Green financing-oriented credit large model joint modeling system and method

Abstract

The invention discloses a green financing-oriented credit large model joint modeling system and method, and belongs to the technical field of green financing, the system comprises a multi-source green data management module, a credible reasoning and calculation module, a large model interaction module and a dynamic knowledge evolution module, and the method comprises the following steps: 1, obtaining original data of a green financing project, generating a knowledge constraint signal; 2, analyzing the structured data to generate a logic decision chain, executing tasks related to environmental benefit measurement and calculation in the structured data, and generating quantitative indexes; 3, integrating the quantitative index and the logic decision chain into a reasoning context, converting the knowledge constraint signal into an output constraint instruction, and generating a credit evaluation report; and 4, receiving feedback information of the credit evaluation report, and updating and optimizing the system. According to the invention, by generating the logic decision chain and the quantitative index, the illusion of the large model is effectively reduced, and the credit assessment accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of green financing, and particularly relates to a credit large model joint modeling system and method for green financing. BACKGROUND

[0002] With the rapid development of green finance, using large-scale language models to assess the credit risk of green financing projects has become an important research direction. Large models can process massive amounts of unstructured text data such as enterprise environmental responsibility reports and project feasibility studies, thereby assisting financial institutions in making credit decisions.

[0003] In the prior art, the text data related to green financing projects, such as enterprise environmental reports and policy documents, are usually directly input into a pre-trained large language model, and the model directly generates judgments on the green attributes of the projects, assessment opinions on the credit risks, or generates preliminary credit evaluation reports based on its internal parameterized knowledge. This method relies on the knowledge and pattern recognition ability obtained by the model during the training stage, and aims to quickly realize the automated processing of unstructured information.

[0004] However, the existing green credit evaluation technology based on large models still has obvious defects, mainly in the following two aspects: first, the inherent "hallucination" problem of large models leads to the possibility of factual errors or false content in the generated information, and in the high-risk evaluation scenario of green credit, such uncertainty will directly cause credit risk, for example, the model may falsely fabricate or confuse the environmental performance data of the project, leading to misjudgment of the "green" attributes of the project; second, the "catastrophic forgetting" phenomenon commonly existing in large models makes it difficult for them to adapt to the rapid iteration of green financial policies, further limiting their practicality. When new green industry directories or environmental benefit accounting standards are released, the model is difficult to effectively integrate new knowledge, which may lead to outdated evaluation basis and affect the accuracy and compliance of the evaluation results.

[0005] Therefore, there is an urgent need to provide a credit large model joint modeling system and method for green financing to solve the above problems. SUMMARY

[0006] The technical problem this invention aims to solve is to overcome the following shortcomings of existing technologies: First, the inherent "illusion" problem of large models leads to the possibility of factual errors or false information in the information they generate. In the high-risk assessment scenario of green credit, such uncertainty will directly trigger credit risk. For example, the model may incorrectly fabricate or confuse the environmental performance data of a project, leading to a misjudgment of the "green" attribute of the project. Second, the "catastrophic amnesia" phenomenon that is common in large models makes it difficult for them to adapt to the rapid iteration of green finance policies, further limiting their practicality. When a new green industry catalog or environmental benefit accounting standard is released, the model has difficulty in effectively integrating new knowledge, which may lead to outdated assessment basis and affect the accuracy and compliance of the assessment results. This invention provides a joint modeling system and method for credit large models oriented towards green financing.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a credit large-scale joint modeling system for green financing, including a multi-source green data governance module, a trusted reasoning and computing module, a large-scale model interaction module, and a dynamic knowledge evolution module. The multi-source green data governance module is used to acquire raw data of green financing projects, generate structured data from the raw data of green financing projects, and generate knowledge constraint signals based on a pre-built green finance policy knowledge base. The credible reasoning and computation module includes a logical reasoning unit and a numerical computation unit. The logical reasoning unit analyzes the structured data to generate a logical decision chain and sends the environmental benefit calculation tasks in the structured data to the numerical computation unit. The numerical computation unit executes the environmental benefit calculation tasks and generates quantitative indicators. The large model interaction module receives the quantitative indicators, the logical decision chain, and the knowledge constraint signal, integrates the quantitative indicators and the logical decision chain into a reasoning context, converts the knowledge constraint signal into an output constraint instruction, and generates a credit assessment report based on the reasoning context and the output constraint instruction through a preset large language model. The dynamic knowledge evolution module receives feedback information from the credit assessment report and updates and optimizes the original data, rules, and large language models of green financing projects in the multi-source green data governance module, the trusted reasoning and computing module, and the large model interaction module.

[0008] The present invention is further configured such that: the original data of the green financing project in the multi-source green data governance module is obtained from corporate environmental reports, government regulatory information and third-party certification data; The steps for generating the structured data are as follows: S1. Extract environmental performance parameters from corporate environmental reports, obtain compliance indicators from government regulatory information, collect certification level information from third-party certification data, and aggregate environmental performance parameters, compliance indicators and certification level information into a pre-set unified data platform. S2. Standardize the data format and unify the units for environmental performance parameters, compliance indicators and certification level information in the unified data platform to generate standard format data; S3. Perform cross-validation on the standard format data, verify the consistency between the environmental performance parameters in the enterprise's environmental report and the compliance indicators matched in the government regulatory information, and use the certification level information in the third-party certification data to corroborate the verification results, and generate the verified data. S4. Based on the rules in the pre-set green finance policy knowledge base, the verified data is verified for compliance. The verified data that passes the verification is used as green attribute data. The green attribute data is labeled with a preset policy compliance tag. The green attribute data is classified and integrated according to the policy compliance tag, and structured data is output.

[0009] The present invention is further configured such that: the specific content of the compliance verification of the verified data based on the rules in the pre-set green finance policy knowledge base in step S4 is as follows: S401. Extract policy clause texts related to the verified data from a pre-set green finance policy knowledge base, and parse the policy clause texts into a set of structured rule units, wherein the set of structured rule units contains the condition judgment logic and compliance judgment benchmark of each structured rule unit. S402. Link each structured rule unit in the set of structured rule units according to the dependency relationship of the condition judgment logic to construct a rule relationship graph. The rule relationship graph includes nodes and connecting edges. The nodes represent the structured rule units, and the connecting edges represent the logical dependency relationship between the structured rule units. S403. Map the key data fields of the verified data to the corresponding nodes in the rule relationship graph, activate the nodes affected by the key data fields, perform logical reasoning propagation along the connection edges of the rule relationship graph, calculate the logical coverage of the activated nodes to the verified data, and perform compliance judgment on the key data fields mapped thereto according to the compliance judgment benchmarks in each node, and generate a preliminary compliance status mark for each key data field. S404. Based on the logical dependency relationship between nodes in the rule relationship diagram, perform preset consistency verification and conflict resolution on multiple preliminary compliance status markers that have logical associations, generate a global consistency verification conclusion for the verified data, and bind the global consistency verification conclusion with the key data field that triggered the global consistency verification conclusion and the information of the corresponding node, as the verified data that has passed the verification. The calculation steps for the logical coverage in step S403 are as follows: S4031. Based on the set of activated nodes in the rule relationship graph, determine the range of all downstream nodes that each activated node can reach in the rule relationship graph, form the maximum influence domain starting from each activated node, merge the maximum influence domains of all activated nodes to obtain the total activation influence domain, calculate the ratio of the total number of nodes in the total activation influence domain to the total number of nodes in the rule relationship graph, and generate the initial node coverage rate. S4032. Analyze the directed paths formed by the connecting edges between the nodes in the total activation influence domain, filter out the effective inference paths starting from the activated node and ending at a point where there are no subsequent nodes and all nodes on the directed path belong to the total activation influence domain, count the ratio of the number of all effective inference paths to the total number of theoretically complete inference paths predefined in the rule relationship graph, generate the effective inference path ratio, and combine the initial node coverage rate with the preset first weight coefficient multiplied by the initial node coverage rate and the preset second weight coefficient multiplied by the effective inference path ratio to obtain the logical coverage.

[0010] The present invention is further configured such that the generation step of the knowledge constraint signal in the multi-source green data governance module is as follows: S5. Based on the verified data, extract the key data fields and node information bound therein, and combine them with the effective reasoning path formed by the activated nodes in the rule relationship graph to identify the set of key constraint elements that play a decisive role in the current green financing project evaluation. S6. Map each constraint element in the set of key constraint elements to a predefined constraint instruction template library, match the constraint instruction template according to the type of the constraint element, and fill the parameter bits of the constraint instruction template with the specific numerical value or logical condition of the constraint element to generate a constraint instruction unit. Aggregate all constraint instruction units to form a knowledge constraint signal.

[0011] The present invention is further configured such that the method for generating the logical decision chain in the trusted reasoning and computation module is as follows: Q1. Extract the green attribute data and the policy compliance label from the structured data as key data elements; Q2. Match the key data elements with the business rules in the preset green finance rule base to identify the set of evaluation rules applicable to the current green financing project; Q3. Based on the evaluation rule set, perform logical reasoning analysis on the key data elements to generate a preliminary logical reasoning sequence. Organize the preliminary logical reasoning sequence in a structured manner according to a preset decision chain format, embed the key data elements as the reasoning basis, and link them to the quantitative indicators generated by the numerical calculation unit to form a logical decision chain. The green finance rule base includes a green and low-carbon transformation industry guidance catalog and a green credit statistical system; The steps for generating the quantitative indicators are as follows: Q4. Receive the evaluation rule set from the key data elements and links embedded in the logical decision chain, identify the environmental benefit parameters and financial risk parameters that need to be accurately calculated, match the corresponding calculation model from the preset environmental benefit calculation model library and financial risk indicator model library according to the parameter type of the environmental benefit parameters and the financial risk parameters, and generate a calculation task instruction set. Q5. Assign the measurement task instruction set to the numerical calculation unit, perform batch processing on the measurement task instruction set, monitor the result confidence level of each numerical calculation unit in real time during the calculation process, recalculate or adjust the parameters of the numerical calculation unit for calculation results whose result confidence level does not reach the preset confidence threshold, until the result confidence level of all calculation results reaches the preset confidence threshold, and generate quantitative indicators.

[0012] The present invention is further configured such that: the specific content of integrating the quantitative indicators and the logical decision chain into the inference context in the large model interaction module is as follows: W1. Parse the logical decision chain, and parse the key data elements, the evaluation rule set and the logical derivation relationship between them into reasoning steps. Then, bind the specific indicator values ​​associated with each reasoning step in the quantitative indicators to the corresponding reasoning steps to generate a binding sequence. W2. Based on the order of the reasoning steps and the data flow in the binding sequence, construct a reasoning topology graph with the reasoning steps as initial nodes and the logical deduction relationship and the data flow as edges; W3. Based on the traversal order of the initial nodes in the inference topology graph, organize the inference steps of each initial node and the specific index values ​​bound to them into a linear narrative flow with temporal logic, and encapsulate it to generate an inference context. The method for generating the output constraint instructions is as follows: W4. Analyze the constraint instruction units contained in the knowledge constraint signal, extract the type, logical condition and specific value of the constraint element in each constraint instruction unit, classify the constraint instruction units into mandatory constraint units and guiding constraint units according to the type of the constraint element and the logical condition, and assign a unique initial constraint strength weight to each constraint instruction unit based on the specific value and the preset constraint strength mapping table. W5. Based on the semantic structure of the reasoning context, determine the applicable positions of the mandatory constraint unit and the guiding constraint unit in the semantic structure, fuse the initial constraint strength weight with the relevant weight of the reasoning context calculated based on the applicable position to generate the final constraint strength, and convert the constraint instruction unit into a natural language instruction fragment according to the final constraint strength, and aggregate all the natural language instruction fragments to form an output constraint instruction.

[0013] The present invention is further configured such that the step of generating the final constraint strength is as follows: W501. For each constraint instruction unit, based on its applicable position in the semantic structure of the inference context, analyze the logical level of the applicable position in the inference topology graph and the degree of connection between the applicable position and the key initial node, and calculate the position weight factor. W502. The position weight factor and the initial constraint strength weight are weighted and fused to obtain an initial fusion value. The initial fusion value is normalized to generate a basic constraint strength. The logical dependency, conditional mutual exclusion, or conclusion progression relationship between the mandatory constraint unit and the guiding constraint unit is identified. When the mandatory constraint unit and the guiding constraint unit have a logical dependency, the initial constraint strength weight of the constraint instruction unit that depends on other mandatory constraint units or other guiding constraint units is multiplied by a predefined enhancement coefficient. When the mandatory constraint unit and the guiding constraint unit have a conditional mutual exclusion, the initial constraint strength weight of the mutually exclusive constraint instruction units is multiplied by a predefined weakening coefficient. When multiple constraint instruction units have a conclusion progression relationship, the initial constraint strength of the subsequent constraint instruction unit is enhanced and adjusted based on its progressive logical association with the preceding constraint instruction unit. The final constraint strength is generated based on the result calculated by the above relationship and the result of the enhancement adjustment.

[0014] The present invention is further configured such that the specific content of the credit assessment report generated in the large model interaction module is as follows: the reasoning context and the output constraint instruction are input into a preset large language model, driving the large language model to perform semantic understanding and logical deduction based on the linear narrative flow in the reasoning context, and the text generation process is regulated by the final constraint strength of the natural language instruction fragment in the output constraint instruction, and the credit assessment report is output. The credit assessment report includes a green credit rating, environmental benefit assessment, risk warning, and policy compliance statement.

[0015] The present invention is further configured such that: the specific content of the dynamic knowledge evolution module is as follows: receiving external verification feedback information of the credit assessment report in actual application scenarios, comparing and analyzing the feedback information with the reasoning context and output constraint instructions that generated the report, identifying new green attributes or policy update points, triggering the adjustment of data source weights and verification rule updates of the original data of green financing projects in the multi-source green data governance module, optimizing business rules of the green finance rule base and calibrating parameters of the environmental benefit calculation model base in the trusted reasoning and calculation module, and incrementally training the pre-set large language model in the large model interaction module based on the labeled data generated by the comparison analysis.

[0016] A joint modeling method for credit large-scale models for green financing includes the following steps: Step 1: Obtain raw data of green financing projects, generate structured data from the raw data of green financing projects, and generate knowledge constraint signals based on a pre-built green finance policy knowledge base; Step 2: Analyze the structured data to generate a logical decision chain, and execute the tasks related to environmental benefit calculation in the structured data to generate quantitative indicators; Step 3: Integrate the quantitative indicators and the logical decision chain into a reasoning context, convert the knowledge constraint signals into output constraint instructions, and generate a credit assessment report based on the reasoning context and the output constraint instructions using a pre-set large language model; Step 4: Receive feedback information from the credit assessment report, and update and optimize the original data, rules, and large language model of green financing projects in the multi-source green data governance module, the trusted reasoning and computing module, and the large model interaction module.

[0017] The beneficial effects of this invention are as follows: 1. This invention generates structured data through multi-source green data governance, combines logical reasoning and numerical calculation to generate logical decision chains and quantitative indicators, and constructs a reasoning context based on these. At the same time, the large model interaction module uses output constraint instruction specification text generation to effectively reduce the "illusion" of the large model and improve the accuracy of credit assessment. 2. This invention receives feedback from credit assessment reports through a dynamic knowledge evolution module, updates and optimizes data sources, verification rules, business rules, and calculation models, and performs incremental training on the large language model to overcome "catastrophic forgetting" and enable it to adapt to the rapid iteration of green finance policies. Attached Figure Description

[0018] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0020] Please see Figure 1 - Figure 2 A credit big data joint modeling system for green financing includes a multi-source green data governance module, a trusted reasoning and computing module, a big data interaction module, and a dynamic knowledge evolution module. The multi-source green data governance module is used to acquire raw data of green financing projects, generate structured data from the raw data of green financing projects, and generate knowledge constraint signals based on a pre-built green finance policy knowledge base. The credible reasoning and computation module includes a logical reasoning unit and a numerical computation unit. The logical reasoning unit analyzes the structured data to generate a logical decision chain and sends the tasks related to environmental benefit measurement in the structured data to the numerical computation unit. The numerical computation unit executes the environmental benefit measurement tasks and generates quantitative indicators. The large model interaction module receives quantitative indicators, logical decision chains, and knowledge constraint signals. It integrates the quantitative indicators and logical decision chains into a reasoning context, transforms the knowledge constraint signals into output constraint instructions, and generates a credit assessment report based on the reasoning context and output constraint instructions through a pre-built large language model. The dynamic knowledge evolution module receives feedback information from credit assessment reports and updates and optimizes the original data, rules, and large language models of green financing projects in the multi-source green data governance module, the trusted reasoning and computing module, and the large model interaction module.

[0021] This system ensures data credibility through multi-source green data governance, generates accurate quantitative indicators and logical decision chains through credible reasoning and computation, outputs reliable credit assessment reports under knowledge constraints through a large model interaction module, and achieves continuous system optimization through a dynamic knowledge evolution module, effectively improving the accuracy, reliability and adaptability of green financing credit assessment.

[0022] One embodiment of the present invention is as follows: the original data of green financing projects in the multi-source green data governance module are obtained from corporate environmental reports, government regulatory information and third-party certification data; The steps for generating structured data are as follows: S1. Extract environmental performance parameters from corporate environmental reports, obtain compliance indicators from government regulatory information, collect certification level information from third-party certification data, and aggregate environmental performance parameters, compliance indicators and certification level information into a pre-set unified data platform. The unified data platform is a central data processing environment that integrates data collection and standardized preprocessing engines. The platform has built-in multi-source adapters for accessing heterogeneous data sources such as enterprise environmental reports, government regulatory information and third-party certification data, and is equipped with a metadata management component to uniformly catalog and trace the imported environmental performance parameters, compliance indicators and certification level information. S2. Standardize the data format and unify the units of environmental performance parameters, compliance indicators and certification level information in the unified data platform to eliminate semantic ambiguity and measurement differences between multi-source heterogeneous data and generate standard format data. S3. Perform cross-validation on standard format data, verify the consistency between the environmental performance parameters in the enterprise's environmental report and the compliance indicators matched in the government regulatory information, and use the certification level information in the third-party certification data to corroborate the verification results and generate verified data. S4. Based on the rules in the pre-built green finance policy knowledge base, the verified data is verified for compliance. The verified data that passes the verification is used as green attribute data, and a preset policy compliance label is attached to the green attribute data. The green attribute data is classified and integrated according to the policy compliance label, and structured data is output.

[0023] If the verification fails in step S4, the verified data is marked as data to be processed and stored in a temporary isolation area. At the same time, the abnormal handling process is triggered to generate a detailed audit report containing specific violation clauses and data conflict points. The audit report is then fed back to the data collection end for source verification and data cleaning. The cleaned data is then re-entered into the unified data platform for standard format data generation steps. Policy compliance label: A machine-readable semantic identifier generated based on the global consistency verification conclusion. The label data structure includes the code of the policy clause that is complied with, the verification timestamp, the compliance level score, and the hash value of the key data field to which it is bound. It is used to quickly retrieve and match the policy basis corresponding to the green attribute data in subsequent processes. Specifically, step S4 involves verifying the compliance of the validated data based on rules from a pre-built green finance policy knowledge base. S401. Extract policy clause texts related to the verified data from the pre-built green finance policy knowledge base, parse the policy clause texts into a set of structured rule units, and the set of structured rule units contains the condition judgment logic and compliance judgment benchmark of each structured rule unit. Green Finance Policy Knowledge Base: A policy knowledge storage system organized using a semantic graph structure. This knowledge base decomposes various green finance policy provisions into structured rule units containing conditional judgment logic and compliance judgment benchmarks, and links them through logical dependencies to form a traversable rule relationship graph. It is also equipped with a version control mechanism to track the revision history of policy provisions. S402. Link the structured rule units in the set of structured rule units according to the dependency relationship of the condition judgment logic to construct a rule relationship graph. The rule relationship graph includes nodes and connecting edges. Nodes represent structured rule units, and connecting edges represent the logical dependency relationship between structured rule units. S403. Map the key data fields of the verified data to the corresponding nodes in the rule relationship graph, activate the nodes affected by the key data fields, perform logical reasoning propagation along the connection edges of the rule relationship graph, calculate the logical coverage of the activated nodes to the verified data, and perform compliance judgment on the key data fields mapped to them according to the compliance judgment benchmarks in each node, and generate a preliminary compliance status mark for each key data field. The steps for calculating the logic coverage in step S403 are as follows: S4031. Based on the set of activated nodes in the rule relationship graph, determine the range of all downstream nodes that each activated node can reach in the rule relationship graph, form the maximum influence domain starting from each activated node, merge the maximum influence domains of all activated nodes to obtain the total activation influence domain, calculate the ratio of the total number of nodes in the total activation influence domain to the total number of nodes in the rule relationship graph, and generate the initial node coverage rate. S4032. Analyze the directed paths formed by the connecting edges between nodes in the total activation influence domain, filter out the effective inference paths that start from the activated node and have no subsequent nodes and whose nodes all belong to the total activation influence domain, count the ratio of the number of all effective inference paths to the total number of theoretically complete inference paths predefined in the rule relationship graph, generate the effective inference path ratio, combine the initial node coverage rate, multiply the preset first weight coefficient with the initial node coverage rate and add the preset second weight coefficient with the effective inference path ratio to obtain the logic coverage. The first and second weight coefficients are obtained as follows: The first weight coefficient is determined based on the basic importance level of the structured rule unit represented by the node in the total activation influence domain in the policy logic, and the second weight coefficient is determined based on the logical depth and decision-making influence level of the key policy clauses covered by the effective reasoning path. S404. Based on the logical dependencies between nodes in the rule relationship graph, perform pre-set consistency checks and conflict resolution on multiple preliminary compliance status markers that have logical connections, generate a global consistency verification conclusion for the verified data, and bind the global consistency verification conclusion with the key data fields that triggered the global consistency verification conclusion and the information of the corresponding nodes as the verified data that has passed the verification. Based on the logical dependencies between nodes in the rule relationship graph, the specific content of the pre-set consistency verification and conflict resolution for multiple preliminary compliance status markers with logical connections is as follows: First, a conflict detection graph is constructed with the preliminary compliance status markers as vertices and the logical dependencies between nodes as edges. The conflict detection graph is traversed to identify the set of vertex pairs with logical contradictions. For each contradictory vertex pair, according to the hierarchical depth of its corresponding node in the rule relationship graph and the priority of the cited policy clauses, the pre-set conflict resolution rules (such as higher-level laws taking precedence over lower-level laws, and new regulations taking precedence over old regulations) are applied to coordinate and cover the status markers, and finally generate a globally consistent verification conclusion without contradictions.

[0024] The steps for generating knowledge constraint signals in the multi-source green data governance module are as follows: S5. Based on the verified data, extract the information of the key data fields and nodes bound therein, and combine them with the effective reasoning path formed by the activated nodes in the rule relationship graph to identify the set of key constraint elements that play a decisive role in the current green financing project evaluation. The specific content of step S5 is as follows: Based on the verified data and its bound key data fields and node information, calculate the centrality index of each activated node in the rule relationship graph and the policy importance weight of the rule unit it represents. Sort and select the top N nodes according to the product of the centrality index and the policy importance weight to form a set of key constraint elements. Each element in the set of key constraint elements encapsulates the core policy variables that affect the final compliance judgment and their value constraints. The steps for calculating the centrality index of each activated node in the rule relationship graph and the policy importance weight of the rule unit it represents are as follows: a. Based on the topological connectivity of nodes in the regular relationship graph, calculate the local centrality index of each activated node. The local centrality index is obtained by calculating the ratio of the number of first-order neighbor nodes directly connected to the node to the average degree of nodes in the entire regular relationship graph. At the same time, the sum of the reciprocals of the shortest path lengths from the node to all other reachable nodes in the regular relationship graph is calculated as the proximity centrality. The initial connectivity and proximity centrality are weighted and summed to obtain the basic centrality index. b. Based on the basic centrality indicator, the influence factors of the preset policy logic dimension are introduced for correction. The policy clause effectiveness level (such as national level, ministerial level, etc.) corresponding to the structured rule unit represented by the node in the green finance policy knowledge base is extracted and quantified into a level weight coefficient. At the same time, the importance level of the green finance goal (such as carbon emission reduction, pollution prevention and control) associated with the structured rule unit is considered and quantified into a target weight coefficient. The basic centrality indicator, the level weight coefficient, and the target weight coefficient are weighted and integrated to generate a comprehensive centrality indicator. c. Calculate the policy importance weight of the rule unit. First, obtain its policy priority calibration value based on the attribute definition of the rule unit in the green finance policy knowledge base. Then, calculate its activity coefficient by combining the frequency of triggering of the rule unit in historical verification data. Multiply the policy priority calibration value and the activity coefficient and normalize them to generate the policy importance weight. d. Multiply the comprehensive centrality index of each activated node by the policy importance weight of its representative rule unit to obtain the final importance score of the node, and sort all activated nodes in descending order according to the final importance score; S6. Map each constraint element in the set of key constraint elements to a predefined constraint instruction template library. Match the constraint instruction template according to the type of constraint element, and fill the parameter bits of the constraint instruction template with the specific numerical value or logical condition of the constraint element to generate a constraint instruction unit with clear semantics. Aggregate all constraint instruction units to form a knowledge constraint signal. The knowledge constraint signal will be used as a boundary condition input to subsequent modules to reduce the generation space of the large language model.

[0025] A predefined constraint instruction template library contains a collection of parameterized natural language statement templates for various constraint types. Each template consists of a constraint type identifier, a fixed sentence structure, and slottable parameter bits. For example, for threshold constraints, the template structure is "The value of item [parameter A] must be [comparison operator] [threshold X]". Once the specific values ​​of the constraint elements are filled into the parameter bits, machine-readable and human-understandable constraint instruction units can be generated. Example: Suppose a green financing project has an environmental report showing a carbon emission intensity of 85 kg CO2 / 10,000 yuan. The compliance judgment benchmark set for the corresponding node in the rule relationship diagram is ≤100 kg CO2 / 10,000 yuan and ≥50 kg CO2 / 10,000 yuan to obtain the "Grade A" label. When calculating the logical coverage, the initial node coverage rate is 85% (e.g., 17 / 20 nodes are activated), and the effective inference path ratio is 70% (e.g., 7 / 10 paths are effective). Let the first weight coefficient be 0.6 and the second weight coefficient be 0.4, then the logical coverage = 0.6 * 85% + 0.4 * 70% = 79%.

[0026] Since 85 kg CO2 / 10,000 yuan is within the compliance range, the data was marked as initially compliant, and the constraint instruction unit "the project's carbon emission intensity must be maintained at no more than 100 kg CO2 / 10,000 yuan and no less than 50 kg CO2 / 10,000 yuan" was finally generated.

[0027] Beneficial effects: This embodiment transforms qualitative policy provisions into calculable indicators by quantitatively calculating the coverage of logic and comparing precise thresholds, effectively eliminating the uncertainty of large models in judging environmental data; the use of rule relationship diagrams and constraint instruction templates injects dynamic policy requirements into the assessment process in a structured manner, significantly improving the accuracy, interpretability and adaptability to policy changes in green credit assessment.

[0028] One embodiment of the present invention is as follows: the method for generating the logical decision chain in the trusted reasoning and computation module is as follows: Q1. Extract green attribute data and policy compliance labels from structured data as key data elements; The key data element extraction steps are as follows: scan all data items carrying policy compliance labels from structured data, group and classify the data items according to the policy clause codes in the policy compliance labels, and then vertically aggregate them according to the classification dimensions of green attribute data (such as environmental benefits, compliance, and certification levels) to form a set of key data elements indexed by financing projects; Q2. Match key data elements with business rules in the pre-set green finance rule base to identify the set of assessment rules applicable to the current green financing project; Q3. Based on the evaluation rule set, perform logical reasoning analysis on key data elements to generate a preliminary logical reasoning sequence. Organize the preliminary logical reasoning sequence in a structured manner according to the preset decision chain format, embed key data elements as the basis for reasoning, and link them to the quantitative indicators generated by the numerical calculation unit to form a logical decision chain. Step Q3 involves logical reasoning analysis of key data elements, which includes determining the green attributes of the project, analyzing the correlation between environmental benefits and potential risks. First, the green attribute compliance of key data elements is determined according to the industry catalog matching rules in the assessment rule set. Then, the correlation between environmental benefit indicators and financing scale is analyzed based on the determination results. Finally, potential risk points of the project in terms of technology route, policy dependence, etc. are scanned through risk identification rules to generate a preliminary logical reasoning sequence that includes determination results, correlation analysis and risk marking.

[0029] The green finance rules library includes a green and low-carbon transformation industry guidance catalog and a green credit statistical system; The steps for generating quantitative indicators are as follows: Q4. Receive the evaluation rule set from the key data elements and links embedded in the logical decision chain, identify the environmental benefit parameters and financial risk parameters that need to be accurately calculated, match the corresponding calculation model from the preset environmental benefit calculation model library and financial risk indicator model library according to the parameter type of the environmental benefit parameters and financial risk parameters, and generate a calculation task instruction set. The pre-set environmental benefit measurement model library specifically includes a collection of standardized environmental benefit measurement models such as life cycle assessment models, Clean Development Mechanism methodology models, and equal emission coefficient method models; the financial risk indicator model library specifically includes a combination of traditional financial risk quantification models such as cash flow coverage ratio models, sensitivity analysis models, and Monte Carlo simulation models; the calculation models include statistical learning models such as linear regression models, time series analysis models, and logistic regression models. Q5. Assign the measurement task instruction set to the numerical calculation unit, perform batch processing on the measurement task instruction set, monitor the confidence level of the results of each numerical calculation unit in real time during the calculation process, recalculate or adjust the parameters of the numerical calculation unit for the calculation results whose confidence level does not reach the preset confidence threshold, until the confidence level of all calculation results reaches the preset confidence threshold, and generate quantitative indicators.

[0030] The numerical computation unit is specifically based on a GPU-accelerated parallel computing architecture, equipped with a floating-point operation coprocessor and an error control module; the optimal value of the preset confidence threshold is: determined by the distribution of the degree of agreement between the model output results and the actual observations in the historical validation data, the minimum confidence threshold of 0.92 required to achieve a statistical significance level of 95%.

[0031] This embodiment establishes a dynamic correlation mechanism between logical decision-making chains and quantitative indicators, deeply integrating qualitative reasoning and quantitative calculation. This ensures both the professionalism of environmental benefit assessment and the accuracy of risk identification, significantly improving the interpretability and reliability of green financing credit assessment.

[0032] One embodiment of the present invention is as follows: the specific content of integrating quantitative indicators and logical decision chains into an inference context in the large model interaction module is as follows: W1. Parse the logical decision chain, analyze the key data elements, evaluation rule set and logical derivation relationship contained therein into reasoning steps, and bind the specific index values ​​associated with each reasoning step in the quantitative indicators to the corresponding reasoning steps to generate a binding sequence. The reasoning steps are analyzed as follows: The logical decision chain is analyzed, and each reasoning step is decomposed into specific values ​​of key data elements, specific business rule clauses referenced by the evaluation rule set, and preconditions and conclusion statements in the logical deduction relationship. Then, a unique identifier is created for each reasoning step, and the specific indicator values ​​related to the reasoning step in the quantitative indicators are bound through the unique identifier to form a binding sequence containing data, rules and logical relationships. W2. Based on the order of reasoning steps and the direction of data flow in the binding sequence, construct a reasoning topology graph with reasoning steps as initial nodes and logical deduction relationships and data flow as edges. W3. Based on the traversal order of the initial nodes in the reasoning topology graph, organize the reasoning steps of each initial node and the specific index values ​​bound to them into a linear narrative flow with temporal logic, and encapsulate it to generate a reasoning context. The method for generating output constraint instructions is as follows: W4. Analyze the constraint instruction units contained in the knowledge constraint signal, extract the type, logical condition, and specific value of the constraint element in each constraint instruction unit, classify the constraint instruction units into mandatory constraint units and guiding constraint units according to the type and logical condition of the constraint element, and assign a unique initial constraint strength weight to each constraint instruction unit based on the specific value and a preset constraint strength mapping table; the preset constraint strength mapping table is as follows: Constraint Strength Mapping Table: A mapping table based on the constraint element type and the predefined weight values ​​of the specific numerical range. Its rows correspond to the constraint element type (such as numeric type, logical type), and its columns correspond to the normalized segment interval of the specific numerical value. The initial constraint strength weight value stored in each cell is obtained by statistically analyzing the correction effect of this type of constraint on the model output deviation in historical data, and is used for fast matching and assignment. W5. Based on the semantic structure of the reasoning context, determine the applicable positions of mandatory constraint units and guiding constraint units in the semantic structure, fuse the initial constraint strength weights with the relevant weights of the reasoning context calculated based on the applicable positions to generate the final constraint strength, and convert the constraint instruction units into natural language instruction fragments with specific formats and scopes according to the final constraint strength, and aggregate all natural language instruction fragments to form the output constraint instruction.

[0033] The relevant weight calculation steps are as follows: First, determine the applicable position of the constraint instruction unit in the semantic structure of the inference context. Then, calculate the product of the logical level number of the position in the inference topology graph (the shortest path length from the root node to the position) and the number of key initial nodes connected to the position. Then divide by the total number of nodes in the inference topology graph to obtain the unnormalized relevant weight value. Finally, scale the value to the [0,1] interval using the minimum-maximum normalization method to obtain the relevant weight. The initial constraint strength weights are fused with the relevant weights of the inference context calculated based on the applicable location to generate the final constraint strength. The specific content is as follows: a weighted average method is used for fusion, where the weight coefficient of the initial constraint strength weights is set to 0.6 and the weight coefficient of the relevant weights of the inference context is set to 0.4. After fusion, the original strength value is obtained, and then a sigmoid function is applied for smoothing so that the final constraint strength value falls within the range of 0 to 1 and highlights the impact of key constraints.

[0034] The steps for generating the final constraint strength are as follows: W501. For each constraint instruction unit, based on its applicable position in the semantic structure of the reasoning context, analyze the logical level of the applicable position in the reasoning topology graph and the degree of connection with the key initial node, and calculate the position weight factor. The position weight factor is used to characterize the importance of its reasoning context. The location weight factor is calculated as follows: the level factor is calculated based on the logical level of the applicable location in the inference topology (the formula is 1 / logical level). At the same time, the connection tightness between the location and the key initial node is calculated (the formula is the number of common neighbor nodes divided by the total number of key initial nodes). The level factor and the connection tightness are multiplied to obtain the original weight, and then divided by the maximum possible weight value of all locations in the inference topology for normalization, and the location weight factor is output. W502: The position weight factor and the initial constraint strength weight are weighted and fused to obtain the initial fusion value. The initial fusion value is normalized to generate the basic constraint strength. The logical dependency, conditional mutual exclusion, or conclusion progression relationship between mandatory constraint units and guiding constraint units is identified. When there is a logical dependency between mandatory constraint units and guiding constraint units, the initial constraint strength weight of the constraint instruction unit that depends on other mandatory constraint units or other guiding constraint units is multiplied by a predefined enhancement coefficient. When there is a conditional mutual exclusion between mandatory constraint units and guiding constraint units, the initial constraint strength weight of the mutually exclusive constraint instruction units is multiplied by a predefined weakening coefficient. When there is a conclusion progression relationship between multiple constraint instruction units, the initial constraint strength of the subsequent constraint instruction unit is enhanced and adjusted based on its degree of progressive logical association with the preceding constraint instruction unit. The final constraint strength is generated based on the results calculated by the above relationships and the results of the enhancement adjustment.

[0035] Predefined enhancement coefficient: A constant determined by regression analysis of the effect of logical dependencies on the constraint strength improvement in historical data. Its value is fixed at 1.3. When there are logical dependencies between constraint instruction units, this coefficient is multiplied by the initial constraint strength weight to achieve enhancement. Predefined weakening coefficient: A constant calculated based on the frequency of constraint failure caused by mutually exclusive conditions in historical data. Its value is fixed at 0.6. When there are mutually exclusive conditions in the constraint instruction unit, this coefficient is multiplied by the initial constraint strength weight to achieve weakening. When multiple constraint instruction units have a progressive relationship in their conclusions, the specific content of enhancing and adjusting the initial constraint strength of the subsequent constraint instruction unit based on its progressive logical association with the preceding constraint instruction unit is as follows: First, calculate the shortest path length between the subsequent constraint instruction unit and the preceding constraint instruction unit in the reasoning topology graph as the association distance. The smaller the association distance, the higher the degree of progressive logical association. Set the enhancement adjustment coefficient as (1 - association distance / maximum possible distance). Multiply the initial constraint strength of the subsequent constraint instruction unit by this adjustment coefficient to obtain the enhanced strength value.

[0036] This embodiment generates a structured reasoning context by parsing the logical decision chain, and generates precise output constraint instructions based on constraint strength mapping and position weights. This ensures that the large language model strictly follows data logic and business rules in credit assessment, effectively eliminating model illusions and improving the accuracy and reliability of assessment reports. At the same time, dynamic weight adjustment enhances the system's adaptability.

[0037] The specific content of the credit assessment report generated in the large model interaction module is as follows: the reasoning context and output constraint instructions are input into the pre-set large language model, which drives the large language model to perform semantic understanding and logical deduction based on the linear narrative flow in the reasoning context, and its text generation process is regulated by the final constraint strength of the natural language instruction fragments in the output constraint instructions, and outputs the credit assessment report. The credit assessment report includes a green credit rating, environmental benefit assessment, risk warning, and policy compliance statement.

[0038] One embodiment of the present invention is as follows: The specific content of the dynamic knowledge evolution module is as follows: receiving external verification feedback information of credit assessment reports in actual application scenarios, comparing and analyzing the feedback information with the reasoning context and output constraint instructions that generated the report, identifying new green attributes or policy update points, triggering the adjustment of data source weights and verification rule updates of the original data of green financing projects in the multi-source green data governance module, optimizing business rules of the green finance rule base in the trusted reasoning and calculation module and calibrating parameters of the environmental benefit measurement model base, and incrementally training the pre-set large language model in the large model interaction module based on the labeled data generated by the comparison analysis, forming a closed-loop model self-optimization capability.

[0039] This embodiment dynamically optimizes data weights, business rules, and model parameters through closed-loop comparison between external verification feedback and internal system reasoning, enabling the system to have continuous learning capabilities and effectively improving the accuracy, adaptability, and responsiveness to policy changes in green credit assessment.

[0040] A joint modeling method for credit large-scale models for green financing includes the following steps: Step 1: Obtain raw data of green financing projects, generate structured data from the raw data of green financing projects, and generate knowledge constraint signals based on the pre-built green finance policy knowledge base; Step 2: Analyze the structured data to generate a logical decision chain, and execute the tasks involving environmental benefit calculation in the structured data to generate quantitative indicators; Step 3: Integrate quantitative indicators and logical decision chains into a reasoning context, transform knowledge constraint signals into output constraint instructions, and generate a credit assessment report based on the reasoning context and output constraint instructions using a pre-built large language model; Step 4: Receive feedback information from the credit assessment report and update and optimize the original data, rules, and large language models of green financing projects in the multi-source green data governance module, trusted reasoning and computing module, and large model interaction module.

[0041] The above are merely embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A green finance-oriented credit large model joint modeling system, characterized in that: The system comprises a multi-source green data management module, a trusted reasoning and calculation module, a large model interaction module, and a dynamic knowledge evolution module. The multi-source green data management module is configured to obtain green financing project raw data, generate structured data from the green financing project raw data, and generate knowledge constraint signals based on a preconfigured green finance policy knowledge base. The trusted reasoning and calculation module comprises a logical reasoning unit and a numerical calculation unit. The logical reasoning unit analyzes the structured data to generate a logical decision chain and sends tasks related to environmental benefit calculation in the structured data to the numerical calculation unit. The numerical calculation unit executes the environmental benefit calculation tasks to generate quantitative indicators. 2.The green finance-oriented credit model joint modeling system according to claim 1, characterized in that: The large model interaction module receives the quantitative indicators, the logical decision chain, and the knowledge constraint signals, integrates the quantitative indicators and the logical decision chain into reasoning context, converts the knowledge constraint signals into output constraint instructions, and generates a credit evaluation report based on the reasoning context and the output constraint instructions through a preconfigured large language model. The dynamic knowledge evolution module receives feedback information of the credit evaluation report and updates and optimizes green financing project raw data, rules, and large language models in the multi-source green data management module, the trusted reasoning and calculation module, and the large model interaction module. The green financing project raw data in the multi-source green data management module is obtained from enterprise environmental reports, government regulatory information, and third-party certification data. The structured data is generated by: S1. Extracting environmental performance parameters from enterprise environmental reports, obtaining compliance indicators from government regulatory information, and collecting certification level information from third-party certification data. S2. Standardizing data formats and units of environmental performance parameters, compliance indicators, and certification level information in the unified data platform to generate standard format data. 3.The green finance-oriented credit model joint modeling system according to claim 2, characterized in that: S3. Cross-verifying the standard format data, checking the consistency of environmental performance parameters in enterprise environmental reports and matching compliance indicators in government regulatory information, and using certification level information in third-party certification data to verify the results to generate verified data. S4. Verifying the verified data based on rules in the preconfigured green finance policy knowledge base, and classifying and integrating the green attribute data based on the policy compliance label to output structured data. The specific content of verifying the verified data based on rules in the preconfigured green finance policy knowledge base in step S4 is: S401. Extracting policy clause texts related to the verified data from the preconfigured green finance policy knowledge base, parsing the policy clause texts into a structured rule unit set, and the structured rule unit set contains the condition judgment logic and compliance judgment benchmark of each structured rule unit. S402, link each of the structured rule units in the structured rule unit set according to the dependency relationship of the conditional judgment logic, and construct a rule relationship graph, wherein the rule relationship graph comprises nodes and connecting edges, the nodes represent the structured rule units, and the connecting edges represent the logical dependency relationship between the structured rule units; S403, map the key data fields of the verified data to the corresponding nodes in the rule relationship graph, activate the nodes affected by the key data fields, perform logical reasoning propagation along the connecting edges of the rule relationship graph, calculate the logical coverage degree of the activated nodes on the verified data, and perform compliance judgment on the key data fields mapped thereto according to the compliance judgment criteria in each node, to generate a preliminary compliance state mark of each key data field; S404, based on the logical dependency relationship between the nodes in the rule relationship graph, perform preset consistency checking and conflict resolution on a plurality of preliminary compliance state marks with logical association, to generate a global consistency verification conclusion for the verified data, and bind the global consistency verification conclusion with the key data fields triggering the global consistency verification conclusion and the information of the corresponding nodes, as the verified data passing the verification; The calculation step of the logical coverage degree in the step S403 is: S4031, according to the set of activated nodes in the rule relationship graph, determine the range of all downstream nodes that can be reached by each activated node in the rule relationship graph, form the maximum influence domain with each activated node as the starting point, merge the maximum influence domains of all activated nodes to obtain a total activated influence domain, and count the ratio of the total number of nodes contained in the total activated influence domain to the total number of nodes in the rule relationship graph, to generate an initial node coverage rate; S4032, analyze the directed paths formed by the connecting edges between the nodes in the total activated influence domain, filter out valid reasoning paths starting from the activated nodes to nodes without subsequent nodes, and the nodes on the directed paths all belong to the total activated influence domain, count the proportion of the number of all valid reasoning paths to the total number of theoretical complete reasoning paths predefined in the rule relationship graph, to generate a valid reasoning path proportion, combine the initial node coverage rate, multiply a preset first weight coefficient by the initial node coverage rate, and multiply a preset second weight coefficient by the valid reasoning path proportion to obtain a logical coverage degree.

4. The green finance-oriented credit macro model joint modeling system according to claim 3, characterized in that: The generation step of the knowledge constraint signal in the multi-source green data governance module is: S5, based on the verified data, extract the information of the key data fields and the nodes bound therein, combine the valid reasoning paths formed by the activated nodes in the rule relationship graph, and identify a key constraint element set that plays a decisive role in the current green financing project evaluation; S6, mapping each constraint element in the set of key constraint elements to a pre-defined constraint instruction template library, matching a constraint instruction template according to the type of the constraint element, and filling the specific numerical value or logical condition of the constraint element to the parameter bit of the constraint instruction template to generate a constraint instruction unit, aggregating all constraint instruction units to form a knowledge constraint signal. 5.The green finance-oriented credit model joint modeling system according to claim 4, characterized in that: The generation method of the logical decision chain in the trusted reasoning and calculation module is as follows: Q1, extracting the green attribute data and the policy compliance label from the structured data as key data elements; Q2, matching the key data elements with the business rules in the pre-set green finance rule library to identify the evaluation rule set applicable to the current green financing project; Q3, based on the evaluation rule set, performing logical reasoning analysis on the key data elements to generate a preliminary logical reasoning sequence, structuring the preliminary logical reasoning sequence according to the pre-set decision chain format, embedding the key data elements as reasoning basis, and linking to the quantitative indicators generated by the numerical calculation unit to form a logical decision chain; The green finance rule library includes the green low-carbon transformation industry guidance directory and the green credit statistical system; The generation steps of the quantitative indicators are: Q4, receiving the key data elements embedded in the logical decision chain and the evaluation rule set linked, identifying the environmental benefit parameters and financial risk parameters that need to be accurately calculated, matching the corresponding calculation model from the pre-set environmental benefit calculation model library and financial risk indicator model library according to the parameter type of the environmental benefit parameters and the financial risk parameters, and generating a calculation task instruction set; Q5, distributing the calculation task instruction set to the numerical calculation unit, performing batch processing operation on the calculation task instruction set, and monitoring the result confidence of each numerical calculation unit in real time during the operation process, re-operating or adjusting the parameters of the numerical calculation unit for the calculation results whose result confidence does not reach the pre-set confidence threshold, until the result confidence of all calculation results reaches the pre-set confidence threshold, and generating quantitative indicators. 6.The green finance-oriented credit model joint modeling system according to claim 5, characterized in that: The specific content of integrating the quantitative indicators and the logical decision chain into reasoning context in the large model interaction module is: W1, analyzing the logical decision chain, parsing the key data elements, the evaluation rule set and the logical derivation relationship therebetween into reasoning steps, and binding the specific indicator values associated with each reasoning step in the quantitative indicators to the corresponding reasoning step to generate a binding sequence; W2, based on the order and data flow direction between the reasoning steps in the binding sequence, constructing a reasoning topology graph with the reasoning steps as initial nodes and the logical derivation relationship and the data flow direction as edges; W3, according to the traversal order of the initial nodes of the reasoning topology graph, organizing the reasoning steps of each initial node and the specific indicator values bound thereto into a linear narrative stream with time sequence logic, and encapsulating to generate a reasoning context; The generation method of the output constraint instruction is: W4, parsing the constraint instruction units contained in the knowledge constraint signal, extracting the type, logical condition and specific value of the constraint elements in each constraint instruction unit, classifying the constraint instruction units into mandatory constraint units and guiding constraint units according to the type and logical condition of the constraint elements, and assigning a unique initial constraint strength weight to each constraint instruction unit based on the specific value and a preset constraint strength mapping table; W5, determining the applicable positions of the mandatory constraint units and the guiding constraint units in the semantic structure of the reasoning context, fusing the initial constraint strength weight with the relevant weight of the reasoning context calculated based on the applicable positions, generating a final constraint strength, and converting the constraint instruction units into natural language instruction segments according to the final constraint strength, and aggregating all the natural language instruction segments to form an output constraint instruction.

7. The green finance-oriented credit macro model joint modeling system according to claim 6, characterized in that: The generation step of the final constraint strength is: W501, for each constraint instruction unit, according to its applicable position in the semantic structure of the reasoning context, analyzing the logical level of the applicable position in the reasoning topology graph and its connection closeness to key initial nodes, and calculating a position weight factor; W502, weighting and fusing the position weight factor and the initial constraint strength weight to obtain an initial fusion value, and generating a basic constraint strength by normalizing the initial fusion value; identifying the logical dependency relationship, conditional mutual exclusion relationship or conclusion progressive relationship between the mandatory constraint units and the guiding constraint units, when the mandatory constraint units and the guiding constraint units have a logical dependency relationship, multiplying the initial constraint strength weight of the constraint instruction units dependent on other mandatory constraint units or other guiding constraint units by a predefined enhancement coefficient, when the mandatory constraint units and the guiding constraint units have a conditional mutual exclusion relationship, multiplying the initial constraint strength weights of the constraint instruction units that mutually exclude each other by a predefined weakening coefficient, when multiple constraint instruction units have a conclusion progressive relationship, enhancing and adjusting the initial constraint strength of the subsequent constraint instruction units based on their progressive logical association degree with the preceding constraint instruction units, and generating a final constraint strength based on the results of the above relationship calculation and enhancement adjustment. 8.The green finance-oriented credit model joint modeling system according to claim 7, characterized in that: The specific content of the generation of the credit evaluation report in the large model interaction module is: inputting the reasoning context and the output constraint instruction into a pre-set large language model, driving the large language model to perform semantic understanding and logical deduction based on the linear narrative flow in the reasoning context, and regulating its text generation process based on the final constraint strength of the natural language instruction segments in the output constraint instruction, and outputting a credit evaluation report; The credit evaluation report includes green credit rating, environmental benefit evaluation, risk warning and policy compliance statement. 9.The green finance-oriented credit model joint modeling system according to claim 8, characterized in that: The specific content of the dynamic knowledge evolution module is: receiving the external verification feedback information of the credit evaluation report in the actual application scene, comparing and analyzing the feedback information with the reasoning context and the output constraint instruction generated by the report, identifying new green attributes or policy update points, triggering the data source weight adjustment and verification rule update of the green financing project original data in the multi-source green data governance module, the business rule optimization of the green finance rule library and the parameter calibration of the environmental benefit measurement model library in the credible reasoning and calculation module, and based on the labeled data generated by the comparison and analysis, the pre-set large language model in the large model interaction module is incrementally trained. 10.The green finance-oriented credit model joint modeling method according to claim 9, characterized in that: Comprising the following steps: Step one: obtain green financing project original data, generate structured data through the green financing project original data, and generate knowledge constraint signals based on the pre-set green finance policy knowledge base; Step two: analyze the structured data to generate a logical decision chain, and execute tasks involving environmental benefit measurement in the structured data to generate quantitative indicators; Step three: integrate the quantitative indicators and the logical decision chain into a reasoning context, convert the knowledge constraint signals into output constraint instructions, and generate a credit evaluation report through a pre-set large language model according to the reasoning context and the output constraint instructions; Step four: receive feedback information of the credit evaluation report, and update and optimize the green financing project original data, rules and large language model in the multi-source green data governance module, the credible reasoning and calculation module and the large model interaction module.

Images