Multi-stage large model hybrid retrieval-based bank contract compliance risk intelligent review method

Abstract

This invention relates to the fields of artificial intelligence and fintech, specifically to a multi-stage, large-scale model-based intelligent review method for bank contract compliance risks. The method includes: receiving contracts to be reviewed, configuration parameters, and specification documents to construct a review dataset; parsing the contracts using a large language model to extract necessary risk elements containing constraints such as legal obligations and effective dates, generating review points; performing multi-dimensional searches based on these points to obtain candidate evidence and constructing an evidence association graph; verifying the timeliness and scope of the evidence, selecting complete evidence groups with the minimum number of entries while meeting constraints on the upper and lower limits of length and coverage; extracting factual units based on the evidence groups to determine whether the risk elements are valid; performing evidence back-check verification, and finally outputting a risk conclusion package or a pending review marker and location information; thereby constraining the review conclusions to the scope of locatable and verifiable evidence.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and fintech, specifically to a method for intelligent review of bank contract compliance risks using a multi-stage, large-scale model hybrid retrieval system. Background Technology

[0002] In the compliance review of bank contracts, contract terms, related attachments and internal and external regulatory basis are usually scattered in different documents, and there are complex cross-references and version changes;

[0003] Existing review techniques mainly rely on keyword matching, text similarity calculation, or human experience for risk screening and evidence screening; however, such a single text comparison method cannot effectively analyze the structural relationships across documents, and it is difficult to simultaneously determine the complete coverage of risk elements and the effective applicability of normative basis.

[0004] Technical defects can lead to inaccurate identification of standard versions and broken chains of evidence, resulting in insufficient basis for risk assessment and inability to effectively trace review conclusions, thus reducing the accuracy and reliability of bank contract compliance reviews. Summary of the Invention

[0005] The purpose of this invention is to propose a multi-stage, large-scale model hybrid retrieval-based intelligent review method for bank contract compliance risks. This method avoids the problems of insufficient risk assessment basis, inaccurate application of norms, or difficulty in retrospective review conclusions caused by existing review methods, such as single retrieval methods, inaccurate identification of normative versions, and incomplete evidence association. Furthermore, it can simultaneously determine the complete coverage of necessary risk elements and the effective applicability of normative basis. By extracting risk elements, constructing evidence association diagrams, and verifying execution timeliness and scope of application, a complete evidence set is selected, and the final conclusion is constrained within the scope of locatable and verifiable evidence, thereby improving the accuracy and reliability of bank contract compliance risk review.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] A multi-stage, large-scale model-based intelligent review method for bank contract compliance risks, executed by a compliance review server, includes:

[0008] Receive contract documents to be reviewed, along with configuration parameters including the review base date, upper limit of length, lower limit of coverage, lower limit of risk establishment, and lower limit of effective application.

[0009] Obtain the standard documents and index data, and construct the review dataset together with the contract documents to be reviewed;

[0010] The contract documents to be reviewed are analyzed by a large language model to obtain contract terms and extract risk elements. Risk elements with legal obligations, authorization boundaries and effective time constraints are classified into necessary risk elements to generate review points.

[0011] Based on the review criteria, keyword, content similarity, and clause relationship searches were performed on the review dataset to obtain candidate evidence. The candidate evidence was then used as evidence items to construct an evidence association diagram.

[0012] The validity period and scope of application of the evidence items are verified, the evidence score and text length are calculated, and the coverage score is determined based on the degree of matching between the evidence items and the necessary risk elements.

[0013] Under the constraints that the sum of the text lengths of the evidence items does not exceed the upper limit of the length limit and the coverage score reaches the lower limit of the coverage, the complete evidence group is selected from the evidence association diagram with the goal of minimizing the number of items.

[0014] Based on the complete evidence set, the risk elements for the execution of factual units and the judgment of normative application are obtained to determine the risk elements for the establishment of risk.

[0015] For risk elements that establish a risk, perform evidence review and verification to output a risk conclusion package, or output a pending review mark and location information.

[0016] Preferably, the contract documents to be reviewed include loan contracts, guarantee contracts, mortgage contracts, credit agreements, supplementary agreements, application forms and attachments;

[0017] The standard documents include statutory norms and regulations, regulatory provisions, industry standards, internal systems and authorization documents. The standard documents include standard clauses, which contain the standard subject identifier, effective date, expiration date, scope of applicable institutions, applicable business types, applicable geographical scope and applicable product scope.

[0018] The index data includes keyword indexes, content vector indexes, clause relationship indexes, and specification version indexes.

[0019] Preferably, extracting risk elements includes: extracting elements of the parties, actions, conditions, time limits, amounts or proportions, liabilities or obligations, and exceptions or exemptions from the contract terms;

[0020] If any element is not present in the contract terms, the missing elements shall be extracted and supplemented in the following order: the referenced terms in the contract document to be reviewed, the superior terms containing the contract terms, and the attached mapping terms associated with the attachment identifier.

[0021] Based on the supplemented elements, determine the risk event categories, risk elements to be verified, business scope, time constraints, and required evidence types, and classify the necessary risk elements according to legal obligations, authorization boundaries, and effective time constraints.

[0022] Preferably, candidate evidence is obtained by performing keyword, content similarity, and clause relationship searches on the review dataset based on the review criteria, including:

[0023] A keyword set is generated based on the review criteria, and the first candidate set is obtained by retrieving the contract documents and specification documents to be reviewed from the review dataset using the keyword set.

[0024] Based on the contract terms and review points, a content retrieval vector is generated using a preset text encoding model. The contract document and specification document to be reviewed are then converted into candidate clause vectors. A second candidate set is obtained by searching the content similarity between the content retrieval vector and the candidate clause vector.

[0025] Based on the hierarchical inclusion relationship, clause reference relationship, condition triggering relationship, exception restriction relationship, attachment mapping relationship, normative basis reference relationship, and business correspondence between the contract and authorization document, the structure of the first and second candidate sets is expanded to obtain candidate evidence.

[0026] Preferably, the timeliness and scope of application verification includes: obtaining a set of specification versions for the same specification topic identifier;

[0027] The effective date and expiration date of the regulatory clauses shall be determined based on the review reference date;

[0028] If the effective date is earlier than or equal to the review reference date, and the expiry date is empty or later than or equal to the review reference date, the validity period mark is determined to be valid; if the effective date is later than the review reference date, or the expiry date is earlier than the review reference date, the validity period mark is determined to be invalid.

[0029] If a specification version exists but lacks an effective date, the validity period will be marked as pending verification, and the specification version will not be used as an independent basis for determination.

[0030] Preferably, the evidence association diagram includes contract clause nodes, attachment nodes, regulatory nodes, authorization basis nodes, related clause nodes, and risk element nodes as evidence items;

[0031] Each evidence entry preserves the source document identifier, clause text, original text location information, clause hierarchy path, statute of limitations parameters, scope of application parameters, and evidence function tags; evidence function tags include fact support tags, fact rebuttal tags, and verification tags.

[0032] In the evidence association diagram, the association edges represent reference relationships, supplementary relationships, superior-subordinate relationships, conditional triggering relationships, exception exclusion relationships, normative application relationships, and factual support relationships.

[0033] Preferably, calculating the evidence score includes:

[0034] The similarity of the content of the evidence items, the matching degree of keywords, the structural relevance, the timeliness and validity, the matching degree of the applicable scope, and the coverage of risk elements are used as evaluation factors;

[0035] Each evaluation factor is multiplied by its corresponding preset weight and summed to obtain the initial score;

[0036] If there are conflicts in the statute of limitations, scope of application, or direction of support and rebuttal of evidence items, a preset conflict deduction value is determined; if there are no conflicts, the preset conflict deduction value is 0.

[0037] The evidence score is obtained by subtracting the preset conflict deduction value from the initial score.

[0038] Preferably, screening a complete set of evidence includes: for necessary risk elements, constraining the complete set of evidence to contain at least one evidence item with an evidence score greater than zero to cover the necessary risk elements;

[0039] For the combined elements that require joint proof among the necessary risk elements, the constraint shall be jointly covered by at least two different evidence items with supplementary or superior-subordinate relationship edges, and the cumulative coverage score of the jointly covered items shall reach the lower limit of coverage.

[0040] Preferably, the judgment of risk requirements and regulatory applicability includes: extracting factual units from the evidence items of the complete evidence group, and binding each factual unit to the corresponding evidence item and the original text of the contract clause;

[0041] The applicability matching degree of the standard is determined based on the timeliness verification results and the scope of application verification results corresponding to the evidence items.

[0042] If the evidence score reaches the lower limit for risk establishment, the matching degree of normative application reaches the lower limit for effective application, and there is a complete path in the evidence association diagram that starts from the contract clause node, is connected to the related clause node through the reference relationship edge, is connected to the normative node through the normative application relationship edge, and is connected to the risk element node through the factual support relationship edge, then the risk conclusion package is output.

[0043] Otherwise, output a flag indicating that it needs to be reviewed.

[0044] Preferably, the evidence review and verification includes: reverse locating the corresponding evidence item for each risk element that establishes the risk;

[0045] If the necessary risk elements cannot be proven to be covered by the evidence items, or the normative clauses are not in effect, or the scope of application of the normative clauses is below the effective application limit, then the output of the risk conclusion package and the generation of a pending review mark are prohibited.

[0046] If there are supporting and rebuttal evidence in the evidence item, and the supporting and rebuttal evidence come from different versions of the standard, the version of the standard that is valid on the review benchmark date shall be given priority.

[0047] When different versions of the standard are all valid, they shall be compared in the following order of priority: applicable organization scope, applicable business type, applicable geographical scope, and applicable product scope, and the standard clause with the smallest applicable scope shall be adopted first.

[0048] If the comparisons are still consistent and conflicting in direction after priority order, an evidence conflict status marker is output, and location information containing conflict clauses and comparison dimensions is generated.

[0049] The beneficial effects of this invention are:

[0050] 1. This invention parses the contract document to be reviewed using a large language model, extracting combinations of main elements, behavioral elements, etc. If any element does not appear in the contract terms, it is extracted and supplemented in the following order: the cited clause, the superior-level clause containing the contract terms, and the attached mapping clause associated with the attachment identifier. This mechanism effectively addresses the reality that contract terms, related attachments, and normative basis are usually scattered in different documents, and solves the defect of existing technology that is difficult to determine the complete coverage of necessary risk elements.

[0051] 2. Based on the key points of review, this invention performs keyword, content similarity, and clause relationship retrieval on the review dataset to obtain candidate evidence, and uses these as evidence items to construct an evidence association graph containing related edges such as reference relationship edges, supplementary relationship edges, and normative application relationship edges. This method overcomes the limitations of traditional review that relies on a single keyword or text similarity to screen evidence, and connects scattered clauses through structural expansion, solving the problems of incomplete evidence association and inconsistent clause descriptions leading to retrieval omissions in the prior art.

[0052] 3. In risk review, this invention verifies the validity period and scope of application of evidence items, and determines the validity period mark by combining the review benchmark date; when encountering supporting and rebuttal evidence from different versions of the standard, it compares them in order of priority: scope of applicable institutions, applicable business types, applicable geographical scope, and applicable product scope; this measure clarifies the valid applicable standard benchmark and solves the background technical problem of inaccurate identification of version changes and differences in the scope of application during the review process, which leads to inappropriate application of the standard. Attached Figure Description

[0053] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0054] Figure 1 This is a flowchart of an intelligent review method for bank contract compliance risks using a multi-stage, large-model hybrid retrieval approach. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 A multi-stage, large-scale model-based intelligent review method for bank contract compliance risks, executed by a compliance review server, includes:

[0057] Receive contract documents to be reviewed, along with configuration parameters including the review base date, upper limit of length, lower limit of coverage, lower limit of risk establishment, and lower limit of effective application.

[0058] Obtain the standard documents and index data, and construct the review dataset together with the contract documents to be reviewed;

[0059] The contract documents to be reviewed are analyzed by a large language model to obtain contract terms and extract risk elements. Risk elements with legal obligations, authorization boundaries and effective time constraints are classified into necessary risk elements to generate review points.

[0060] Based on the review criteria, keyword, content similarity, and clause relationship searches were performed on the review dataset to obtain candidate evidence. The candidate evidence was then used as evidence items to construct an evidence association diagram.

[0061] The validity period and scope of application of the evidence items are verified, the evidence score and text length are calculated, and the coverage score is determined based on the degree of matching between the evidence items and the necessary risk elements.

[0062] Under the constraints that the sum of the text lengths of the evidence items does not exceed the upper limit of the length limit and the coverage score reaches the lower limit of the coverage, the complete evidence group is selected from the evidence association diagram with the goal of minimizing the number of items.

[0063] Based on the complete evidence set, the risk elements for the execution of factual units and the judgment of normative application are obtained to determine the risk elements for the establishment of risk.

[0064] For the risk elements that establish a risk, perform evidence review and verification to output a risk conclusion package, or output a mark and location information to be reviewed;

[0065] In a preferred embodiment of the present invention, the contract documents to be reviewed include loan contracts, guarantee contracts, mortgage contracts, credit agreements, supplementary agreements, application forms, and attachments;

[0066] The standard documents include statutory norms and regulations, regulatory provisions, industry standards, internal systems and authorization documents. The standard documents include standard clauses, which contain the standard subject identifier, effective date, expiration date, scope of applicable institutions, applicable business types, applicable geographical scope and applicable product scope.

[0067] The index data includes keyword indexes, content vector indexes, clause relationship indexes, and specification version indexes;

[0068] Risk extraction criteria include:

[0069] Extract the following elements from the contract terms: subject elements, behavioral elements, condition elements, time limit elements, amount or proportion elements, liability or obligation elements, and exception or exemption elements;

[0070] If any element is not present in the contract terms, the missing elements shall be extracted and supplemented in the following order: the referenced terms in the contract document to be reviewed, the superior terms containing the contract terms, and the attached mapping terms associated with the attachment identifier.

[0071] Based on the supplemented elements, determine the risk event categories, risk elements to be verified, business scope, time constraints, and required evidence types, and classify the necessary risk elements according to legal obligations, authorization boundaries, and effective time constraints.

[0072] In specific bank credit contract compliance reviews, the compliance review server organizes the main text of the contract, supplementary agreements, application forms, attachments, legal norms and regulations, regulatory provisions, internal bank rules and authorization documents around the same contract terms to be reviewed, and forms an evidentiary basis that can be used for risk assessment within the constraints of limited space.

[0073] The banking system submits the loan contract, guarantee contract, mortgage contract, credit agreement, supplementary agreement, application form and attachments corresponding to a certain credit business to the compliance review server, along with the review benchmark date, upper limit of length, lower limit of coverage, lower limit of risk establishment and lower limit of effective application.

[0074] After receiving the data, the server organizes the batch of contract documents together with the pre-stored specification documents and index data into a review dataset corresponding to a review task. The specification documents contain the specification subject identifier, effective date, expiration date, applicable organization scope, applicable business type, applicable geographical scope, and applicable product scope for each specification clause. The index data contains keyword index, content vector index, clause relationship index, and specification version index for subsequent use.

[0075] In this implementation, the compliance review server calls a preset large language model to perform structured parsing of the contract document to be reviewed, and divides the main clauses, heading clauses, reference clauses and appendix clauses into locatable contract clauses to be reviewed.

[0076] The large language model employs a decoder network based on a transformer architecture, containing 12 self-attention layers and hidden layer feature vectors. The input consists of clause text arranged in document order, along with its title, hierarchical path, document identifier, and adjacent clause identifiers. This is transformed into a feature vector with a dimension of [missing information - likely a specific dimension]. The input tensor, Given the length of the text sequence, after the hidden layer sequence is output through forward propagation with fixed parameters pre-trained, the sequence is then labeled and decoded using a conditional random field classification head. The output is a set of structured fields with fixed dimensions that correspond one-to-one with each contract clause to be reviewed.

[0077] The structured field set includes at least the clause identifier, subject element, behavior element, condition element, time limit element, amount or proportion element, liability or obligation element, exception or exemption element, reference clause identifier, and attachment identifier;

[0078] The server performs field validation on the model output: if the field name does not belong to the preset field set, the field value cannot be located in the original text of the clause or related clauses, or if the same field has multiple contradictory values, the field is marked as pending validation and will not be used directly as an extracted element.

[0079] For the same clause, the fields output by the model are saved in a fixed order: subject, action, condition, term, amount or proportion, liability or obligation, exception or exemption; when a field is missing, it is left blank without changing the position of the remaining fields, so that the parsing results of different clauses have the same data structure.

[0080] For any of the contract clauses to be reviewed, the server extracts the subject elements, behavioral elements, condition elements, time limit elements, amount or proportion elements, liability or obligation elements, and exception or exemption elements from the original text of the clause, the clause title, the clause hierarchy path, and adjacent clauses.

[0081] The server performs original text positioning for each feature output by the model, binding the feature value with document identifier, page number or paragraph number, and character start and end position; if the feature value output by the model fails the original text positioning verification, the feature value is deleted and left in a pending verification state.

[0082] If an element does not appear in the current clause, it will not be marked as missing directly. Instead, it will be extracted and supplemented in the following order: the referenced clause, the superior clause containing the current clause, and the attached mapping clause associated with the attachment identifier.

[0083] After each layer of supplementation is completed, the server performs clause identifier matching and original text location verification; the supplementary element is only written into the element set if the identifier of the supplementary clause can be found in the current document's reference relationship or attachment mapping relationship.

[0084] If the same element yields different values ​​from multiple related clauses, the priority value is retained in the order of the cited clause, the superior clause, and the attached mapping clause, while the remaining values ​​are saved as conflict values ​​for subsequent verification.

[0085] Based on this, the server determines the risk category, risk elements to be verified, business scope, time constraints and required evidence type corresponding to the clause according to the preset element judgment table;

[0086] The element determination table takes as input a combination of subject elements, behavioral elements, condition elements, time limit elements, amount or proportion elements, liability or obligation elements, and exception or exemption elements, and outputs the risk event category, risk elements to be verified, business scope, time constraints, and required evidence type.

[0087] Among them, when both the subject element and the behavioral element exist, and the liability or obligation element also exists, a risk element of legal obligation or authorization boundary is generated; when the term element and the statement of effective, invalid or duration exist simultaneously, a risk element of effective time restriction is generated.

[0088] When an amount or percentage element exists, it is used as a quantity limit field for risk requirements; when an exception or exemption element exists, it is used as an exclusion condition field for risk requirements; if a necessary field is still pending verification, the corresponding risk requirement is marked as pending verification and not marked as satisfied.

[0089] The server then categorizes the risk elements that have legal obligations, authorization boundaries, and effective time restrictions into necessary risk elements, and writes the risk element identifier, components, constraints, scope of application, time constraints, and required evidence types into the intermediate record of the current review task, which serves as the direct basis for subsequent retrieval and screening, avoiding the need for subsequent modules to revisit the original full contract text.

[0090] The review points are directly generated from the constituent fields of the necessary risk elements; for each necessary risk element, the server sequentially reads the risk event category, subject element, behavioral element, condition element, time limit element, amount or proportion element, liability or obligation element, exception or exemption element, business scope, time constraint and required evidence type.

[0091] Fields with values ​​are concatenated in a fixed order to form review points, while fields without values ​​retain their field identifiers and are given a pending verification status. The review points generated in this way simultaneously save the risk element identifier and the original text location information. Subsequent candidate evidence must be able to correspond to at least one component of the risk element in order to be included in the coverage calculation of the risk element.

[0092] For an example clause, if the subject element is the borrower, the behavior element is the submission of materials, the term element is the credit period, the liability or obligation element is continuous extraction, and the exception element is empty, then the generated review points will include at least four search fields: borrower, continuous submission of materials, credit period, and continuous obligation; if the term element only exists in the referenced clause, then the term field in the review points will also store the location information of the referenced clause.

[0093] During the retrieval phase, the server generates review points based on necessary risk elements, and performs keyword retrieval, content similarity retrieval, and clause relationship retrieval sequentially in the same review dataset accordingly; keyword retrieval is used to recall candidate clauses with clear terminology from contract documents and specification documents in advance;

[0094] Content similarity retrieval is used to supplement candidate clauses that are inconsistent in expression but similar in meaning; clause relationship retrieval, based on the first two types of results, continues to expand along clause hierarchical inclusion relationships, clause reference relationships, condition triggering relationships, exception restriction relationships, attachment mapping relationships, normative basis reference relationships, and business correspondence relationships between contracts and authorization documents to obtain a set of candidate evidence.

[0095] The expanded candidate evidence is not directly included in the risk assessment, but is first written into the evidence association diagram as evidence items; each evidence item retains the original text content, source identifier and location information, and performs timeliness verification and scope of application verification respectively;

[0096] The server calculates the evidence score and text length of each piece of evidence, and determines the coverage score by combining the degree to which the evidence piece supports each necessary risk element.

[0097] Then, with the goal of minimizing the number of evidence items, and provided that the sum of the lengths of the individual evidence texts of the selected evidence items does not exceed the upper limit of the length limit, and the coverage score of the necessary risk elements reaches the lower limit of the coverage, a complete evidence group is selected from the evidence association diagram.

[0098] After forming a complete set of evidence, the server extracts factual units from it one by one and binds each factual unit to the corresponding evidence item and the original text position. Based on the factual unit, the server performs risk element judgment and normative application judgment respectively to obtain the risk elements for determining the establishment of risk.

[0099] The server performs evidence back-check verification on the risk elements for determining the establishment of risk, checking each element in reverse to see if it has corresponding evidence coverage, whether the corresponding standard is valid, and whether the corresponding scope of application has reached the lower limit of the effective application threshold; if the verification results meet the requirements, a deterministic risk conclusion data packet is output.

[0100] If there is insufficient evidence, the validity of the standard cannot be confirmed, or the scope of application does not meet the requirements, then output the status mark pending review and the location information for manual review.

[0101] This location information points to at least the original text of the contract clauses to be reviewed and the location of related regulatory clauses, so as to support subsequent manual review to directly access the corresponding clauses without having to search again;

[0102] In the case of missing elements in this implementation, if a necessary element cannot be extracted after parsing the contract clauses, and no valid result is obtained in the order of supplementing the cited clauses, higher-level clauses, and appendix mapping clauses, the risk element corresponding to that element remains in a state of pending verification and continues to participate in the subsequent candidate evidence retrieval, but it shall not be directly treated as a fulfilled element.

[0103] If a normative clause exists in the candidate evidence set but its time field or scope of application field is incomplete, the clause may be retained in the candidate evidence set to participate in the coverage calculation, but it shall not be used as the sole basis for a definitive risk conclusion during the evidence review and verification stage.

[0104] The above mechanism is used to overcome the technical defects in bank contract review, such as the scattered description of the main text and attachments, mutual reference between different documents, and the coexistence of standard versions. If missing items are directly excluded in the early stage, it will lead to the omission of necessary risk elements.

[0105] However, if the retrospective review restriction is not implemented at the final conclusion stage, inapplicable grounds may be introduced. Through the above-mentioned processing, under the same credit contract review, the original contract text, related clauses and regulatory grounds can be continuously circulated under the same review task, and the final conclusion can be constrained within the scope of locatable and verifiable evidence.

[0106] Based on the review criteria, candidate evidence was obtained by performing keyword, content similarity, and clause relationship searches on the review dataset, including:

[0107] A keyword set is generated based on the review criteria, and the first candidate set is obtained by retrieving the contract documents and specification documents to be reviewed from the review dataset using the keyword set.

[0108] Based on the contract terms and review points, a content retrieval vector is generated using a preset text encoding model. The contract document and specification document to be reviewed are then converted into candidate clause vectors. A second candidate set is obtained by searching the content similarity between the content retrieval vector and the candidate clause vector.

[0109] Based on the hierarchical inclusion relationship, clause reference relationship, condition triggering relationship, exception restriction relationship, attachment mapping relationship, normative basis reference relationship, and business correspondence between the contract and authorization document, the first and second candidate sets are structurally expanded to obtain candidate evidence;

[0110] The timeliness and scope of application verification includes: obtaining a set of specification versions for the same specification topic identifier;

[0111] The effective date and expiration date of the regulatory clauses shall be determined based on the review reference date;

[0112] If the effective date is earlier than or equal to the review reference date, and the expiry date is empty or later than or equal to the review reference date, then the validity period mark is determined to be valid; if the effective date is later than the review reference date, or the expiry date is earlier than the review reference date, then the validity period mark is determined to be invalid.

[0113] If a specification version exists but lacks an effective date, the validity period will be marked as pending verification, and the specification version will not be used as an independent basis for determination.

[0114] The evidence association diagram includes contract clause nodes, attachment nodes, regulatory nodes, authorization basis nodes, related clause nodes, and risk element nodes that serve as evidence items;

[0115] Each evidence entry preserves the source document identifier, clause text, original text location information, clause hierarchy path, statute of limitations parameters, scope of application parameters, and evidence function tags; evidence function tags include fact support tags, fact rebuttal tags, and verification tags.

[0116] In the evidence association diagram, the association edges represent citation relationships, supplementary relationships, superior-subordinate relationships, conditional triggering relationships, exception exclusion relationships, normative application relationships, and factual support relationships.

[0117] The calculation of evidence scores includes: obtaining the content similarity of evidence items, keyword matching degree, structural relevance, timeliness and validity, scope of application matching degree, and risk element coverage as evaluation factors;

[0118] Each evaluation factor is multiplied by its corresponding preset weight and summed to obtain the initial score;

[0119] If there are conflicts in the statute of limitations, scope of application, or direction of support and rebuttal of evidence items, a preset conflict deduction value is determined; if there are no conflicts, the preset conflict deduction value is 0.

[0120] The evidence score is obtained by subtracting the preset conflict deduction value from the initial score.

[0121] After identifying the necessary risk elements, a set of candidate evidence is formed around these necessary risk elements, and the following steps are taken to further complete the screening of normative timeliness, evidence organization, and evidence score calculation;

[0122] The server establishes an independent retrieval task context for each necessary risk element and processes them in the order of keyword recall, vector recall, and relation expansion. After the retrieval task enters the running state, it first generates hierarchical keywords from the subject element, behavior element, condition element, time limit element, amount or proportion element, responsibility or obligation element, and exception or exemption element.

[0123] The contract terms and key points to be reviewed are then encoded into content retrieval vectors in a fixed order. After each candidate term is encoded, it will only be compared for similarity if the vector dimension is consistent with the content retrieval vector, the source location is complete, and it is not marked as parsing failure. Candidate terms with inconsistent dimensions or missing location are transferred to the verification queue, which does not affect the processing of other candidate terms.

[0124] Keyword retrieval is performed in the order of first-level keywords followed by restrictive keywords; subject elements and behavioral elements are used as first-level keywords, condition elements and time limit elements are used as constraint keywords, and amount or proportion elements, responsibility or obligation elements, and exception or exemption elements are used as restrictive keywords.

[0125] When a candidate clause matches at least one primary keyword and one constraint keyword or restriction keyword, it is directly included in the first candidate set.

[0126] Clauses that only hit the restricted keywords are retained as low-priority candidates and will only be included in the supplementary search if the corresponding necessary risk elements have not yet reached the preset recall quantity; the keyword matching degree is calculated according to the level of the hit words, the number of consecutive occurrences, and the position in the original text, and repeated hits of the same term are not counted repeatedly;

[0127] Content similarity retrieval uses a pre-defined bidirectional long short-term memory network as a text encoding model to process text fragments with a fixed field order. The word embedding layer dimension of the encoding model is set to 256, and the number of neurons in the hidden layer is set to 512. Each text fragment is first concatenated in the order of subject, behavior, condition, time limit, responsibility or obligation, exception or exemption.

[0128] Missing fields are filled with null values ​​of fixed length, without being filled by adjacent fields; the encoding model outputs a real vector of fixed dimensions, and the server performs norm and dimension checks on the vector. If the norm of any vector is zero or the dimensions are inconsistent, the corresponding candidate clause is marked as pending verification and the similarity calculation is skipped in this round.

[0129] The content similarity of the verified vectors is calculated using cosine similarity, and the formula is as follows:

[0130]

[0131] In the formula, For content retrieval vectors, For candidate clause vectors, This represents the dot product operation of vectors. The norm of a vector;

[0132] The result is mapped to the 0 to 1 interval using the following linear scaling formula:

[0133]

[0134] In the formula, This is the scaled similarity. The cosine similarity before the aforementioned mapping is used. Candidate terms that reach the similarity screening threshold are entered into the second candidate set, while terms that are below the threshold are not entered into the candidate set in this round.

[0135] If the number of valid candidate clauses is lower than the minimum recall required for the necessary risk element, the server will successively broaden the synonym range and lower the vector screening threshold, but a minimum threshold boundary will be set for each adjustment; if the minimum boundary is still insufficient, the current result will be retained and the necessary risk element will be marked as insufficient recall.

[0136] The server places the first and second candidate sets into a bounded candidate processing window and removes duplicates according to the source document and original text position. When a candidate clause enters the window, it is marked as pending expansion. After the relationship identification is completed, it is changed to the expanded state. If the relationship type cannot be determined or the referenced clause cannot be located, it is changed to the pending verification state and the expansion along the clause is stopped.

[0137] Relationship expansion is executed sequentially according to the following relationships: clause hierarchy inclusion relationship, clause reference relationship, condition triggering relationship, exception restriction relationship, attachment mapping relationship, standard basis reference relationship, and business correspondence relationship between contract and authorization document;

[0138] For each relationship, an associated edge is established only when the two ends of the relationship can be uniquely matched through clause number, attachment identifier, specification document number, authorization document number, or business attribute.

[0139] Candidate clauses from the same source and in the same original text location will not be added repeatedly; when the expansion depth reaches the preset limit, or when the new clauses fail to increase the necessary risk element coverage in two consecutive rounds, the expansion will be stopped to prevent the relationship chain from growing indefinitely.

[0140] The candidate processing window is executed in batches, with each batch processing only a fixed number of candidate terms. After the current batch completes deduplication and internal status updates, the next batch is released, and asynchronous cross-batch relationship recognition network expansion is executed uniformly after all batches enter the processing window.

[0141] If multiple necessary risk elements trigger the retrieval simultaneously, the server allocates processing resources in descending order of the risk establishment lower limit, ascending order of the required number of evidence types, and ascending order of the current coverage.

[0142] When the candidate window occupied by a single task reaches its limit, its low-priority relationship expansion is suspended, and the keyword and vector retrieval of other tasks is completed first. Any candidate clause retains its source, location, current processing status and the reason for entering the next stage. Only clauses that are in the expanded or verified pending verification state are allowed to enter the evidence association graph.

[0143] After obtaining the set of candidate evidence, the server generates corresponding relation edges based on the clause hierarchy path, explicit clause number reference, condition trigger words, exception trigger words, attachment identifier, specification name or document number, and authorization document number;

[0144] If the same candidate clause satisfies multiple relationship conditions, then each type of relationship edge is retained; after the relationship edge is generated, the candidate clause changes from the pending expansion state to the associated state; if multiple possible targets appear at both ends of the relationship, then no presumed connection is made, but the clause changes to the pending verification state.

[0145] The nodes in the evidence association diagram include at least contract clause nodes, attachment nodes, regulatory nodes, and authorization basis nodes. Each evidence item saves the source document identifier, clause text, original text location information, clause hierarchy path, statute of limitations parameter, scope of application parameter, and evidence function label.

[0146] Clauses that support the establishment of necessary risk requirements are marked with a fact support label, clauses that directly exclude or limit the requirements are marked with a fact refutation label, and clauses that lack an effective date, applicable subject, or complete condition field are marked with a verification pending label.

[0147] The timeliness verification is performed after the relationship expansion is completed at the specification node; the server first obtains the specification version set according to the specification topic identifier, and then compares the review base date, effective date and expiration date one by one;

[0148] If the effective date is earlier than or equal to the review reference date and the expiration date is empty or later than or equal to the review reference date, it is marked as valid; if the effective date is later than the review reference date or the expiration date is earlier than the review reference date, it is marked as invalid.

[0149] When the effective date is missing, it is marked as pending verification; invalid specifications can be retained in the diagram for rebuttal relationship and version conflict analysis, but their validity period is set to 0 and cannot be used as a deterministic basis; when the specification version set is empty, the date cannot be parsed, or there are contradictory dates in the same version, they are all changed to pending verification status, and the corresponding specification node is sent to the manual review and location queue.

[0150] During the evidence score calculation phase, content similarity, keyword matching, structural relevance, timeliness, applicability, and risk element coverage are all converted to a range of 0 to 1; the initial score is determined as follows:

[0151]

[0152] in, This is the initial score. to These represent six evaluation factors: content similarity, keyword matching, structural relevance, timeliness and effectiveness, applicability matching, and risk element coverage. to They are respectively to The corresponding non-negative weights, and satisfying Keyword matching score is determined based on the hit rate of primary keywords, constrained keywords, and restricted keywords.

[0153] The structural relevance is determined based on the shortest effective path between contract clause nodes and candidate clause nodes. Direct reference or normative application relationship takes the first preset relationship value, one-level supplementary or superior-subordinate relationship takes the second preset relationship value, and relationship of two or more levels takes the third preset relationship value. The first preset relationship value is greater than the second preset relationship value, and the second preset relationship value is greater than the third preset relationship value.

[0154] The risk element coverage is calculated based on the number of necessary risk element components covered. Repeated coverage of the same element is counted only once. The number of necessary risk element components covered is divided by the total number of required components preset for the risk element to obtain the risk element coverage value mapped in the range of 0 to 1.

[0155] The scope of application matching degree is compared item by item according to the applicable institution, business, region and product. If there is a clear conflict in any dimension, the value is 0. If the field is missing but there is no conflict, the value to be verified is used.

[0156] If there are conflicts in the time limit, scope of application, or direction of support and rebuttal of evidence items, the server determines the penalty value for each conflict according to the type of conflict, and adds up the penalty values ​​of each conflict and limits them to a preset upper limit to determine the total penalty value; the final evidence score is the initial score minus the total penalty value, and if it is less than 0, it is treated as 0.

[0157] The validity period is set to 1 for valid, below 1 for pending verification, and 0 for expired; items pending verification can participate in the candidate coverage ranking, but cannot meet the lower limit of validity application on their own.

[0158] If any evaluation factor required for the calculation of the evidence score is unavailable, the server will not fill it with the preset maximum value, but will mark the entry as pending verification and calculate it according to the limit value corresponding to the missing factor; after the scoring is completed, the evidence entry will enter the filterable state; entries that have not been scored or are in the parsing failure state will not be included in the complete evidence group.

[0159] Screening a complete set of evidence includes: for necessary risk elements, ensuring that the complete set of evidence contains at least one evidence item with a score greater than zero to cover the necessary risk elements;

[0160] For the combined elements that require joint proof among the necessary risk elements, the constraint shall be jointly covered by at least two different evidence items with supplementary or superior-subordinate relationship edges, and the cumulative coverage score of the joint coverage shall reach the lower limit of coverage.

[0161] The judgment of risk requirements and applicable regulations includes: extracting factual units from the evidence items of the complete evidence set, and binding each factual unit to the corresponding evidence item and the original text of the contract clause;

[0162] The applicability matching degree of the standard is determined based on the timeliness verification results and the scope of application verification results corresponding to the evidence items.

[0163] If the evidence score reaches the lower limit for risk establishment, the matching degree of normative application reaches the lower limit for effective application, and there is a complete path in the evidence association diagram that starts from the contract clause node, is connected to the related clause node through the reference relationship edge, is connected to the normative node through the normative application relationship edge, and is connected to the risk element node through the factual support relationship edge, then the risk conclusion package is output.

[0164] Otherwise, output a flag indicating that it needs to be reviewed;

[0165] Evidence review and verification includes: for each risk element that establishes a risk, locating the corresponding evidence item in reverse;

[0166] If the necessary risk elements cannot be proven to be covered by the evidence items, or the normative clauses are not in effect, or the scope of application of the normative clauses is below the effective application limit, then the output of the risk conclusion package and the generation of a pending review mark are prohibited.

[0167] If there are supporting and rebuttal evidence in the evidence item, and the supporting and rebuttal evidence come from different versions of the standard, the version of the standard that is valid on the review benchmark date shall be given priority.

[0168] When different versions of the standard are all valid, they shall be compared in the following order of priority: applicable organization scope, applicable business type, applicable geographical scope, and applicable product scope, and the standard clause with the smallest applicable scope shall be adopted first.

[0169] If the comparisons are still consistent and conflicting in direction after priority order, an evidence conflict status marker is output, and location information containing conflict clauses and comparison dimensions is generated.

[0170] Based on the existing evidence association diagram and the fact that each evidence item has evidence score, statute of limitations status, and scope of application information, the steps for screening complete evidence groups and generating risk conclusions or results pending review are as follows;

[0171] When a clause in a credit agreement involves multiple elements such as the authorizing entity, the purpose of authorization, the authorization period, and the withdrawal mechanism, selecting only the single evidence item with the highest score may only prove some of the elements and cannot form a complete basis sufficient to support a compliance conclusion.

[0172] Therefore, when the server filters a complete set of evidence from the evidence association diagram, it checks for each necessary risk element whether there is at least one independent evidence item with an evidence score greater than zero that covers that necessary risk element.

[0173] If a necessary risk element is of the type that can be proven by a single element, then the evidence item that meets the above conditions can be included in the candidate scope of the complete evidence group.

[0174] If a certain necessary risk element is a combination element that requires joint proof by multiple elements, it is further required that it be covered by at least two different evidence items, and that these two evidence items have a supplementary relationship or a superior-subordinate relationship. Only after the cumulative coverage score reaches the lower limit of coverage can the set of items be regarded as a valid coverage of the combination element.

[0175] During the screening process, among the candidate evidence groups that meet all necessary risk element coverage constraints, the evidence group with the fewest evidence items is selected first; when the number of evidence items is the same, the evidence group with the smallest sum of individual evidence text lengths and the largest total evidence score is selected first.

[0176] If there is no evidence set that simultaneously meets the upper limit of the length limit and the lower limit of the coverage limit, then a complete evidence set will not be forcibly formed, and the unmet necessary risk elements and the missing evidence types will be written into the information pending review.

[0177] In this embodiment, the screening of the complete evidence group is performed item by item according to the necessary risk elements; the server first establishes a coverage record for each necessary risk element. The coverage record includes at least the risk element identifier, the number of required elements, the number of covered elements, the corresponding evidence item list, and the cumulative coverage score.

[0178] For a single-factor necessary risk element, if a certain evidence item has a coverage score greater than zero for the risk element and a greater evidence score than zero, then the evidence item is marked as an available coverage item.

[0179] For combined requirements, the server first checks the set of elements covered by each evidence item, and then checks whether there are supplementary relationship edges or superior-subordinate relationship edges between these evidence items. Only when the union of the elements of multiple evidence items can cover all the required elements of the combined requirement, the cumulative coverage score of their common coverage is calculated.

[0180] If there are multiple optional item combinations for the same combination requirement, they are sorted in ascending order by the number of evidence items; if the number is the same, they are sorted in ascending order by the sum of the lengths of the individual evidence texts; if the above two are still the same, they are sorted in descending order by the sum of the evidence scores, and the combination with the highest score is selected as the candidate coverage result for that combination requirement.

[0181] If a necessary risk element still has no candidate coverage result that meets the conditions after traversing all evidence items, then the necessary risk element is directly recorded as not fully covered, and the subsequent generation of the deterministic risk conclusion path is terminated.

[0182] To avoid confusion between the cumulative coverage score and the sum of evidence scores in terms of calculation methods, the server stores the two separately: the cumulative coverage score is only used to determine whether a certain necessary risk element has reached the coverage limit, and its calculation object is limited to the elements covered by the necessary risk element;

[0183] The sum of evidence scores is only used to prioritize multiple candidate combinations that have met the coverage criteria, and is not used to replace the lower limit of coverage judgment. Thus, the coverage judgment link and the ranking link are kept separate in the same screening process. The former corresponds to the sufficiency of the requirement coverage, and the latter corresponds to the priority of the candidate group, avoiding misjudging a highly relevant but low coverage evidence combination as a complete evidence group.

[0184] Furthermore, the common coverage of the combined elements is calculated according to the following rules: first, the combined element is broken down into the required set of elements from subject, conduct, condition, time limit, liability or obligation, exception or exemption; then, each candidate evidence item is marked with its covered elements.

[0185] When the union of the covered elements of at least two different evidence items equals the required element set, it is determined that all required elements are covered; the cumulative coverage score is only counted for the covered elements within the required element set, and when the same element is covered repeatedly, it is counted once, and it is not accumulated repeatedly because multiple pieces of evidence support the same element.

[0186] Furthermore, to avoid timing conflicts between coverage judgments and length constraints during the screening of complete evidence sets, the server executes the checks in a fixed order on the same candidate combination:

[0187] First, determine whether the candidate combination covers the corresponding necessary risk elements to the lower limit. Then, calculate whether the sum of the lengths of the individual evidence texts after the candidate combination is included in the current complete evidence group exceeds the upper limit of the length limit. Only when the previous judgment is true and the next judgment is true will the candidate combination be retained as a valid candidate.

[0188] If a candidate combination is excluded due to space limitations after passing the coverage judgment, the server continues to search for the next candidate combination of the same necessary risk element, instead of directly recording the necessary risk element as uncovered.

[0189] By adopting the above mechanism, the coverage judgment, length judgment and priority sorting are executed sequentially at the same time, so as to avoid the same set of evidence items from different judgment stages yielding contradictory screening results due to unclear order.

[0190] Once the complete set of evidence is determined, the server extracts fact units from each of the evidence items in the set and binds each fact unit to the corresponding evidence item and the original text location of the contract clause to be reviewed; the fact unit, along with its original text location information, is written into the current review task record.

[0191] The extraction of fact units employs a named entity recognition algorithm based on conditional random fields to extract sequence labels according to element type, without using a free generation method; for subject elements, the subject name, subject role, and correspondence with the contract clauses to be reviewed are extracted; for behavioral elements, the behavioral type, behavioral object, and behavioral direction are extracted.

[0192] For conditional and time-limited elements, extract the triggering condition, start time, end time, or duration; for liability or obligation elements and exception or exemption elements, extract the constraint content and exclusion content respectively.

[0193] If multiple elements appear in the same evidence entry, the server will split them into multiple fact units according to the order of appearance in the original text, and keep these fact units pointing to the same original text location information;

[0194] The server determines the applicability of the standard based on the timeliness verification results and scope verification results of each evidence item, and judges the risk elements item by item in conjunction with the evidence score.

[0195] The applicability and matching degree of the standard are determined based on the matching results of the scope of applicable institutions, types of applicable businesses, geographical scope of applicable areas, and scope of applicable products.

[0196] For each of the above-mentioned scopes of application, the server will compare the applicable objects recorded in the evidence clauses with the institutions, businesses, regions and product attributes corresponding to the contract under review item by item, and record the status as fully matched, partially matched, mismatched or pending verification according to the preset matching level.

[0197] Complete match and partial match correspond to preset matching values, respectively. No match corresponds to an invalid value. The pending verification state cannot satisfy the effective applicable lower limit on its own.

[0198] The server determines the applicability of the standard based on the matching results of each dimension and their preset weights, and saves the results of the dimensions involved in the calculation to the evidence entries for subsequent review and reuse.

[0199] The matching values ​​and weights for each dimension are preset in the review configuration. The matching values ​​are all in the range of 0 to 1, the weights are non-negative numbers and the sum of the weights is 1.

[0200] A perfect match is set to 1, a no match to 0, a partial match to a preset value between 0 and 1, and a pending verification status to a limit value lower than the effective applicable lower limit.

[0201] The weights are configured according to the impact of the applicable institutional scope, applicable business type, applicable geographical scope, and applicable product scope on the current business risk, and are calibrated by compliance personnel based on historical manual review records; the same weight is used when there are no historical records.

[0202] During the evidence review phase, the effective lower limit of application in the review configuration parameters is used to verify the scope of application; the risk element nodes in the evidence association diagram are task-level nodes generated based on the necessary risk elements, used to aggregate the corresponding factual supporting relationships;

[0203] The paths between this node and contract clause nodes, related clause nodes, and specification nodes all retain specific edge types and original text location information.

[0204] If the evidence score corresponding to a certain risk element reaches the lower limit for risk establishment, the matching degree of normative application reaches the lower limit for effective application, and there is a complete path in the evidence association diagram that starts from the contract clause node, is connected to the related clause node through the reference relationship edge, is connected to the normative node through the normative application relationship edge, and is connected to the risk element node through the factual support relationship edge, then the server forms a preliminary certainty judgment on the risk element and writes it, along with the corresponding contract clause position, the related normative clause position, and the set of evidence items supporting the judgment, into the pending review conclusion record;

[0205] To ensure a clear definition of the complete path, the server verifies the node and edge types in the path in a fixed order:

[0206] A complete path is considered to exist only when the starting point is the contract clause node corresponding to the contract clause currently under review, the intermediate nodes include at least one related clause node that is connected to the contract clause by reference and at least one normative node that is connected to the related clause by normative application, and the ending point is a risk element node that is connected to the normative node by factual support.

[0207] If there are multiple nodes in the path but the edge type order does not meet the above requirements, or if the path is only indirectly connected by supplementary relation edges or superior-subordinate relation edges and lacks reference relation edges or normative application relation edges, then it is not considered to meet the path conditions for a deterministic risk conclusion.

[0208] If any of the above conditions are not met, the server will not make a preliminary definitive judgment, but will instead output a pending review status marker; furthermore, the standard applicability matching degree is calculated according to the following formula:

[0209]

[0210] in, Indicates the degree of applicability of the specification. to These represent the matching values ​​for the applicable institutional scope, applicable business type, applicable geographical scope, and applicable product scope, respectively. to Let each represent the corresponding dimension weight, and satisfy the following conditions: When there is an explicit mismatch in any dimension, the matching value for that dimension is set to 0.

[0211] When any dimension is in a pending verification state, it is only used for calculation based on the corresponding limit value, and cannot be used alone as the basis for reaching the lower limit of the effective applicable threshold.

[0212] For connectivity in a complete path, the server checks each segment according to the edge types already saved between nodes, without using presumed connections based solely on text similarity or thematic similarity; only when the corresponding edge record for each segment connection can be located in the evidence association graph is the path considered to truly exist.

[0213] In this embodiment, the evidence back-check and verification is performed after the initial judgment of the risk elements; for each risk element that is determined to be established, the server reverses to locate the corresponding evidence item and verifies whether the necessary risk element has indeed been covered by the evidence item.

[0214] During reverse positioning, the server traces back in the order of risk element nodes, factual support relationship edges, normative nodes, normative application relationship edges, related clause nodes, and contract clause nodes, and reads the original positioning information, validity period markers, and scope of application matching results stored on each node and edge;

[0215] If it is found that the necessary risk elements are not proven to be covered, or the corresponding regulatory clauses are not in effect, or the scope of application of the regulatory clauses is lower than the preset effective application limit, then it is prohibited to convert the pending review conclusion record into a definitive risk conclusion data package and generate a pending review status mark.

[0216] The applicability matching degree is calculated using the same results as the aforementioned standard applicability matching degree, thereby ensuring that risk assessment and evidence review use a unified applicability judgment standard;

[0217] If an evidence item contains both supporting and rebuttal evidence, and the two come from different versions of the standard, the version of the standard that is valid on the review benchmark date shall be given priority; when different versions of the standard are all valid, they shall continue to be compared in the following order of priority: applicable institutional scope, applicable business type, applicable geographical scope, and applicable product scope.

[0218] Among them, the smallest scope of application means that the scope of the subject of review is the smallest subset of the normative clauses. The priority is compared step by step according to the scope of applicable institutions, applicable business types, applicable geographical scope, and applicable product scope. If the applicable subject can be distinguished in the previous dimension, the comparison result will not be changed in the subsequent dimension.

[0219] If the comparison still shows consistency and conflicting directions, an evidence conflict status marker is output, and manual review positioning information containing conflict clauses and comparison dimensions is generated. Only after passing the above-mentioned evidence coverage, normative validity, scope of application, and conflict handling review conditions will the server output a deterministic risk conclusion data packet.

[0220] Furthermore, when locating the corresponding evidence items in reverse, the server outputs at least one set of traceable link identifiers for each risk element. The link identifiers sequentially record the risk element node identifier, the factual support relationship edge identifier, the normative node identifier, the normative application relationship edge identifier, the related clause node identifier, and the contract clause node identifier.

[0221] When a risk element has multiple supporting links, the link identifiers of each group are retained, and the corresponding links are checked back in descending order of evidence score and standard applicability matching degree. If the link ranked first fails the check but the link ranked last passes the check, the link ranked last that has passed the check is allowed to retain the certainty of the risk element. If all links fail the check, the risk element is put into a pending review status.

[0222] In boundary cases, if an evidence item in the complete evidence set has been marked as pending verification in the preliminary stage, but the item is not the sole basis for covering the relevant necessary risk elements, the server allows it to be retained in the complete evidence set as supplementary evidence. However, in the stage of determining the application of the norm and verifying the evidence, it will not be used as the sole basis for meeting the lower limit of effective application.

[0223] If there is a content relationship between two or more evidence items required for the combined requirements, but no supplementary relationship edge or superior-subordinate relationship edge is formed in the evidence relationship diagram, it is not considered to meet the common proof conditions of the combined requirements, so as to avoid making an unstable judgment based solely on discrete clauses with similar content.

[0224] If neither supporting nor rebuttal evidence can resolve the conflict through version validity and scope of application priority, the server will remain in a pending review state and will not continue to output definitive risk conclusions.

[0225] By adopting the above mechanism, under the same credit contract review, the entire process from screening complete evidence sets, extracting facts, assessing risks, to reviewing conclusions revolves around the same set of evidence items, which can limit the final output to the scope of evidence that has been covered, located, and verified for standard validity.

Claims

1. A multi-stage, large-scale model-based intelligent review method for bank contract compliance risks, executed by a compliance review server, characterized in that... include: Receive contract documents to be reviewed, along with configuration parameters including the review base date, upper limit of length, lower limit of coverage, lower limit of risk establishment, and lower limit of effective application. Obtain the standard documents and index data, and construct the review dataset together with the contract documents to be reviewed; The contract documents to be reviewed are analyzed by a large language model to obtain contract terms and extract risk elements. Risk elements with legal obligations, authorization boundaries and effective time constraints are classified into necessary risk elements to generate review points. Based on the aforementioned review criteria, keyword, content similarity, and clause relationship searches are performed on the review dataset to obtain candidate evidence. The candidate evidence is then used as evidence items to construct an evidence association graph. The validity period and scope of application of the evidence items are verified, the evidence score and text length are calculated, and the coverage score is determined based on the degree of matching between the evidence items and the necessary risk elements. Under the constraints that the sum of the text lengths of the evidence items does not exceed the upper limit of the length limit and the coverage score reaches the lower limit of the coverage limit, a complete evidence group is selected from the evidence association diagram with the goal of minimizing the number of items. Based on the complete evidence set, the risk elements for extracting factual units and the application of norms are determined, thereby obtaining the risk elements for determining the establishment of risk. The evidence back-check verification is performed on the risk elements for determining the establishment of the risk, so as to output a risk conclusion package, or output a mark to be reviewed and location information.

2. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 1, characterized in that, The contract documents to be reviewed include loan contracts, guarantee contracts, mortgage contracts, credit agreements, supplementary agreements, application forms and attachments; The standard documents include statutory norms and regulations, regulatory provisions, industry standards, internal systems and authorization documents. The standard documents include standard clauses, which store the standard subject identifier, effective date, expiration date, applicable institutional scope, applicable business type, applicable geographical scope and applicable product scope. The index data includes keyword index, content vector index, clause relationship index, and specification version index.

3. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 1, characterized in that, Risk extraction criteria include: Extract the subject elements, behavioral elements, condition elements, time limit elements, amount or proportion elements, liability or obligation elements, and exception or exemption elements from the contract terms; If any of the aforementioned elements does not appear in the contract terms, the missing elements shall be extracted and supplemented in the following order: the referenced terms in the contract document to be reviewed, the superior terms containing the contract terms, and the attachment mapping terms associated with the attachment identifier. Based on the supplemented elements, determine the risk event category, the risk elements to be verified, the scope of business application, the time constraints, and the required evidence types, and classify the necessary risk elements according to the legal obligations, authorization boundaries, and effective time constraints.

4. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 1, characterized in that, Based on the aforementioned review criteria, candidate evidence is obtained by performing keyword, content similarity, and clause relationship searches on the review dataset, including: A keyword set is generated based on the review criteria, and a first candidate set is obtained by retrieving the contract documents to be reviewed and the specification documents in the review dataset using the keyword set. Based on the contract terms and the review points, a content retrieval vector is generated using a preset text encoding model. The contract document to be reviewed and the specification document are then converted into candidate clause vectors. A second candidate set is obtained by searching the content similarity between the content retrieval vector and the candidate clause vector. Based on the hierarchical inclusion relationship, clause reference relationship, condition triggering relationship, exception restriction relationship, attachment mapping relationship, normative basis reference relationship, and business correspondence between contract and authorization document, the structure of the first candidate set and the second candidate set is expanded to obtain the candidate evidence.

5. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 2, characterized in that, The timeliness and scope of application verification includes: Retrieve a set of specification versions for the same specification topic identifier; The validity period marker is determined based on the review reference date, the effective date of the specification clause, and the expiration date; If the effective date is earlier than or equal to the review reference date, and the expiration date is empty or later than or equal to the review reference date, then the validity period mark is determined to be valid; if the effective date is later than the review reference date, or the expiration date is earlier than the review reference date, then the validity period mark is determined to be invalid. If a specification version exists but the effective date is missing, the validity period marker will be determined as pending verification, and the specification version will not be used as an independent basis for determination.

6. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 1, characterized in that, The evidence association diagram includes contract clause nodes, attachment nodes, specification nodes, authorization basis nodes, related clause nodes, and risk element nodes, which serve as the evidence items. The evidence entries store source document identifiers, clause texts, original text location information, clause hierarchy paths, time limits parameters, scope of application parameters, and evidence function tags; the evidence function tags include fact support tags, fact rebuttal tags, and tags to be verified. The edges in the evidence association diagram represent reference relationships, supplementary relationships, superior-subordinate relationships, conditional triggering relationships, exception exclusion relationships, normative application relationships, and factual support relationships.

7. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 1, characterized in that, The calculation of the evidence score includes: The content similarity, keyword matching degree, structural relevance, timeliness and validity, scope of application matching degree, and risk element coverage of the evidence items are obtained as evaluation factors. Each evaluation factor is multiplied by its corresponding preset weight and summed to obtain the initial score. If the evidence item has conflicts in terms of time limit, scope of application, or direction of support and rebuttal, a preset conflict deduction value is determined; if there are no such conflicts, the preset conflict deduction value is 0. The evidence score is obtained by subtracting the preset conflict deduction value from the initial score.

8. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 6, characterized in that, The complete evidence set includes: Regarding the necessary risk element, the complete evidence set is required to contain at least one evidence item with an evidence score greater than zero to cover the necessary risk element; For the combined elements that require joint proof among the necessary risk elements, the constraint is jointly covered by at least two different evidence entries that have the supplementary relationship edge or the superior-subordinate relationship edge, and the cumulative coverage score of the joint coverage reaches the lower limit of coverage.

9. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 6, characterized in that, The assessment of risk requirements and regulatory applicability includes: Extract fact units from the evidence entries of the complete evidence group, and bind each fact unit to the corresponding evidence entry and the original text of the contract clause; The applicability matching degree of the standard is determined based on the timeliness verification results and the scope of application verification results corresponding to the evidence items; If the evidence score reaches the lower limit for the establishment of risk, the matching degree of the norm applicability reaches the lower limit for effective application, and there is a complete path in the evidence association graph that starts from the contract clause node, is connected to the associated clause node through the reference relationship edge, is connected to the norm node through the norm applicability relationship edge, and is connected to the risk element node through the fact support relationship edge, then the risk conclusion package is output; Otherwise, output the flag to be reviewed.

10. The intelligent review method for bank contract compliance risk based on multi-stage large-scale model hybrid retrieval according to claim 2, characterized in that, The evidence review and verification includes: For each risk element that constitutes a risk, the corresponding evidence item is located in reverse order. If the necessary risk elements cannot be proven to be covered by the evidence items, or the normative clauses are not in effect, or the scope of application of the normative clauses is lower than the effective application limit, then the output of the risk conclusion package and the generation of the pending review mark shall be prohibited. If there are supporting and rebuttal evidence in the evidence item, and the supporting and rebuttal evidence comes from different versions of the specification, the version of the specification that is valid on the review benchmark date shall be used first. When all the different versions of the specification are valid, they shall be compared in the following order of priority: scope of applicable institutions, type of applicable business, geographical scope, and scope of applicable products, and the specification clause with the smallest scope of application shall be adopted first. If the comparisons are still consistent and the directions conflict after being ordered according to the priority, an evidence conflict status marker is output, and the positioning information containing the conflict clauses and comparison dimensions is generated.

Images