing the protection relationship corresponding to the overdue acceptance processing result, the undeducted relationship message between the performance obligation and performance result in the acceptance confirmation and overdue processing path reappeared; after restoring the corresponding protection relationship, the undeducted relationship message disappeared. After withdrawing the protection relationship corresponding to the written initiation condition, the undeducted relationship message between the performance prerequisite and performance obligation in the deployment initiation and service obligation path reappeared; after restoring the corresponding protection relationship, the undeducted relationship message disappeared. Both protection relationships were determined to be necessary protection relationships. The performance position corresponding to the written initiation condition received 1 point, and the performance position corresponding to the overdue acceptance processing result received 2 points, for a total score of 3 points, and the risk level was determined to be high risk.

[0151] This implementation method traces back the historical performance path along the necessary protection relationships, identifying the clause nodes and cross-clause rights and obligations within that path as the basis for risk assessment. The risk basis records the source contract number, clause number, deployment initiation conditions, service obligations, acceptance requirements, and the results of handling overdue acceptance. The modifications are extracted from the starting and ending clause nodes constituting the necessary protection relationships. While maintaining the original subject names and clause terminology of the contract under review, the method adds that deployment service takes effect upon receipt of a written initiation notice, and also adds the liability and cost handling details for overdue acceptance. The deployment resource preparation conditions and the results of the re-inspection after supplementing the acceptance data were not identified as necessary protection relationships because the number of relationship messages was not reduced after inserting the corresponding protection relationship; therefore, this implementation method did not include the corresponding rights and obligations in the modification results.

[0152] The verification process employs conventional keyword rule review as a comparison method. Conventional methods filter clauses based on combinations of risk terms, negative terms, and liability terms, then reviewers determine whether the clauses lack preconditions or processing results. This invention directly outputs the risk level, historical performance path, necessary protection relationships, risk basis, and modification content. Both methods use the same 30 contracts to be reviewed and the same 74 confirmed risk points. Review time is calculated from the import of the contract text to the generation of the risk level, risk basis, and modification content. Manual review time is calculated from the delivery of the review results to the reviewers to the confirmation of the risk basis and modification content. The data in the table are exemplary implementation records generated under controlled verification conditions, used to illustrate the execution effect of the technical solution.

[0153] Table 1: Comparison of Cross-Term Risk Review Effectiveness in Anonymous Contract Samples

[0154] Evaluation indicators Regular keyword rule review This invention Comparison results Number of samples reviewed 30 copies 30 copies Sample consistency Manual confirmation of risk points 74 74 Benchmark consistency Correctly detect risk points 52 71 19 more Compliance Gap Detection Rate 70.3% 95.9% Increased by 25.6 percentage points Risk points of missed detection 22 3 19 fewer False alarm risk points 17 5 Reduce 12 Risk based on positioning accuracy 39 / 52，75.0% 69 / 71，97.2% An increase of 22.2 percentage points The modified content can be directly adopted. 30 / 52，57.7% 62 / 71，87.3% Increased by 29.6 percentage points Average review time per document 15.6 minutes 6.1 minutes Shortened by 9.5 minutes Average time for manual review 8.9 minutes 3.2 minutes Shortened by 5.7 minutes

[0155] Table 1 records the comparison results of 30 contracts awaiting review under the same confirmation criteria. Conventional keyword rule review detected 52 confirmed risk points, while this invention detected 71 confirmed risk points, increasing the performance gap detection rate from 70.3% to 95.9%. The number of missed risk points decreased from 22 to 3. These results demonstrate that forward combination propagation, backward combination propagation, and paired deduction can integrate the performance prerequisites, obligations, and results scattered across different clauses into a single performance path for verification, and the detection results do not depend on the appearance of a single risk keyword.

[0156] The number of false alarm risk points has been reduced from 17 to 5. After inserting a protection relationship, the number of undeducted relationship messages needs to be reduced, and frozen protection relationships also need to undergo reverse withdrawal verification. Similar clauses that do not participate in path closure are withdrawn during the insertion phase. Protection relationships that can be restored to closure but do not cause the recurrence of undeducted relationship messages after withdrawal are not identified as necessary protection relationships. Two screening processes reduce the risk of erroneous judgments caused solely by textual similarity.

[0157] The accuracy rate of risk basis positioning is calculated as the ratio of correctly positioned risk points to correctly detected risk points; the direct usability rate of modified content is calculated as the ratio of modified content that can be used without changing the obligor, performance conditions, and liability outcomes to correctly detected risk points. The accuracy rate of risk basis positioning increased from 75.0% to 97.2%. This invention traces the historical performance path along the necessary protection relationship, outputting the source contract number, clause number, clause node, and cross-clause rights and obligations, allowing reviewers to directly verify the source of risk. The direct usability rate of modified content increased from 57.7% to 87.3%. The modified text comes from clause nodes at both ends of the necessary protection relationship and is compared with the contract under review according to the same performance object and the same clause category, ensuring that the risk basis corresponds to the written content.

[0158] The average time for a single review decreased from 15.6 minutes to 6.1 minutes, and the average time for manual review decreased from 8.9 minutes to 3.2 minutes. The reduction in time came from the automatic extraction of historical performance paths, clustering of performance gaps, screening of protection relationships, confirmation of necessary protection relationships, and retrospective analysis of risk evidence.

[0159] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for risk review of contract terms based on knowledge graphs and cluster analysis, characterized in that, Includes the following steps: Analyze historical contracts, connect cross-clause rights and obligations, and form a knowledge graph of historical contracts; Historical performance paths are extracted from the historical contract knowledge graph, and forward and backward combination propagation are performed through the Combinatorial Graph Convolutional Network (CompGCN). Pairing deduction is performed on the relationship messages formed by forward and backward combination propagation according to the fulfillment location. Clustering is performed on the historical fulfillment paths according to the arrangement of the undeducted relationship messages to form complete closed clusters and fulfillment gap clusters. The pending contract is analyzed to form a pending performance path. Forward combination propagation, backward combination propagation and pairing deduction are performed on the pending performance path. The cluster affiliation of the pending performance path is determined based on the arrangement of the undeducted relational messages. When a pending performance path is included in a performance gap cluster, the first unclosed position is marked according to the first undeducted relationship message. The protection relationship between the nodes on both sides of the first unclosed position is extracted from the complete closed cluster and inserted into the pending performance path to form a trial path. The trial path is repeatedly subjected to forward combination propagation, reverse combination propagation and pairing deduction, freezing protection relationships that reduce the number of undeducted relationship messages, and withdrawing protection relationships that do not reduce the number of undeducted relationship messages, until there are no undeducted relationship messages in the trial path; Recall frozen protection relationships in the reverse order of insertion, identify protection relationships that cause the recurrence of undeducted relationship messages as necessary protection relationships, and generate risk levels, risk basis, and modification content based on necessary protection relationships.

2. The method for risk review of contract terms based on knowledge graphs and cluster analysis according to claim 1, characterized in that, The formation of the historical contract knowledge graph specifically includes: Historical contracts are broken down into clause nodes, and clause nodes corresponding to the performance prerequisites, performance obligations, and performance results distributed across different clauses are extracted according to the same performance object. Connect the corresponding clause nodes in the order of performance premises, performance obligations, and performance results to form cross-clause rights and obligations relationships; A knowledge graph of historical contracts is formed based on cross-clause rights and obligations.

3. The method for risk review of contract terms based on knowledge graphs and cluster analysis according to claim 2, characterized in that, The execution of forward combination propagation and reverse combination propagation specifically includes: Using the node representations in the historical performance path and the relationship representations of cross-clause rights and obligations as shared inputs, the node representations and relationship representations are combined along the node order of the historical performance path to generate positive relationship messages; Reverse the direction of cross-clause rights and obligations to generate a reverse relationship representation. Combine the node representation and the reverse relationship representation along the reverse node order of the historical performance path to generate a reverse relationship message. Bind forward and reverse relationship messages to the same fulfillment position.

4. The method for reviewing contract terms based on knowledge graphs and cluster analysis according to claim 3, characterized in that, The execution of paired deduction specifically includes: Establish message pairs between forward and reverse relationship messages that are bound to each other, based on the condition that the relationship type is consistent and the node order is reversed. The relationship direction of the reverse relationship message in the reverse relationship pair is calculated, and the difference between the forward relationship message and the reverse relationship message after the reverse relationship direction is calculated dimension by dimension. Message pairs whose differences in each dimension meet the preset consistency conditions are paired and deducted; message pairs whose differences in each dimension do not meet the preset consistency conditions are retained as undeducted relationship messages. Arrange the undeducted relationship messages according to their corresponding performance positions to form an arrangement status of the undeducted relationship messages.

5. The method for reviewing contract terms based on knowledge graphs and cluster analysis according to claim 4, characterized in that, The clustering process specifically includes: Historical fulfillment paths that do not contain undeducted relationship messages are grouped into complete closed clusters; Historical fulfillment paths with undeducted relationship messages are grouped according to the fulfillment position corresponding to the first undeducted relationship message. Historical fulfillment paths with the same arrangement of undeducted relationship messages within the same group are grouped into the same fulfillment gap cluster.

6. The method for risk review of contract terms based on knowledge graphs and cluster analysis according to claim 5, characterized in that, The determination of the cluster affiliation of the pending performance path specifically includes: If there are no undeducted relationship messages in the pending performance path, the pending performance path will be classified into a complete closed cluster; When there are undeducted relationship messages in the pending performance path and there are performance gap clusters that meet the inclusion criteria, the pending performance path will be included in the performance gap cluster with the same performance position as the first undeducted relationship message and the same arrangement of the undeducted relationship messages. If there are undeducted relationship messages in the pending performance path and there is no performance gap cluster that meets the inclusion criteria, a new performance gap cluster will be formed using the pending performance path.

7. The method for reviewing contract terms based on knowledge graphs and cluster analysis according to claim 6, characterized in that, The formation of the trial-matching path specifically includes: Mark the fulfillment position corresponding to the first unclosed relationship message as the first unclosed position; Extract protection relationships from the historical performance paths in the complete closed cluster that connect the nodes on both sides of the first unclosed position and whose relationship direction is consistent with the node order of the performance path to be reviewed; The protection relationship is inserted into the pending performance path according to its corresponding historical performance position, forming a trial cooperation path.

8. The method for reviewing contract terms based on knowledge graphs and cluster analysis according to claim 7, characterized in that, The process of iteratively performing forward combination propagation, backward combination propagation, and pairing deduction on the trial-and-match path specifically includes: For each trial path after inserting a protection relationship, perform forward combination propagation, reverse combination propagation, and pairing deduction, and compare the number of undeducted relationship messages before and after insertion; Freeze protected relationships that reduce the number of undeducted relationship messages, and withdraw protected relationships that do not reduce the number of undeducted relationship messages; The frozen protection relationships are retained, and the trial path is used to continue the loop until there are no undeducted relationship messages in the trial path.

9. The method for risk review of contract terms based on knowledge graphs and cluster analysis according to claim 8, characterized in that, The determination of the protection relationship that causes the reproduction of the undeducted relationship message as a necessary protection relationship specifically includes: After the trial path no longer contains messages with undeducted relationships, the frozen protection relationships are withdrawn one by one in the reverse order of insertion. After each withdrawal, forward combination propagation, backward combination propagation, and pairing deduction are performed on the trial path; If the undeducted relationship message does not reappear, the protection relationship remains in a withdrawn state. If the undeducted relationship message reappears, the withdrawn protection relationship is restored, and the restored protection relationship is determined as a necessary protection relationship, until all frozen protection relationships have been withdrawn and paired for deduction.

10. The method for reviewing contract terms based on knowledge graphs and cluster analysis according to claim 9, characterized in that, The generation of risk levels, risk basis, and modification content specifically include: The risk level is determined based on the number of necessary protection relationships and the corresponding performance location of those relationships. Tracing back the source historical performance path along the necessary protection relationship to the complete closed cluster, the clause nodes and cross-clause rights and obligations relationships in the source historical performance path are identified as the basis for risk assessment; The necessary protection relationships are mapped to the corresponding performance positions in the performance path to be reviewed, and the rights and obligations corresponding to the necessary protection relationships are written into the contract clauses of the contract to be reviewed associated with the corresponding performance positions, thus forming the revised content.

Images