A green credit intelligent identification system, method, device and medium

Abstract

The application provides a green credit intelligent identification system, method, device and medium, effectively solves the defects of insufficient semantic understanding ability of unstructured documents, dynamic semantic retrieval of green industry directory knowledge, and lack of verifiable reasoning basis of matching results in the prior art. The system comprises: a multi-modal understanding module configured to call a corresponding large model based on the document type of an unstructured multi-modal document, extract key information of an enterprise through a prompt engineering, and verify the key information of the enterprise; an intelligent matching module configured to obtain a query vector by quantifying the verified key information of the enterprise, retrieve a retrieval result based on the query vector, splice the query vector to obtain an enhanced context, and process the enhanced context to obtain a preliminary identification result; and a collaborative review module configured to filter a target identification suggestion from the preliminary identification result, provide a visual interface for an approval personnel to perform review, and generate a formal identification report after the review is passed.

Description

Technical Field

[0001] This application relates to the field of credit assessment technology, and more specifically, to a green credit intelligent assessment system, method, device, and medium. Background Technology

[0002] Currently, in green credit operations, the determination of whether loan projects comply with the "Green Industry Guidance Catalogue" relies heavily on manual review. Approving personnel must meticulously review non-standardized documents submitted by companies, such as trade background contracts, VAT invoices, tender documents, and business licenses, and manually compare the technical parameters and explanations listed in the 9 primary, 38 secondary, and 271 tertiary categories of the catalogue. This model suffers from three inherent flaws: First, it has weak unstructured document processing capabilities; existing OCR tools can only recognize text positions and cannot understand the semantic relationship between phrases like "the subject matter of the contract is a waste heat boiler" and "industrial waste heat recovery and utilization under item 3.1.2 of the catalogue." Second, its knowledge representation methods are outdated; traditional rule engines solidify catalogue clauses into keyword matching logic, making it difficult to handle catalog updates, cross-level clause references, and ambiguous boundaries of technical parameters (such as "high-efficiency motor" without a defined energy efficiency level). Third, it lacks interpretable reasoning capabilities; the system can only output binary judgment results and cannot generate a chain of arguments citing original materials and catalogue clauses, nor can it propose technical improvement paths for mismatched projects. Although there have been attempts at simple semantic matching based on BERT, they are limited by single-modal text modeling and cannot handle scanned documents containing charts, seals, or handwritten annotations; while fine-tuning large models faces practical obstacles such as scarce labeled data, a black box in the inference process, and untraceable auditing.

[0003] Therefore, there is an urgent need for a new technical architecture that integrates multimodal perception, semantic knowledge base and verifiable reasoning to break through the technical bottleneck of intelligent identification of green credit. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a green credit intelligent identification system, method, device and medium, which effectively solves the problems of insufficient semantic understanding of unstructured documents, inability to dynamically semantically retrieve green industry catalog knowledge and lack of verifiable reasoning basis for matching results in the existing technology.

[0005] In a first aspect, embodiments of this application provide a smart green credit identification system, the system comprising: The multimodal understanding module is used to receive unstructured multimodal documents for credit, call the corresponding large model based on the document type of the unstructured multimodal document, and extract key enterprise information through prompts and engineering, and verify the key enterprise information; The intelligent matching module is used to quantify and verify key enterprise information to obtain a query vector. Based on the query vector, a search is performed in a green industry knowledge base pre-built based on the target credit file to obtain search results. The search results are then concatenated with the query vector to obtain an enhanced context. The enhanced context is then processed by a large language model guided by prompting engineering to obtain a preliminary identification result. The collaborative review module is used to filter target identification suggestions from the preliminary identification results, provide them to approvers through a visual interface, and generate an interpretable formal identification report for green credit after the review is approved.

[0006] In conjunction with the first aspect, this application provides a first possible implementation of the first aspect, wherein the intelligent matching module is used to process the enhanced context based on the prompting engineering-guided large language model to obtain a preliminary identification result, including: The enhanced context is semantically understood using a large language model to obtain the understood content, and the semantic matching degree between the understood content and each clause in the green industry knowledge base is calculated. Based on the semantic matching degree, target third-level directory names are filtered out to generate matching suggestions, and original text fragments that support the matching suggestions are located. The original text fragment is compared with the corresponding third-level directory clauses to generate matching criteria that include the cited original text and clause comparison explanations, which are then integrated with the matching suggestions to arrive at a preliminary determination result.

[0007] In conjunction with the first aspect, this application provides a second possible implementation of the first aspect, wherein, after generating the matching basis including the original references and the clause comparison description, it further includes: Extract the target third-level directory name and the project information used to describe the loan purpose from the enterprise's key information in the matching suggestions, respectively. The system automatically verifies the consistency between the target third-level directory name and the project information in the enterprise's key information, and integrates the results with the matching suggestions after the consistency is verified to generate a preliminary assessment.

[0008] In conjunction with the first aspect, this application provides a third possible implementation of the first aspect, wherein the intelligent matching module is used to retrieve search results based on the query vector in a green industry knowledge base pre-constructed based on the target credit document, including: A green industry knowledge base is pre-constructed based on the target credit documents, and the text content of the green industry knowledge base is transformed into a high-dimensional vector; The query vector is matched with the high-dimensional vector to obtain the matching results, and the search results are filtered out from the matching results.

[0009] In conjunction with the first aspect, this application provides a fourth possible implementation of the first aspect, wherein a green industry knowledge base is pre-constructed based on the target credit document, including: The target credit document is transformed into a structured knowledge representation, and multiple knowledge documents are extracted based on the knowledge representation; Semantic quantification is performed on the document content of the multiple knowledge documents to establish a green industry knowledge base with a target retrieval method after quantification.

[0010] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein the multimodal understanding module is used to call the corresponding large model based on the document type of the unstructured multimodal document and to prompt the engineering to extract key enterprise information, including: Based on the document type, semantic parsing is performed on the unstructured multimodal document to obtain text information; Based on the document type, select the corresponding prompt template from the prompt template library built through the prompt project based on the business characteristics identified by green credit; Input the text information and prompt template into the large language model to obtain key enterprise information related to green credit certification.

[0011] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein the multimodal understanding module is used to verify the enterprise key information, including: Extract financial data and entity identification information from key corporate information in multiple documents related to the same project or business; The financial data logic is validated, and the entity identification information is validated to verify the key information of the enterprise.

[0012] Secondly, embodiments of this application provide a smart identification method for green credit, characterized in that the method includes: The multimodal understanding module receives unstructured multimodal documents for credit purposes, calls the corresponding large model based on the document type of the unstructured multimodal document and prompts the engineering process to extract key enterprise information, and verifies the key enterprise information. The query vector is obtained based on the key enterprise information after quantitative verification by the intelligent matching module. The query vector is then used to search the green industry knowledge base pre-built based on the target credit file to obtain the search results. The search results are then concatenated with the query vector to obtain the enhanced context. The enhanced context is then processed by the large language model guided by the prompting engineering to obtain the preliminary identification result. The collaborative review module filters target identification suggestions from the preliminary identification results, provides them to approvers for review through a visual interface, and generates an interpretable formal identification report for green credit after the review is approved.

[0013] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the aforementioned green credit intelligent identification method are performed.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the aforementioned intelligent green credit identification method.

[0015] This application provides a green credit intelligent identification system, comprising: a multimodal understanding module, used to receive unstructured multimodal documents for credit, call the corresponding large model and prompt engineering based on the document type of the unstructured multimodal documents to extract key enterprise information, and verify the key enterprise information; an intelligent matching module, used to quantify the verified key enterprise information to obtain a query vector, search in a green industry knowledge base pre-constructed based on the target credit document based on the query vector to obtain search results, concatenate with the query vector to obtain enhanced context, and process the enhanced context based on prompt engineering guided by a large language model to obtain a preliminary identification result; and a collaborative review module, used to filter target identification suggestions from the preliminary identification results, provide a visual interface for review by approvers, and generate an interpretable formal green credit identification report after the review is passed. Based on the above system, this application effectively overcomes the technical defects of existing green credit identification, such as relying on manual verification of unstructured documents item by item, difficulty in understanding semantic relationships, inability to dynamically adapt and update the directory, and lack of interpretable reasoning process. Through the multimodal understanding module, adaptive parsing and semantic-level information extraction of heterogeneous documents such as contracts, invoices, and tender documents are achieved, significantly improving the accuracy and robustness of unstructured data processing. With the help of the semantic knowledge base and retrieval enhancement generation mechanism built by the intelligent matching module, fuzzy matching, hierarchical penetration, and potential correlation mining between project features and the green industry catalog are possible, breaking through the rigid limitations of traditional rule engines in semantic understanding and dynamic adaptation. The collaborative review module adopts a closed-loop design of "machine review and human verification", which not only ensures the authority and traceability of the assessment conclusions, but also strengthens the basis for human decision-making and business guidance capabilities through visual interaction and interpretable output, forming an efficient, reliable, and sustainably evolving intelligent assessment method for green credit. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This paper shows a structural block diagram of a green credit intelligent identification system provided in an embodiment of this application; Figure 2 This paper illustrates a system architecture diagram of a green credit intelligent identification system provided in an embodiment of this application. Figure 3 A flowchart illustrating the multimodal understanding module provided in an embodiment of this application is shown; Figure 4 A flowchart illustrating the intelligent matching module provided in an embodiment of this application is shown; Figure 5 A schematic diagram illustrating the process of constructing a green industry knowledge base provided in an embodiment of this application is shown; Figure 6 A flowchart illustrating the collaborative review module provided in an embodiment of this application is shown; Figure 7 A flowchart illustrating a smart identification method for green credit provided in an embodiment of this application is shown. Figure 8 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0019] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0021] Existing credit assessment technologies cannot effectively overcome the shortcomings of insufficient semantic understanding of unstructured documents, inability to dynamically retrieve knowledge from the green industry catalog, and lack of verifiable reasoning basis for matching results in green credit assessment scenarios. They also cannot automatically extract key enterprise information from multi-source heterogeneous documents, perform semantic-level accurate matching with the vectorized green catalog, and generate auditable assessment conclusions with original text citations.

[0022] Based on this, the embodiments of this application provide a green credit intelligent identification system, method, device and medium, which are described below through embodiments.

[0023] Example 1 To facilitate understanding of this embodiment, a detailed description of the intelligent green credit identification system disclosed in this application embodiment will be provided first. For example... Figure 1 The diagram shown is a structural block diagram of a green credit intelligent identification system, as follows: Figure 2 The diagram shows a system architecture of a smart green credit identification system. This system can be understood as a server component including a processor. The smart green credit identification system provided in this application includes: The multimodal understanding module 101 is used to receive unstructured multimodal documents for credit, call the corresponding large model based on the document type of the unstructured multimodal document and prompt the engineering to extract key enterprise information, and verify the key enterprise information; The intelligent matching module 102 is used to quantify and verify the key information of the enterprise to obtain a query vector, and to retrieve the query result in a green industry knowledge base pre-constructed based on the target credit file. The result is then concatenated with the query vector to obtain an enhanced context, and the enhanced context is processed by a large language model guided by prompting engineering to obtain a preliminary identification result. The collaborative review module 103 is used to filter target identification suggestions from the preliminary identification results, provide them to the approvers for review through a visual interface, and generate an interpretable formal identification report for green credit after the review is approved.

[0024] The multimodal understanding module 101, such as Figure 3 As shown, this system is used to receive unstructured multimodal documents for credit purposes, including but not limited to text documents (such as application forms, project feasibility reports, tender documents, and enterprise qualification certificates), scanned image documents (such as scanned copies of contracts and invoice images), and system screenshot documents (such as screenshots of query results from enterprise information systems). This application pre-sets corresponding large models based on the document type of the unstructured multimodal document. For example, for image documents, a multimodal large model such as QWenVL-max-72b is used; for text documents, a large language model such as DEEPSEEKv3 is used for semantic parsing. Based on the document type of the unstructured multimodal document (text, scanned image, and system screenshot), the corresponding large model is called, and the large model is guided by a prompting project for processing. This allows the large model to accurately identify and extract key enterprise information through carefully designed prompt templates. The prompt templates in the prompting project are designed with the business characteristics of green credit recognition in mind and set differentiated extraction strategies for different types of documents. Simultaneously, an information verification mechanism is established to verify the consistency of the extracted results, such as the correspondence between invoice amounts and contract amounts, and the consistency of company names across various documents. This ensures the accuracy of the extracted key company information, ultimately enabling the automatic identification and extraction of key company information related to green credit certification from various documents. The extracted key company information includes, but is not limited to: basic project information: project name, project type, loan purpose, etc.; company qualification information: company name, business scope, etc.; trade background information: contract amount, contract subject matter, invoice information, etc.; and key company information in bidding documents: bidding content, technical requirements, evaluation criteria, etc.

[0025] For unstructured multimodal documents related to credit, there are two scenarios: one is the initial green credit identification scenario at the branch level, in which the branch account manager needs to prepare relevant green loan identification materials for the enterprise, and the other is the intelligent identification of green labels by automatically identifying multiple types of approval materials through multimodal big data model technology, combined with vector knowledge base and enhanced retrieval generation technology; the other is the green credit review scenario at the branch level, in which the branch credit reviewer uses multimodal big data model technology to automatically identify multiple types of approval materials, and at the same time combines the information declared by the account manager, and uses vector knowledge base and reasoning big data model to realize intelligent reprinting and review of green labels.

[0026] In some embodiments, the multimodal understanding module is used to call the corresponding large model based on the document type of the unstructured multimodal document and to prompt the engineering team to extract key enterprise information, including: Based on the document type, semantic parsing is performed on the unstructured multimodal document to obtain text information; Based on the document type, select the corresponding prompt template from the prompt template library built through the prompt project based on the business characteristics identified by green credit; Input the text information and prompt template into the large language model to obtain key enterprise information related to green credit certification.

[0027] In this embodiment, the multimodal understanding module performs differentiated preprocessing on the input unstructured multimodal documents based on document type: for scanned image documents (such as environmental impact assessment approvals and energy-saving certification certificates), QWenVL-max-72b is used to extract the original text and spatial layout information from the images, understand and extract information related to the enterprise's loan purpose; for text documents (such as project feasibility study reports), a large language model is called to identify chapter structure, table nesting, and hyperlink context, and to identify project content, loan description, and other information. Subsequently, the system dynamically selects a suitable prompt template from a pre-built "green credit prompt template library," wherein the standard fields of the prompt template include, but are not limited to, basic project information, enterprise qualification information, trade background information, and key information of the bidding enterprise. Finally, the extracted text information and the selected prompt template are input into the large language model (supporting 128K context). Under the guidance of the prompts, the large language model identifies elements strongly related to green certification, including the green industry category to which the project belongs, the description of the advancement of the technical route, the quantitative indicators of environmental benefits (such as annual carbon reduction and water saving), the status and validity period of third-party certification, etc. Among them, strong correlation can be elements with a correlation greater than a preset threshold. The identified elements are processed with standardized field names (such as GREEN_CATEGORY, ENV_BENEFIT_VALUE) to output structured key information of enterprises with green credit certification.

[0028] In some embodiments, the multimodal understanding module is used to verify the enterprise's key information, including: Extract financial data and entity identification information from key corporate information in multiple documents related to the same project or business; The financial data logic is validated, and the entity identification information is validated to verify the key information of the enterprise.

[0029] In this embodiment, the multimodal understanding module further constructs a cross-document collaborative verification mechanism to improve the credibility and business adaptability of key corporate information in green credit. After completing the semantic parsing of a single document, the multimodal understanding module automatically identifies multiple document sets under the same credit project or business-related scenario, such as project feasibility study reports, carbon emission verification reports, and green procurement contracts. The multimodal understanding module extracts two types of core verification objects in a structured manner: first, financial-related data, including but not limited to indicators such as green project investment amount and loan amount; second, entity identification information, covering strong identity anchors such as unified social credit code and project implementation entity name (and other types of data). For financial data, it verifies whether the loan amounts disclosed by the same entity in different documents match. For entity identification information, it initiates entity consistency verification: after standardized cleaning (removing spaces, converting full-width and half-width characters, and unified encoding mapping), it compares unique fields such as unified social credit code among multiple documents; for name fields (such as "XX New Energy Technology Co., Ltd." and "XX New Energy Technology Co., Ltd."), it uses semantic similarity to determine whether they point to the same legal entity. All verification results are labeled with the problem type, the source of the relevant documents, and the confidence level, and a traceable verification log is generated. Only when all verification items pass will the multimodal understanding module mark the group of enterprise key information as "trustworthy and usable" and proceed to the subsequent green credit rating and credit decision-making process.

[0030] The intelligent matching module 102, such as Figure 4As shown, the bge-zh-based model quantifies the received, verified key enterprise information to obtain a highly semantically accurate query vector. Based on this query vector, a search is performed in a pre-constructed green industry knowledge base based on a target credit document to obtain retrieval results including Top-K related knowledge fragments. The target credit document is the "Green Industry Guidance Catalogue (2025 Edition)". The constructed green industry knowledge base is organized using a hierarchical index structure, supporting multi-dimensional recall based on fine-grained domains, national economic industry categories, conditions / standards, etc. The retrieval results are concatenated with the query vector to obtain an enhanced context. This enhanced context is injected with a customized green prompt template. The green prompt template explicitly constrains the large language model to classify and identify applicable clauses based on current regulatory definitions, avoiding subjective inference. Large language models such as DEEPSEEKv3 perform green credit assessment based on this enhanced context and output preliminary assessment results. These preliminary assessment results include the green industry category, matching criteria, points of doubt (e.g., "The project must strictly comply with the ecological protection red line requirements, ensure the construction plan passes environmental impact assessment review, and avoid environmental disputes"), and a list of recommended supplementary materials. The contents of the preliminary determination results can all be traced back to specific entries in the knowledge base and original document evidence, ensuring that the determination process is transparent, auditable, and compliant with regulatory requirements.

[0031] In some embodiments, the intelligent matching module is used to process the enhanced context based on the prompting engineering-guided large language model to obtain a preliminary identification result, including: The enhanced context is semantically understood using a large language model to obtain the understood content, and the semantic matching degree between the understood content and each clause in the green industry knowledge base is calculated. Based on the semantic matching degree, target third-level directory names are filtered out to generate matching suggestions, and original text fragments that support the matching suggestions are located. The original text fragment is compared with the corresponding third-level directory clauses to generate matching criteria that include the cited original text and clause comparison explanations, which are then integrated with the matching suggestions to arrive at a preliminary determination result.

[0032] In this embodiment, the intelligent matching module further introduces a fine-grained semantic alignment mechanism to improve the accuracy and interpretability of green industry identification. First, it drives a large language model to perform deep semantic understanding of the aforementioned enhanced context (including verified key enterprise information and retrieved knowledge fragments), generating structured understanding content. This structured understanding content not only covers key information such as loan purpose but also automatically removes redundant descriptions and vague expressions. Subsequently, this understanding content is compared with all third-level category clauses in the green industry knowledge base, such as "1.2.1 New Energy Vehicle and Key Component Manufacturing" and "3.1.4 Agricultural and Forestry Waste Comprehensive Utilization Equipment Manufacturing," as well as the corresponding conditions / standards for each third-level category, using fine-grained semantic matching degree calculation. The resulting matching degrees are then sorted, and based on this sorting, the target third-level category name with the highest score and meeting the threshold is selected to form preliminary matching suggestions. The intelligent understanding module also simultaneously backtracks the original document, accurately locating key original text fragments supporting the recommendation (such as "This project is a rooftop distributed photovoltaic system, adopting a 'full grid connection' model, with an installed capacity of 2.3MW" in the feasibility study report), and compares this fragment sentence by sentence with the original text of the selected third-level clauses to generate a clause compliance statement. Finally, the matching recommendation, the located original text, and the clause comparison statement are organically integrated to form a preliminary determination result with a complete chain of evidence, available for manual review.

[0033] In some embodiments, after generating the matching criteria that includes the original references and clause comparison descriptions, the method further includes: Extract the target third-level directory name and the project information used to describe the loan purpose from the enterprise's key information in the matching suggestions, respectively. The system automatically verifies the consistency between the target third-level directory name and the project information in the enterprise's key information, and integrates the results with the matching suggestions after the consistency is verified to generate a preliminary assessment.

[0034] In this embodiment, the application also sets up a verification submodule in the intelligent matching module as a key verification step before the preliminary identification result is generated, to ensure that the green industry attribution strictly corresponds to the actual credit use. This module first accurately extracts the target third-level directory name from the matching suggestions, such as "1.3.2 Motor System Energy Efficiency Improvement", and simultaneously locates the structured project information field used to describe the loan purpose from the enterprise's key information, including core semantic elements such as project name, construction content, technical solution, service object, operation mode, and fund usage details. Then, it starts bidirectional semantic alignment verification: on the one hand, it sets the technical scope, applicable scenarios, and restrictions defined by the third-level directory, such as "Building Renewable Energy Application", which clearly requires the following building renewable energy application facility construction activities in energy-saving engineering construction activities, including but not limited to: 1. Activities of installing solar photovoltaic power generation devices on building roofs and walls; 2. Installation and construction activities of building cooling, heating, and hot water supply facilities such as heat pumps and solar water heating devices; on the other hand, it analyzes whether each element in the project information fully supports the benchmark at the semantic level, for example For example, if the project information is described as "building rooftop and wall-mounted solar power generation facilities," then the conditions / standards in the catalog terms are fully covered. The verification process uses a rule-guided semantic parser, combined with a green finance terminology dictionary and a negative identification mechanism, such as automatically filtering text containing exclusionary expressions like "for internal employee use only" or "excluding data platform integration," to avoid mechanical keyword matching. Consistency is considered passed only when all necessary semantic dimensions achieve positive coverage and there are no substantial conflicts. After passing the consistency verification, the system integrates the target third-level catalog name, the verified project information summary, and the aforementioned matching basis explanation to generate a logically consistent, purpose-defined, and traceable preliminary identification result, ensuring that every green credit fund is invested in a truly accurate and compliant manner in accordance with the green industry policy guidance of official institutions.

[0035] In some embodiments, the intelligent matching module is used to retrieve search results based on the query vector in a green industry knowledge base pre-constructed based on the target credit document, including: A green industry knowledge base is pre-constructed based on the target credit documents, and the text content of the green industry knowledge base is transformed into a high-dimensional vector; The query vector is matched with the high-dimensional vector to obtain the matching results, and the search results are filtered out from the matching results.

[0036] In this embodiment, each clause text (including first-level directory, second-level directory, field, national economic industry name, conditions / standards, and remarks) in the green industry knowledge base pre-built in this application is processed by a semantic encoder adapted to the green finance field and transformed into a high-dimensional dense vector with fixed dimensions and business discrimination capabilities. The vectorization process fully preserves the policy intent, technical boundaries, and logical constraints of the clauses, avoiding semantic drift caused by general language models. All vectors are uniformly stored in a vector database that supports efficient nearest neighbor retrieval, and a multi-level index is established, organized according to the directory hierarchy (e.g., first level "energy saving and carbon reduction" → second level "high-efficiency energy-saving equipment manufacturing" → third level "energy-saving agricultural input manufacturing"). After the intelligent matching module generates the query vector, it calls the vector retrieval engine to complete the full database similarity calculation and returns the Top-K semantically most relevant clause vectors and their original texts. The retrieval results are strictly selected according to two standards: first, the semantic matching degree must exceed the preset business threshold to ensure that the recalled results have substantial reference value; second, the results must cover different granular levels (prioritizing accurate hits in the third-level directory, while retaining the context of the upper-level directory to support the classification logic deduction). The final output retrieval results not only include the original text of the clause, but also include the confidence score and the complete path of the clause in the knowledge base, providing authoritative, clearly structured, and verifiable knowledge support for subsequent large language model reasoning guided by the prompting project, fundamentally ensuring the policy compliance and technical rigor of the green certification conclusions.

[0037] In some embodiments, a green industry knowledge base is pre-built based on the target credit document, including: The target credit document is transformed into a structured knowledge representation, and multiple knowledge documents are extracted based on the knowledge representation; Semantic quantification is performed on the document content of the multiple knowledge documents to establish a green industry knowledge base with a target retrieval method after quantification.

[0038] In this embodiment, such as Figure 5As shown, this application constructs a green industry knowledge base based on the "Green Industry Guidance Catalogue (2025 Edition)" as the authoritative standard. This catalogue document contains issues such as tables spanning multiple pages, merged cells, and cells spanning multiple pages. Direct use may lead to problems such as irregular truncation, semantic ambiguity, and delayed updates. Therefore, format conversion and data cleaning are performed first to remove redundant information and format noise, forming a standardized catalogue. A rule-guided document parser is also used to identify the title level, second-level directory, field, national economic industry name, conditions / standards, and remarks of the "Green Industry Guidance Catalogue (2025 Edition)". This generates multiple semantically independent and topic-focused knowledge documents, such as "energy conservation and carbon reduction industry," "green services," and "green trade." Subsequently, semantic quantification is performed on each knowledge document based on the vector model bge-zh. Finally, all quantified knowledge documents are injected into a vector database and bound to target retrieval methods: supporting precise retrieval by third-level directory number and responding to natural language queries (such as "Which clauses apply to the upgrading and renovation of county-level sewage treatment plants?").

[0039] The established green industry knowledge base also has verification and policy tracking mechanisms to regularly check the consistency and timeliness of the knowledge. When the green industry guidance catalog is updated, the knowledge base is maintained through incremental updates to ensure the continued effectiveness of the knowledge.

[0040] The collaborative review module 103, such as Figure 6As shown, the collaborative workflow design employs both machine-based review and human verification, fully leveraging the efficiency advantages of intelligent systems and the flexibility of manual review. First, it automatically filters results with low confidence from the preliminary assessment results, displaying the top three preliminary assessment suggestions with high relevance or matching degree. All target suggestions are uniformly pushed to a dedicated visual review interface, which adopts a three-column layout: the left column displays key pages of the original credit document (supporting zooming and text overlay recognition); the middle column presents the preliminary assessment logic chain in a structured manner (including green directory references, original text location anchors, and consistency verification markers); and the right column provides interactive review tools, including checkboxes for approval / rejection, drop-down selection of supplementary evidence types (such as "supplementary environmental impact assessment approval" and "upload technical contract attachments"), and an embedded annotation box for approvers to fill in professional opinions. During the review process, the collaborative review module records operation behavior, modification traces, and time nodes in real time to ensure compliance of the entire audit process. After the review is approved, the collaborative review module automatically calls the report generation engine, integrates original material evidence, clause comparison tables, manual confirmation signatures, and standardized expressions of regulatory terminology, and outputs the required assessment report. The assessment report includes basic project information, assessment conclusions, matching evidence, approval opinions, etc. For projects that do not meet the green catalog standards, the report also provides suggestions for green enterprise development, guiding companies in their improvement efforts. The report adopts a structure of "conclusions first, evidence layered, and original text readily available," with each assessment conclusion accompanied by a hyperlink to the relevant original text and knowledge base clauses. This fully meets the interpretability requirements of regulatory inspections, internal and external audits, and client disclosures.

[0041] Compared with the prior art, this application has the following significant advantages: (1) Significantly improved document processing capabilities: It enables unified processing of various unstructured data, accurately understands the semantic structure of documents, and automatically extracts key information.

[0042] (2) Significantly improved accuracy in directory matching: This application constructs a semantic green industry knowledge base and uses RAG technology to achieve intelligent matching, which can handle complex situations such as semantic similarity matching and fuzzy matching between project features and directory standards, thus significantly improving matching accuracy. It is particularly worth mentioning that this invention can uncover green directory dimensions that are not recognized by humans and discover potential associations between projects and directories, which is something that existing technologies cannot achieve.

[0043] (3) Significantly improved review efficiency: Manual review has a long cycle and high labor costs, which seriously restricts the scale of business. This application realizes the collaborative process of machine review and human review, which significantly improves review efficiency.

[0044] (4) Significantly enhanced business guidance capabilities: Existing technologies can only provide simple rejection conclusions for projects that do not meet the green catalog standards, without offering valuable suggestions. This application can provide green enterprise development suggestions for non-catalog client projects, guiding enterprises to transform towards green development. This function effectively enhances the business guidance capabilities of account managers.

[0045] (5) Strong system scalability: Existing rule systems have high maintenance costs and are difficult to cope with policy changes and business expansion. The knowledge base-driven architecture adopted in this application supports incremental updates and continuous optimization of knowledge. When the green industry guidance catalog is updated, only the knowledge base content needs to be updated, and the system can adapt to the new standard requirements, demonstrating good scalability and adaptability.

[0046] This application uses the domestic Tongyi Qianwen QWen large model. Other large models are also applicable to this green credit intelligent identification system, such as Zhipu GLM, Wenxin Yiyan, and Pangu VLM.

[0047] Furthermore, the intelligent matching module, if not using retrieval enhancement generation technology, can also employ a fine-tuning approach with a large language model. This involves using historical data related to green credit identification to fine-tune the large language model, enabling it to directly possess matching judgment capabilities. However, this requires a large amount of labeled data, resulting in high model training and maintenance costs, and making it difficult to interpret the model's judgment logic.

[0048] Example 2 This application also provides a smart method for identifying green credit, such as Figure 7 The diagram illustrates a flowchart of a smart green credit identification method. The steps of executing this method on a terminal device correspond to the effects of the aforementioned smart green credit identification system. The smart green credit identification method described in this application includes: S701. Receive unstructured multimodal documents for credit through the multimodal understanding module, call the corresponding large model and prompt the engineering process based on the document type of the unstructured multimodal document to extract the key information of the enterprise, and verify the key information of the enterprise. S702. Obtain a query vector based on the key enterprise information after quantitative verification by the intelligent matching module. Based on the query vector, search the green industry knowledge base pre-built based on the target credit file to obtain the search results. Concatenate the query vector with the enhanced context to obtain the enhanced context. Process the enhanced context based on the prompting engineering guided large language model to obtain the preliminary identification result. S703, the control collaboration review module selects target identification suggestions from the preliminary identification results, provides them to the approvers for review through a visual interface, and generates an interpretable formal identification report for green credit after the review is approved.

[0049] Example 3 This application also provides an electronic device, such as Figure 8 As shown, it includes: a processor 801, a memory 802, and a bus 803. The memory 802 stores machine-readable instructions that can be executed by the processor 801. When the electronic device is running, the processor 801 and the memory 802 communicate through the bus 803. When the machine-readable instructions are executed by the processor 801, the steps of the green credit intelligent identification method are performed.

[0050] Example 4 This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the aforementioned smart identification method for green credit.

[0051] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0052] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0053] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0054] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0055] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A green credit intelligent identification system, characterized in that, The system includes: The multimodal understanding module is used to receive unstructured multimodal documents for credit, call the corresponding large model based on the document type of the unstructured multimodal document, and extract key enterprise information through prompts and engineering, and verify the key enterprise information; The intelligent matching module is used to quantify and verify key enterprise information to obtain a query vector. Based on the query vector, a search is performed in a green industry knowledge base pre-built based on the target credit file to obtain search results. The search results are then concatenated with the query vector to obtain an enhanced context. The enhanced context is then processed by a large language model guided by prompting engineering to obtain a preliminary identification result. The collaborative review module is used to filter target identification suggestions from the preliminary identification results, provide them to approvers through a visual interface, and generate an interpretable formal identification report for green credit after the review is approved.

2. The system according to claim 1, characterized in that, The intelligent matching module is used to process the enhanced context based on the prompting engineering-guided large language model to obtain preliminary identification results, including: The enhanced context is semantically understood using a large language model to obtain the understood content, and the semantic matching degree between the understood content and each clause in the green industry knowledge base is calculated. Based on the semantic matching degree, target third-level directory names are filtered out to generate matching suggestions, and original text fragments that support the matching suggestions are located. The original text fragment is compared with the corresponding third-level directory clauses to generate matching criteria that include the cited original text and clause comparison explanations, which are then integrated with the matching suggestions to arrive at a preliminary determination result.

3. The system according to claim 2, characterized in that, After generating the matching criteria, which includes the original cited text and a comparison description of the clauses, it also includes: Extract the target third-level directory name and the project information used to describe the loan purpose from the enterprise's key information in the matching suggestions, respectively. The system automatically verifies the consistency between the target third-level directory name and the project information in the enterprise's key information, and integrates the results with the matching suggestions after the consistency is verified to generate a preliminary assessment.

4. The system according to claim 1, characterized in that, The intelligent matching module is used to retrieve search results based on the query vector in a green industry knowledge base pre-constructed based on the target credit document, including: A green industry knowledge base is pre-constructed based on the target credit documents, and the text content of the green industry knowledge base is transformed into a high-dimensional vector; The query vector is matched with the high-dimensional vector to obtain the matching results, and the search results are filtered out from the matching results.

5. The system according to claim 4, characterized in that, A green industry knowledge base is pre-built based on the target credit documents, including: The target credit document is transformed into a structured knowledge representation, and multiple knowledge documents are extracted based on the knowledge representation; Semantic quantification is performed on the document content of the multiple knowledge documents to establish a green industry knowledge base with a target retrieval method after quantification.

6. The system according to claim 1, characterized in that, The multimodal understanding module is used to call the corresponding large model based on the document type of unstructured multimodal documents and to prompt the engineering team to extract key enterprise information, including: Based on the document type, semantic parsing is performed on the unstructured multimodal document to obtain text information; Based on the document type, select the corresponding prompt template from the prompt template library built through the prompt project based on the business characteristics identified by green credit; Input the text information and prompt template into the large language model to obtain key enterprise information related to green credit certification.

7. The system according to claim 1, characterized in that, The multimodal understanding module is used to verify the key enterprise information, including: Extract financial data and entity identification information from key corporate information in multiple documents related to the same project or business; The financial data logic is validated, and the entity identification information is validated to verify the key information of the enterprise.

8. A smart identification method for green credit, characterized in that, The method includes: The multimodal understanding module receives unstructured multimodal documents for credit purposes, calls the corresponding large model based on the document type of the unstructured multimodal document and prompts the engineering process to extract key enterprise information, and verifies the key enterprise information. The query vector is obtained based on the key enterprise information after quantitative verification by the intelligent matching module. The query vector is then used to search the green industry knowledge base pre-built based on the target credit file to obtain the search results. The search results are then concatenated with the query vector to obtain the enhanced context. The enhanced context is then processed by the large language model guided by the prompting engineering to obtain the preliminary identification result. The collaborative review module filters target identification suggestions from the preliminary identification results, provides them to approvers through a visual interface, and generates an interpretable formal identification report for green credit after the review is approved.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the green credit intelligent identification method as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the green credit smart identification method as described in claim 8.

Images