Financial contract intelligent approval management system and method

Abstract

The invention relates to the technical field of data processing, in particular to an intelligent financial contract approval management system and method, and the system comprises an obtaining part, an analysis part, a rule part, an evolution part, a learning part, and a decision part. The acquisition component captures an electronic contract document and analyzes the electronic contract document to generate a structured text sequence; the analysis component extracts entity features, condition features and behavior features through legal grammar dependency analysis to form a coupling vector; the rule component constructs a rule topology network and calculates a matching weight; the evolutionary component recombines the nodes to generate a new rule path when the matching weight is insufficient; the learning component executes hierarchical adversarial training on the new rule path to verify compliance; and the decision component decides an approval result based on the compliance confidence coefficient, and feeds back a manual recheck result to drive system optimization. According to the method, dynamic evolution and closed-loop feedback of the rule network are realized, complex logic branches of a contract are accurately identified, and the misjudgment defect of a predefined rule engine is eliminated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a financial contract intelligent approval management system and method. BACKGROUND

[0002] The financial contract intelligent approval system is based on artificial intelligence technology, automatically parses contract text through a natural language processing module, extracts key elements such as amount, clauses, and payment conditions, integrates a pre-defined business rule engine and a machine learning model, analyzes contract compliance and assesses potential risk points such as breach probability or legal conflicts, identifies abnormalities and matches best practices through a logical reasoning chain, and the decision-making module can automatically approve compliant contracts or forward high-risk cases to manual review, thereby optimizing the approval process and reducing errors.

[0003] In the financial contract intelligent approval system management scenario, the technical pain points are mainly the lack of dynamic adaptation capability of the rule engine and the limitation of context understanding, which is due to the fact that the pre-defined rule system cannot cover the complex semantic combination of contract clauses and industry dynamic changes. For example, when there is an exception clause in the contract that allows a breach of the regular amount limit under certain conditions, the system may misjudge the compliant clause as a high-risk abnormality due to the lack of dynamic semantic association analysis capability, which is caused by the fact that the rule engine only supports static threshold matching and cannot perform multi-level logical reasoning on context conditions (such as "effective after special approval by the board of directors"), resulting in legal contracts being incorrectly forwarded to manual review, causing redundant approval processes. SUMMARY

[0004] To address the deficiencies in the prior art, the present application provides a financial contract intelligent approval management system and method, which solves the technical problem of misjudgment caused by the lack of dynamic adaptation capability of the pre-defined rule engine, which cannot accurately identify the complex logical branches of contract clauses.

[0005] To solve the above technical problems, the specific application of the present application is as follows: In a first aspect, the present application provides a financial contract intelligent approval management system, comprising: An acquisition module for connecting to an enterprise business system port to capture electronic contract documents, performing structural analysis on the documents to generate a structured text sequence, and outputting the structured text sequence to an analysis module; An analysis module for receiving the structured text sequence from the acquisition module, extracting entity features, condition features, and behavior features in the clauses through legal syntax dependency analysis, forming a coupling vector, and pushing the coupling vector to a rule module; A rule module is configured to load a historical compliance case construction rule topology network, the rule topology network including nodes and edges, the nodes storing rule types and thresholds, and the edges storing logical relationships and trigger frequencies, and the rule module is configured to perform matching weight calculation after receiving the coupling vector parsed by the parsing module; An evolution module is configured to detect the matching weight of the rule module, trigger a node recombination operation when the matching weight is lower than a critical value, perform cross-recombination of a selected main rule node and a logically adjacent node to generate a new rule path with fused node features, and output the new rule path to the learning module. A learning module is configured to receive the new rule path output by the evolution module, perform hierarchical adversarial training to verify compliance, generate adversarial samples through primary grammar replacement sample generation and advanced legal conflict injection, filter the adversarial samples through a legal knowledge base, and output verification results, the verification results being the new rule path that complies and being input to the rule module to update the rule topology network of the rule module. A ruling module is configured to obtain a compliance confidence coefficient of the rule path from the rule module, perform automatic contract approval when the compliance confidence coefficient is higher than a preset threshold, output a manual review instruction when the compliance confidence coefficient is lower than the preset threshold, and feed back a manual review annotation result to the learning module to adjust the training strategy, wherein the learning module inputs the training strategy adjusted by the manual review annotation to the evolution module to optimize the node recombination operation.

[0006] Further, the financial contract intelligent approval management system provided by the present application, the obtaining module comprises: A document parser is deployed to identify tables and clause blocks in the document format to obtain initial parsed text; An OCR skew correction algorithm is performed on the scanned document to process character misplacement and generate corrected text; Based on the initial parsed text and the corrected text, regular expression matching is performed to extract the contract number and the name of the contracting party, and a standard JSON structure data stream is output as a structured text sequence to the parsing module.

[0007] Further, the financial contract intelligent approval management system provided by the present application, the parsing module comprises: The structured text sequence is input to perform part-of-speech tagging and named entity recognition to obtain basic semantic features; Based on the basic semantic features, a bidirectional LSTM is used to capture long-distance dependency relationships to output enhanced semantic features; The enhanced semantic features are focused on logical connection words through an attention mechanism, the clauses are disassembled into coupling vectors of entities, operators and constraint conditions, and the coupling vectors are pushed to the rule module in real time.

[0008] Further, the financial contract intelligent approval management system provided by the present application, the rule module comprises: The node features include rule types and threshold ranges, and the edge attributes include logical relationships and historical trigger frequencies, to form an initial rule network; When loading the historical case library in real time, the association relationship of the contract type, the abnormal marker and the adjudication result from the case library is extracted, the association relationship is converted into a connection edge with a weight between nodes, and the topology structure of the initial rule network is updated; When the coupling vector of the receiving and analyzing module is received, a matching thread is started to calculate the cosine similarity, and the output similarity value determines whether to trigger the evolution module.

[0009] Further, the financial contract intelligent approval management system provided by the application comprises: When the cosine similarity is detected to be less than a set threshold, a main rule node is selected and a logically adjacent node is retrieved to obtain a candidate node set; Based on the candidate node set, a cross-recombination process is performed to generate fusion node features, and the fusion node features constitute a new rule path; The new rule path is output to the learning module in real time to perform hierarchical adversarial training.

[0010] Further, the financial contract intelligent approval management system provided by the application comprises: The learning module receives the new rule path output by the evolution module and starts hierarchical adversarial training: The primary generator synthesizes a grammatical equivalent sample based on the new rule path using a synonym replacement and a sentence restructuring model; The advanced generator injects a legal numerical conflict into the grammatical equivalent sample to generate an adversarial sample that violates a legal prohibitive rule; The legal clause boundary detector connects a legal regulation knowledge base to filter the adversarial sample, and outputs the filtered variant sample to the analysis module to strengthen abnormality recognition; The compliance of the filtered adversarial sample is verified, and when the verification is passed, the rule module updates its topology network.

[0011] Further, the financial contract intelligent approval management system provided by the application comprises: The rule path matching similarity is obtained from the rule module in real time: When the similarity is higher than a set value, an automatic approval result is output through an automatic adjudication channel; When the rule path is a new rule path verified by the learning module, a cooperative verification channel is activated to perform rapid approval; When the compliance confidence coefficient is lower than a threshold, an artificial review instruction is output and an artificial correction result is obtained, a negative sample feature vector is generated according to the error type marked by the correction result; The negative sample feature vector is input into the rule module in reverse to adjust the node weight, and the learning module is driven to reconstruct the generation strategy of the adversarial sample.

[0012] Further, the financial contract intelligent approval management system provided by the present application comprises: The artificial review correction result is analyzed, and threshold value misjudgment or logical omission error types are identified; Based on the error type, a negative sample feature vector is generated and transmitted to the rule module in real time to update the node weight value; The error type is input into the learning module in real time, and the learning module is driven to adjust the statutory numerical value conflict injection strategy of the advanced generator; The updated weight value of the rule module and the adjusted strategy of the learning module jointly act on the node recombination operation of the evolution module to form a cross-module collaborative optimization closed loop.

[0013] Further, the financial contract intelligent approval management system provided by the present application further comprises: The learning module verifies the new rule path in compliance and writes the topology network of the rule module in real time; The filtered and mutated sample generated by the learning module is transmitted to the analysis module to enhance the identification ability of the analysis module to the logical conflict clauses; The feature vector output by the analysis module in the abnormal detection scene is input into the rule module in reverse to update the node threshold value range and the logical relationship attribute, forming a cross-reinforcement loop of rule optimization and semantic analysis.

[0014] In a second aspect, the present application provides a financial contract intelligent approval management method, applied to the financial contract intelligent approval management system as described, comprising: Step 1, connect the enterprise business system port to capture the electronic contract document, perform structure analysis on the document to generate a structured text sequence, and output the structured text sequence to step 2; Step 2, receive the structured text sequence from step 1, extract entity features, condition features and behavior features in the clauses through legal syntax dependency analysis, form a coupling vector, and push the coupling vector to step 3; Step 3, load historical compliance cases to construct a rule topology network, the rule topology network comprises nodes and edges, the nodes store rule types and threshold values, and the edges store logical relationships and trigger frequencies; after receiving the coupling vector of step 2, step 3 performs matching weight calculation; Step 4, detect the matching weight of step 3, trigger node recombination operation when the matching weight is lower than the critical value, perform cross-recombination of the selected main rule node and the logically adjacent node, generate a new rule path with fused node features, and output the new rule path to step 5; Step 5: Receive the new rule path output from Step 4, perform hierarchical adversarial training to verify compliance, generate adversarial samples through primary grammar replacement sample generation and advanced legal conflict injection, output the verification result after filtering by the legal knowledge base, input the new rule path that is compliant into Step 3, and update the rule topology network of Step 3. Step 6: Obtain the compliance confidence coefficient of the rule path from Step 3. When the compliance confidence coefficient is higher than the preset threshold, execute automatic contract approval. When the compliance confidence coefficient is lower than the preset threshold, output the manual review instruction and feed back the manual review annotation result to Step 5 to adjust the training strategy. Step 5 inputs the training strategy adjusted by manual review annotation into Step 4 to optimize the node reorganization operation.

[0015] Beneficial effects of this invention; This invention deconstructs contract terms into a three-element coupled vector of entity features, conditional features, and behavioral features through a semantic coupling parsing engine, accurately capturing complex logical branches. A dynamic rule network calculates matching weights in real time based on the coupled vectors. When the weights fall below a critical value, an evolutionary topology operator is triggered to perform node cross-reorganization, generating a new logical path that integrates rule features and directly adapts to the semantic structure of contract exception clauses. A compliance boundary pressure learning framework performs hierarchical adversarial training on the new rule path. The primary generator simulates semantic variants, while the advanced generator injects legal boundary conflicts. After filtering by a legal knowledge base, the feasibility of the rules is verified, and the anomaly recognition capability of the parsing engine is simultaneously strengthened. An adaptive adjudication module processes contracts based on the confidence coefficient of the rule matching. The results of manual review are converted into negative sample feature vectors and fed back to the rule network and adversarial training framework, forming a closed-loop optimization mechanism for rule generation, verification, and application. This effectively improves the system's dynamic adaptability and recognition accuracy for non-standard contract terms and significantly reduces the need for manual review. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0017] Figure 1 A flowchart of a financial contract intelligent approval management method provided in an embodiment of the present invention. Detailed Implementation

[0018] In order to make the technical solutions of the present application clearer, the present application will be described below in detail with reference to the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. The present application provided by each embodiment of the present application will be described in detail below with reference to the drawings. For the purpose of better understanding the present application, the present application will be further described in detail below.

[0019] In a first aspect, the present application provides a financial contract intelligent approval management system, comprising: The acquisition module is configured to connect to a port of an enterprise business system to capture an electronic contract document, perform structural analysis on the document to generate a structured text sequence, and output the structured text sequence to the analysis module. The analysis module is configured to receive the structured text sequence from the acquisition module, extract entity features, condition features, and behavior features in the clauses through legal syntax dependency analysis, form a coupling vector, and push the coupling vector to the rule module. The rule module is configured to load historical compliance cases to construct a rule topology network, the rule topology network includes nodes and edges, the nodes store rule types and threshold values, and the edges store logical relationships and trigger frequencies. The rule module performs matching weight calculation after receiving the coupling vector from the analysis module. The evolution module is configured to detect the matching weight of the rule module, trigger node reorganization when the matching weight is below a critical value, perform selection of primary rule nodes and logical adjacent nodes for cross-reorganization, generate a new rule path with fused node features, and output the new rule path to the learning module. The learning module is configured to receive the new rule path output by the evolution module, perform hierarchical adversarial training to verify compliance, generate adversarial samples through primary grammar replacement samples and high-level legal conflict injection, filter the adversarial samples through a legal knowledge base, and output verification results. The verification results are input into the rule module as compliant new rule paths, and the rule topology network of the rule module is updated. The decision module is configured to obtain a compliance confidence coefficient of the rule path from the rule module, perform automatic contract approval when the compliance confidence coefficient is higher than a preset threshold, output a manual review instruction when the compliance confidence coefficient is lower than the preset threshold, and feed back the manual review annotation result to the learning module to adjust the training strategy. The learning module inputs the training strategy adjusted by the manual review annotation into the evolution module to optimize the node reorganization operation.

[0020] After the acquisition module connects the enterprise business system port to capture the electronic contract document, the document structure parsing process is executed. The document structure parsing first deploys the document parser to identify the table and clause blocks in the different format documents to generate the initial parsed text; the OCR skew correction algorithm is executed simultaneously for the scanned document type to correct the misplacement of the text to generate the corrected text; based on the initial parsed text and the corrected text, the contract number and the signing party name key fields are extracted through the regular expression pattern matching, and finally the standard JSON structure data stream is output as the structured text sequence. This process realizes the transformation of the original contract data into a machine-processable standardized sequence.

[0021] After the parsing module receives the structured text sequence output by the acquisition module, the legal grammar dependency analysis process is started. The analysis process first performs part-of-speech tagging and named entity recognition on the text sequence to obtain basic semantic features; based on the basic semantic features, a bidirectional long short-term memory network is used to capture long-distance dependency relationships in the clauses to output enhanced semantic features; the attention mechanism is applied to the enhanced semantic features to focus on the logical connection words, and the complex clauses are disassembled into three-element coupled vectors of entity features, condition features and behavior features. The completed coupled vectors are pushed to the subsequent processing unit in real time.

[0022] The rule module loads the historical compliance case library to construct a dynamic rule topology network. The rule topology network is based on a graph database to realize storage function, the network nodes store the core attributes of rule type and threshold range, and the network edges store the key parameters of logical relationship and historical trigger frequency; when the case library is loaded in real time, the contract type, abnormal mark and decision result association relationship in the case are extracted, and the association relationship is converted into a connection edge with weight between nodes; when the coupled vector pushed by the parsing module is received, the matching thread is immediately started, and the cosine similarity value of the coupled vector and the rule node is calculated. The similarity value output result determines whether to trigger the reorganization operation.

[0023] The evolution module continuously monitors the similarity value output by the rule module. When the similarity value is detected to be lower than the set threshold, the node reorganization operation is immediately triggered: first, the main rule node is selected and its logical adjacent nodes are retrieved to generate a candidate node set; based on the candidate node set, the cross-reorganization process is executed to fuse the rule features of different nodes to generate a new rule path. The new rule path represents the logic of the composite business rule, and is output to the verification system in real time.

[0024] The learning module receives the new rule path output by the evolution module and performs hierarchical adversarial training verification. The hierarchical adversarial training first generates a syntactically equivalent sample based on the new rule path through a primary generator, and uses synonym replacement and sentence restructuring techniques to keep the semantics unchanged; the advanced generator injects numerical conflicts that violate the legal prohibitive rules into the syntactically equivalent sample; the legal clause boundary detector connects the legal regulation knowledge base to filter the adversarial samples, and outputs the compliance variation samples to the analysis module to strengthen the abnormal recognition ability; the compliance new rule path verified passes is used to update the rule topology network structure in real time.

[0025] The decision module obtains the compliance confidence coefficient of the rule path from the rule module in real time. When the compliance confidence coefficient is higher than the preset threshold, the automatic decision channel is used to execute the contract approval; when the rule path is a new rule verified by the learning module, the cooperative verification channel is activated for rapid approval; when the compliance confidence coefficient is lower than the threshold, an artificial review instruction is output, and a negative sample feature vector is generated according to the error type marked by the artificial correction result. The negative sample feature vector is transmitted reversely to the rule module to adjust the node weight and drive the learning module to reconstruct the adversarial strategy, thereby forming a closed-loop optimization mechanism.

[0026] The adjusted adversarial strategy of the learning module is synchronously input into the evolution module to optimize the logical constraint conditions of the node recombination operation. The updated node weight of the rule module and the reconstructed generation strategy of the learning module jointly act on the evolution module to form cross-system cooperative optimization. The analysis module receives the filtered variation sample output by the learning module to strengthen the identification accuracy of the logical conflict clause; the feature vector output by the analysis module in the abnormal detection is reversely input into the rule module to update the node threshold range and the logical relationship attribute, thereby realizing the cross-strengthening of rule optimization and semantic analysis. The entire system continuously reduces the misjudgment rate of the contract clause through a data flow closed loop.

[0027] Specifically, the financial contract intelligent approval management system provided by the application comprises: A document parser is deployed to identify tables and clause blocks in the document format, and initial parsed text is obtained; An OCR skew correction algorithm is performed on the scanned document to correct mispositioned characters, and corrected text is generated; Based on the initial parsed text and the corrected text, regular expression matching is performed to extract the contract number and the name of the contracting party, and standard JSON structure data streams are output as structured text sequences to the analysis module.

[0028] The document structure analysis unit of the acquisition module deploys a document parser to divide the table blocks and the clause text blocks in the PDF or Word format to form an initial parsed text; the OCR and rectification algorithm is synchronously started for the scanned document, the character displacement is corrected by the glyph comparison and position correction to generate a corrected text; the regular expression pattern matching is performed based on the initial parsed text and the corrected text to extract the contract number field and the signatory name field key metadata; the metadata and the text content are integrated to generate a standard JSON structure data stream, which is transmitted to the analysis module as a structured text sequence.

[0029] Specifically, the financial contract intelligent approval management system provided by the present application comprises: The structured text sequence is input to perform part-of-speech tagging and named entity recognition to obtain basic semantic features; Based on the basic semantic features, a bidirectional LSTM is used to capture long-distance dependency relationships to output enhanced semantic features; The enhanced semantic features are focused on logical connection words through an attention mechanism, and the clauses are disassembled into entity, operator and constraint condition coupling vectors, which are pushed to the rule module in real time.

[0030] The analysis module performs part-of-speech tagging and recognition of verbs and nouns in the input structured text sequence, and performs named entity recognition to mark legal subjects and amount clauses in parallel; based on the marking results, a bidirectional long short-term memory network is used to analyze cross-sentence dependency relationships in the clauses to output enhanced semantic features carrying context semantics; the attention mechanism is applied to the enhanced semantic features to weight the logical connection words, and the clauses are disassembled into triplets of entity features, condition operators and constraint behavior features according to the weighting results; the generated triple coupling vectors are pushed to the rule module data processing interface in real time.

[0031] Specifically, the financial contract intelligent approval management system provided by the present application comprises: Based on the graph database, the node features and edge attributes are stored, the node features include rule types and threshold ranges, and the edge attributes include logical relationships and historical trigger frequencies, to form an initial rule network; When the historical case library is loaded in real time, the association relationship of the contract type, the abnormal mark and the decision result in the case library is extracted, the association relationship is converted into a connection edge with weight between nodes, and the topology structure of the initial rule network is updated; When the coupling vector of the analysis module is received, a matching thread is started to calculate the cosine similarity, and the output similarity value determines whether to trigger the evolution module.

[0032] The rule module constructs an initial rule network based on a graph database, network nodes store payment rule type and amount threshold range attributes, and network edges store clause exclusion relationships and historical trigger frequency attributes; when loading a historical case library, automatically extract the contract classification label and abnormal mark associated decision results in the case; convert the association relationship into a weighted connection edge between nodes to update the network topology logic; immediately start a cosine similarity calculation thread after receiving the coupling vector from the parsing module, and output the similarity value to the state monitoring port.

[0033] Specifically, the financial contract intelligent approval management system provided by the present application comprises: When the cosine similarity is detected to be less than a set threshold, a main rule node is selected and logical adjacent nodes are retrieved to obtain a candidate node set; Based on the candidate node set, a cross-recombination process is performed to generate a fusion node feature, and the fusion node feature constitutes a new rule path; The new rule path is output in real time to the learning module to perform hierarchical adversarial training.

[0034] The evolution module continuously receives the similarity value output by the rule module; when it is monitored that the similarity is lower than the set threshold, the node retrieval program is activated, and the main rule node is selected according to the rule type similarity; the candidate node set is generated by traversing the logical adjacent nodes centered on the main rule node; the feature cross-recombination is performed on the candidate node set to generate a composite rule logic by fusing the rule type and threshold features of different nodes; the new rule path carrying the fusion feature label is transmitted in real time to the learning module verification interface.

[0035] Specifically, the financial contract intelligent approval management system provided by the present application comprises: Receive the new rule path output by the evolution module and start hierarchical adversarial training: The primary generator synthesizes a grammatical equivalent sample based on the new rule path using a synonym replacement and sentence restructuring model; The advanced generator injects a legal numerical value conflict into the grammatical equivalent sample to generate an adversarial sample that violates the legal prohibitive rules; The legal clause boundary detector connects the legal knowledge base to filter the adversarial sample, and outputs the filtered variant sample to the parsing module to enhance abnormality identification; Verify the compliance of the filtered adversarial sample, and trigger the rule module to update its topology network when the verification is passed.

[0036] The learning module initiates hierarchical adversarial training after receiving the new rule path, a primary generator generates syntactic equivalent samples by using a synonym replacement model and a sentence restructuring model, a senior generator injects conflict parameters that violate legal numerical boundaries into the syntactic equivalent samples, a legal clause boundary detector filters the conflict samples by connecting a civil code knowledge base, the filtered variant samples are input into the analysis module to enhance the abnormality detection capability, and the compliance new rule path is written back to the rule module to update the node threshold and the connection edge weight.

[0037] Specifically, the financial contract intelligent approval management system provided by the application comprises a rule module, a learning module, an analysis module and an evolution module. The rule module is configured to acquire a rule path matching similarity in real time. When the similarity is higher than a set value, an automatic approval result is output through an automatic adjudication channel. When the rule path is a new rule path verified by the learning module, a collaborative verification channel is activated to perform rapid approval. When the compliance confidence coefficient is lower than a threshold value, an artificial review instruction is output and an artificial correction result is obtained, a negative sample feature vector is generated according to the error type marked by the correction result, and the negative sample feature vector is reversely input into the rule module to adjust the node weight. The negative sample feature vector is reversely input into the rule module to adjust the node weight, and the learning module is driven to reconstruct the adversarial sample generation strategy.

[0038] The rule module is configured to acquire a rule path matching similarity in real time.

[0039] Specifically, the financial contract intelligent approval management system provided by the application comprises a rule module, a learning module, an analysis module and an evolution module. The artificial review correction result is analyzed, and a threshold misjudgment or a logical omission error type is identified. A negative sample feature vector is generated based on the error type and is transmitted to the rule module in real time to update the node weight value. The error type is input into the learning module synchronously, and the learning module is driven to adjust the legal numerical conflict injection strategy of the senior generator. The updated weight value of the rule module and the adjusted strategy of the learning module jointly act on the node restructuring operation of the evolution module to form a cross-module collaborative optimization closed loop.

[0040] The receiving artificial labeling result analyzes the threshold configuration error or the logical relationship missing type; a negative sample feature vector is generated based on the error type characteristics; the negative sample feature vector is input into the rule module to perform a node weight attenuation operation; the weight attenuation data stream synchronously triggers the learning module to reconstruct the countermeasures, adjust the legal conflict injection rules of the advanced generator and the sentence restructuring parameters of the primary generator; and the reconstructed countermeasures are transmitted to the evolution module to optimize the node retrieval threshold.

[0041] Specifically, the financial contract intelligent approval management system provided by the application further comprises: The learning module verifies the new rule path in compliance and writes the topology network of the rule module in real time; The filtered variation samples generated by the learning module are transmitted to the analysis module to strengthen the identification ability of the analysis module to the logical conflict clauses; The feature vector output by the analysis module in the abnormal detection scene is input into the rule module in reverse, the node threshold range and the logical relationship attribute are updated, and a cross-strengthening loop of rule optimization and semantic analysis is formed.

[0042] The rule module updates the network topology after receiving the new rule path verified by the learning module, increases the fusion feature nodes and creates new logical connection edges; the filtered variation samples output by the learning module are input into the analysis module to train the abnormal detection model and improve the identification sensitivity to the exceptional clauses; the feature vector output by the analysis module in the abnormal detection is input into the rule module in reverse, the node threshold range and the connection edge trigger frequency parameters are dynamically adjusted; and a cross-strengthening mechanism of rule logic optimization and semantic analysis ability enhancement is formed.

[0043] In a second aspect, the application provides a financial contract intelligent approval management method applied to the financial contract intelligent approval management system, comprising: Step 1, connecting the enterprise business system port to capture the electronic contract document, performing structure analysis on the document to generate a structured text sequence, and outputting the structured text sequence to step 2; Step 2, receiving the structured text sequence from step 1, extracting entity features, condition features and behavior features in the clauses through legal syntax dependency analysis, forming a coupling vector, and pushing the coupling vector to step 3; Step 3, loading historical compliance cases to construct a rule topology network, the rule topology network comprising nodes and edges, the nodes storing rule types and thresholds, and the edges storing logical relationships and trigger frequencies; after receiving the coupling vector of step 2, step 3 performs matching weight calculation; Step 4, detecting the matching weight of step 3, triggering node recombination operation when the matching weight is lower than the critical value, performing cross-recombination of the selected main rule node and the logically adjacent node, generating a new rule path with fusion node features, and outputting the new rule path to step 5; Step 5, receiving the new rule path output in step 4, performing hierarchical adversarial training to verify compliance, generating adversarial samples through primary grammar replacement sample generation and high-level legal conflict injection, outputting verification results after filtering through the legal knowledge base, and inputting the verification results of the compliant new rule path into step 3 to update the rule topology network in step 3; Step 6, obtaining the compliance confidence coefficient of the rule path from step 3, when the compliance confidence coefficient is higher than the preset threshold, performing automatic contract approval, when the compliance confidence coefficient is lower than the preset threshold, outputting manual review instructions, and feeding back the manual review annotation results to step 5 to adjust the training strategy, wherein step 5 inputs the training strategy adjusted by the manual review annotation into step 4 to optimize the node reorganization operation.

[0044] The application solves the adaptability defects of the pre-defined rule engine by constructing a dynamic rule network and a closed-loop feedback mechanism. Specifically, the semantic coupling analysis engine first decomposes the contract clauses into a three-element coupling vector of entity features, condition features and behavior features, breaking through the complexity of natural language combination; the dynamic rule network is initially constructed by loading historical compliance cases to generate node and connection edge topology, but unlike the static rule base, the network calculates the matching weight in real time by receiving the coupling vector, and when the weight is lower than the critical value, it indicates that the present rule cannot cover the clause semantics, and immediately triggers the evolutionary topology operator to perform node crossover and reorganization, for example, the "board approval" node and the "exchange rate fluctuation threshold" node are fused to generate a new logic path "exchange rate exceeds the threshold and requires special approval", which directly adapts to the exception clause logic branch in the contract.

[0045] The new rule path is input into the compliance boundary pressure learning framework to verify its feasibility, and the framework implements hierarchical adversarial training: the primary generator generates a semantically equivalent clause based on synonym replacement to simulate semantic variants, and the advanced generator injects conflict parameters that violate legal mandatory boundaries, such as non-statutory low penalty rate; after the generated samples are cleaned by the legal knowledge base filtering device, the output logical variant samples are used to train the analysis engine in reverse, enhancing its robustness to abnormal semantic combinations; the compliance rule path that passes the verification is updated to the dynamic rule network, allowing the topology structure to continuously expand the non-standard rule space.

[0046] A double-channel adjudication mechanism is provided at the end of the system, which automatically approves low-risk contracts when the rule matching confidence coefficient is higher than the threshold; for high-risk clauses and new rule paths, the manual review results generate negative sample feature vectors, including threshold misjudgment or logical omission marker types, and the vectors are used to real-time correct the rule node weight and the reorganization strategy of the evolutionary operator. The whole process forms a data-driven closed loop: semantic analysis abnormalities trigger rule reorganization, reorganized rules update the network after adversarial verification, network updates reduce the subsequent misjudgment rate, manual review optimizes rule generation logic, and finally realizes the adaptive identification of the rule engine for complex contract branches.

Claims

1. A financial contract intelligent approval management system, characterized in that, include: The acquisition module is used to connect to the enterprise business system port to capture electronic contract documents, perform structural parsing on the documents to generate a structured text sequence, and output the structured text sequence to the parsing module; The parsing module receives the structured text sequence from the acquisition module, extracts the entity features, condition features, and behavioral features from the clauses through legal grammar dependency analysis, forms a coupling vector, and pushes the coupling vector to the rules module. The rules module is used to load historical compliance cases to build a rule topology network. The rule topology network includes nodes and edges. Nodes store rule types and thresholds, and edges store logical relationships and trigger frequencies. The rules module receives the coupling vector from the parsing module and then performs matching weight calculation. The evolution module is used to detect the matching weight of the rule module. When the matching weight is lower than the threshold, the node reorganization operation is triggered. The main rule node is selected and reorganized with logically adjacent nodes to generate a new rule path that integrates node features. The new rule path is then output to the learning module. The learning module receives new rule paths output by the evolution module, performs hierarchical adversarial training to verify compliance, generates adversarial samples through primary grammar replacement sample generation and advanced legal conflict injection, outputs verification results after filtering by the legal knowledge base, inputs new rule paths that are compliant into the rule module, and updates the rule topology network of the rule module. The adjudication module is used to obtain the compliance confidence coefficient of the rule path from the rule module. When the compliance confidence coefficient is higher than the preset threshold, automatic contract approval is executed. When the compliance confidence coefficient is lower than the preset threshold, a manual review instruction is output, and the manual review annotation results are fed back to the learning module to adjust the training strategy. The learning module inputs the training strategy adjusted by manual review annotation into the evolution module to optimize the node reorganization operation.

2. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, The acquisition module includes: Deploy a document parser to identify tables and clause blocks in the document format and obtain the initial parsed text; The scanned document is processed using an OCR correction algorithm to correct text misalignment and generate corrected text. Based on the initial parsed text and the corrected text, regular expression matching is performed to extract the contract number and the name of the signatory, and the standard JSON structured data stream is output as a structured text sequence to the parsing module.

3. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, The parsing module includes: Input a structured text sequence and perform part-of-speech tagging and named entity recognition to obtain basic semantic features; Based on basic semantic features, a bidirectional LSTM is used to capture long-distance dependencies and output enhanced semantic features. To enhance semantic features, an attention mechanism is used to focus on logical connectives, decompose terms into coupled vectors of entities, operators, and constraints, and push them to the rules module in real time.

4. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, The rule module includes: Based on the graph database, node features and edge attributes are stored. Node features include rule type and threshold range, and edge attributes include logical relationship and historical trigger frequency, forming an initial rule network. When loading the historical case library in real time, the association between contract types, anomaly markers and adjudication results is extracted from the case library, and the association is converted into weighted connection edges between nodes to update the topology of the initial rule network. When the coupling vector from the parsing module is received, the matching thread is started to calculate the cosine similarity, and the output similarity value determines whether the evolution module is triggered.

5. The intelligent approval management system for financial contracts as described in claim 4, characterized in that, The evolution module includes: When the cosine similarity is less than a set threshold, the main rule node is selected and logically adjacent nodes are retrieved to obtain a set of candidate nodes; Based on the candidate node set, a cross-recombination process is executed to generate fused node features, and the fused node features constitute a new rule path; The new rule path is output to the learning module in real time to perform layered adversarial training.

6. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, The learning module includes: Receive the new rule path output by the evolution module and initiate hierarchical adversarial training: The primary generator is based on a new rule path and uses a synonym replacement and sentence recombination model to synthesize grammatically equivalent samples. Advanced generators inject legal numerical conflicts into syntactically equivalent samples to generate adversarial samples that violate legal prohibitions. The legal clause boundary detector connects to a legal knowledge base to filter adversarial samples and outputs the filtered variant samples to the parsing module to enhance anomaly identification. Verify the compliance of the filtered adversarial examples, and trigger the rules module to update its topology network when the verification passes.

7. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, The adjudication module includes: Real-time acquisition of rule path matching similarity from the rules module: When the similarity is higher than the set value, the automatic approval result is output through the automatic adjudication channel; When the rule path is a new rule path verified by the learning module, activate the collaborative verification channel to perform fast approval; When the compliance confidence coefficient is lower than the threshold, a manual review instruction is output and a manual correction result is obtained. Based on the correction result, the error type is labeled and a negative sample feature vector is generated. The negative sample feature vectors are input into the rule module to adjust the node weights, while simultaneously driving the learning module to reconstruct the adversarial sample generation strategy.

8. The intelligent approval management system for financial contracts as described in claim 7, characterized in that, The adjudication module includes: Analyze the results of manual review and correction to identify types of errors such as threshold misjudgment or logical omission; Negative sample feature vectors are generated based on error types and transmitted to the rule module in real time to update node weight values. The error type is synchronously input into the learning module, which drives the learning module to adjust the statutory numerical conflict injection strategy of the advanced generator. The updated weight values ​​from the rules module and the adjusted strategies from the learning module work together to perform node reorganization operations in the evolution module, forming a cross-module collaborative optimization closed loop.

9. The intelligent approval management system for financial contracts as described in claim 1, characterized in that, Also includes: The new rule paths for compliance verification in the learning module are written into the topology network of the rule module in real time; The filtered variant samples generated by the learning module are transmitted to the parsing module to enhance the parsing module's ability to identify logically conflicting clauses; The feature vectors output by the parsing module in the anomaly detection scenario are input back into the rule module to update the node threshold range and logical relationship attributes, forming a cross-reinforcement loop of rule optimization and semantic parsing.

10. A method for intelligent approval management of financial contracts, applied to the intelligent approval management system for financial contracts as described in any one of claims 1 to 9, characterized in that, include: Step 1: Connect to the enterprise business system port to capture electronic contract documents, perform structural parsing on the documents to generate a structured text sequence, and output the structured text sequence to Step 2; Step 2: Receive the structured text sequence from Step 1, extract the entity features, condition features, and behavioral features from the clauses through legal grammar dependency analysis, form a coupling vector, and push the coupling vector to Step 3; Step 3: Load historical compliance cases to build a rule topology network. The rule topology network includes nodes and edges. Nodes store rule types and thresholds, and edges store logical relationships and trigger frequencies. Step 3 receives the coupling vector from Step 2 and then performs matching weight calculation. Step 4: Detect the matching weight in Step 3. When the matching weight is lower than the threshold, trigger the node reorganization operation, select the main rule node and logically adjacent nodes for cross-reorganization, generate a new rule path that integrates node features, and output the new rule path to Step 5. Step 5: Receive the new rule path output from Step 4, perform hierarchical adversarial training to verify compliance, generate adversarial samples through primary grammar replacement sample generation and advanced legal conflict injection, output the verification result after filtering by the legal knowledge base, input the new rule path that is compliant into Step 3, and update the rule topology network of Step 3. Step 6: Obtain the compliance confidence coefficient of the rule path from Step 3. When the compliance confidence coefficient is higher than the preset threshold, execute automatic contract approval. When the compliance confidence coefficient is lower than the preset threshold, output the manual review instruction and feed back the manual review annotation result to Step 5 to adjust the training strategy. Step 5 inputs the training strategy adjusted by manual review annotation into Step 4 to optimize the node reorganization operation.

Images