Large-model-based loan commission intelligent discrimination system and large-model-based loan commission intelligent discrimination method

Abstract

The invention discloses a loan direction intelligent discrimination system and method based on a large model, and belongs to the technical field of financial credit loan, and the system comprises a multi-source data collection module, a data preprocessing module, a policy knowledge base module, a large model analysis module, a rule engine module, a result fusion and decision-making module, a manual review and feedback module, a knowledge base and model updating module, and a user interaction module. The method comprises the steps of preprocessing after multi-source data collection, policy knowledge loading, big model preliminary judgment, rule engine verification, manual review and feedback after decision fusion, knowledge base and model updating and final output of a display result. The loan investment classification accuracy, efficiency and policy adaptability are improved, the operation complexity is reduced, continuous optimization is supported, and the method is suitable for automatic loan investment classification scenes in the financial field.

Description

Technical Field

[0001] This invention relates to a loan system and method, and more specifically, to a loan allocation intelligent identification system and method based on a large model, belonging to the field of financial credit technology. Background Technology

[0002] Currently, financial institutions primarily rely on manual processes to determine whether loan allocations fall under the "Five Major Articles" (a set of five key areas of financial services). Specific methods include: 1) Manual policy comparison: Account managers and approvers manually review annual reports, business registration information, and project materials submitted by companies, comparing them with relevant policy documents for the "Five Major Articles" (such as R&D investment standards for science and technology finance and environmental certification requirements for green finance) to subjectively determine the loan allocation category; 2) Keyword matching: Keywords related to the "Five Major Articles" (such as "carbon neutrality" and "micro and small enterprises") are manually selected from the data, and categorized based on experience; 3) Indicator verification: Financial data of companies (such as the proportion of R&D expenses and the amount of environmental protection expenditures) are manually extracted and compared with quantitative indicators in policies (such as R&D investment ≥ 5%) to determine the category.

[0003] The existing technology has the following drawbacks: 1) High subjectivity: It relies on the experience of personnel, and different approvers have different understandings of the policies, which can easily lead to inconsistent classification results for similar projects; 2) Low efficiency: Manually processing multi-source texts (such as project feasibility reports of dozens of pages) and complex policy provisions can take several hours to classify a single loan; 3) High risk of misjudgment: Policies are updated frequently (such as the addition of "carbon footprint accounting" requirements to green finance), and it is difficult for humans to grasp all the details in real time, which can easily lead to misjudgment due to information omissions; 4) Poor scalability: When the policy standards or classification dimensions of the "five major articles" are adjusted, the manual process requires retraining of personnel, and the adaptation period is long. Summary of the Invention

[0004] The purpose of this invention is to provide a loan allocation intelligent identification system and method based on a large model, which improves the accuracy, processing efficiency and policy adaptability of loan allocation classification, reduces the complexity of manual operation, and enables continuous system optimization through a feedback mechanism to meet the compliance and efficiency requirements of financial institutions for loan classification in the "five major areas".

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution:

[0006] This invention provides a loan allocation intelligent discrimination system based on a large model, comprising:

[0007] The multi-source data acquisition module can connect with the bank's credit system, enterprise credit platform, business registration database, and patent database to collect structured and unstructured data of enterprises. It supports both batch import and real-time capture modes. The structured data includes registered capital, R&D investment ratio, and environmental certification number. The unstructured data includes annual reports, project feasibility reports, and environmental impact assessment documents.

[0008] The data preprocessing module, connected to the multi-source data acquisition module, is used to clean and standardize the acquired data, verify the integrity of fields and unify the format of structured data, and use OCR to recognize image-type text and use word segmentation tools to split text and extract entity information for unstructured data. The processing flow is as follows: raw data input, data cleaning and standardization, text extraction, recognition and word segmentation, entity recognition, and generation of structured indicators and unstructured topic words.

[0009] The policy knowledge base module includes a policy rule library, a quantitative indicator library, a keyword library, and a case library, which are used to store policy documents related to the "five major articles", hard indicators in various fields, core terms, and historical classified cases.

[0010] The large model analysis module is connected to the data preprocessing module and the policy knowledge base module. It adopts a large model finely tuned in the financial field (a model based on the Transformer architecture, such as the GPT series model), receives preprocessed enterprise data, extracts the core investment direction of the project through semantic understanding, and matches it with the keyword library and case library in the policy knowledge base module to output preliminary classification results and confidence scores.

[0011] The rules engine module, connected to the large model analysis module and the policy knowledge base module, is used to call the quantitative indicator library to perform hard indicator verification on the preliminary classification results output by the large model. If the indicators meet the standards, the classification results are strengthened; if they do not meet the standards, they are marked as "to be reviewed" and the non-compliant items are output.

[0012] The results fusion and decision module is connected to the large model analysis module and the rule engine module. It is used to combine the outputs of the two to generate the final judgment result. When the confidence of the large model is ≥80% and the rule engine verification is passed, the classification result and basis are directly output. When the confidence is <80% or the rule engine verification is not passed, the manual review process is triggered and the doubt analysis is output.

[0013] The manual review and feedback module is connected to the result fusion and decision-making module. It is used to show project data, preliminary model results and doubts to experts, receive final classification and annotations from experts, and automatically synchronize the expert annotation results to the case library and large model training set of the policy knowledge base module.

[0014] The knowledge base and model update module is connected to the policy knowledge base module and the large model analysis module. It can regularly crawl policy updates from the official websites of the People's Bank of China and the National Development and Reform Commission through policy crawlers, automatically parse new clauses and update them to the policy clause library and quantitative indicator library. Furthermore, it can fine-tune and train the large model monthly based on new cases and manual feedback data.

[0015] The user interaction module, connected to the result fusion and decision-making module and the manual review and feedback module, provides a visual interface to show account managers the final classification results, decision basis, and risk warnings. It also supports experts to manually modify the classification results and enter the reasons for the modification.

[0016] Preferably, the data preprocessing module automatically marks and completes the missing "environmental protection expenditure" in structured data, and uniformly converts "R&D investment ratio" into percentage form; for unstructured data, it uses the jieba word segmentation tool to split the text, and the extracted entity information includes "patent number" and "photovoltaic power station project name".

[0017] Preferably, the policy rule base in the policy knowledge base module stores policy documents related to the "five major areas" in the categories of science and technology finance, green finance, inclusive finance, pension finance, and digital finance.

[0018] The quantitative indicator library records: 1) Technology finance requirements: cumulative R&D investment ≥ 10 million yuan in the past 3 years; 2) Inclusive finance requirements: hard indicators for each field for companies with ≤ 300 employees; The keyword library is labeled by field with core terms such as "carbon emission reduction" and "wind power projects" for green finance, and "long-term care insurance" and "senior apartments" for elderly care finance; The case library stores historical categorized cases, including cases of new energy companies being judged as green finance due to photovoltaic projects and ISO14001 certification.

[0019] Preferably, the preliminary classification results output by the large model analysis module include "financial technology", "green finance", "inclusive finance", "pension finance" and "digital finance", with a confidence level range of 0-100%, which are used to assist the result fusion and decision-making module in determining whether to trigger the manual review process.

[0020] Preferably, when the rule engine module performs hard indicator verification on the preliminary classification results, if the large model determines it to be "technology finance", it will automatically verify the indicators of "R&D investment ratio ≥ 5%" and "number of patents ≥ 3"; if the large model determines it to be "digital finance", it will automatically verify the indicator of "digital investment ratio ≥ 8%".

[0021] Preferably, the manual review and feedback module synchronizes the expert annotation results to the case library and the large model training set, including the final category and reason for the expert judgment, which is used to optimize the verification criteria of the model parameters and rule engine.

[0022] Preferably, the knowledge base and model update module crawls policy updates once a week and performs fine-tuning training on the large model once a month, updating the semantic matching parameters and classification weights of the model.

[0023] Preferably, the risk warnings displayed to the account manager by the user interaction module include information about the company's environmental penalty records, and the output classification criteria include the matching policy clause number, keywords, and specific values ​​of quantitative indicators.

[0024] Preferably, when the multi-source data acquisition module collects data, it automatically sends a prompt to the account manager to complete the data for companies that have not provided an environmental impact assessment report, ensuring that unstructured data covers environmental protection-related information of the project.

[0025] This invention provides a method for intelligently identifying loan allocation based on a large model, comprising the following steps:

[0026] Step 101: Multi-source data collection. Automatically capture structured and unstructured data such as business registration information, financial statements, project contracts, and patent certificates of enterprises through interfaces. Automatically send a prompt to the account manager to complete the data if an environmental impact assessment report is not provided.

[0027] Step 102: Data preprocessing. The structured data is formatted and standardized, converting "R&D expenses of 5 million" into "R&D investment ratio of 5.2%". Unstructured data is extracted and entity recognition is performed. "Core business is artificial intelligence chip R&D" is extracted from the annual report, and the technology field to which the patent corresponding to the "patent number" belongs is marked. The processed enterprise feature data is output. The processing flow is as follows: raw data input, data cleaning and standardization, text extraction, recognition and word segmentation, entity recognition, generation of structured indicators and unstructured keywords.

[0028] Step 103: Load policy knowledge. Load the latest policy provisions, quantitative indicators, keywords and historical cases of the "Five Major Articles" from the policy knowledge base. If a policy update is detected that green finance has added a new classification standard for "hydrogen energy projects", the loaded content will be updated automatically.

[0029] Step 104: Initial identification by the large model. Input the enterprise feature data into the fine-tuned large model, identify the core investment direction of the project through semantic matching, compare it with the keyword library and case library of the policy knowledge base, and output the preliminary classification results and the percentage of confidence. The preliminary classification results include "digital finance".

[0030] Step 105: Rule engine verification. Based on the preliminary classification results, the corresponding quantitative indicator library is called for verification. "Digital finance" must meet the requirement of "digital investment ratio ≥ 8%". If all indicators meet the requirements, proceed to step 106. If there are any non-compliant items, mark them as "pending review" and record the reasons for non-compliance.

[0031] Step 106: Results Integration and Decision Making. When the confidence level of the preliminary classification result is ≥80% and the rule engine verification is passed, the final classification result is generated. The final classification result includes "inclusive finance", the policy clauses of the "Management Measures for Special Funds for the Development of Inclusive Finance", the keyword "rural cooperative", and the quantitative indicator of "50 employees of the enterprise". When the confidence level is <80% or the verification fails, the manual review process is triggered.

[0032] Step 107: Manual review and feedback. Experts review the project data, preliminary model results and doubts, manually determine the final category and enter the reasons. The expert's determination results and reasons are synchronized to the case library and used as training data.

[0033] Step 108: Knowledge base and model update. The official policy website is automatically crawled weekly to parse new policy content and update the knowledge base. New cases and human feedback data are used monthly to incrementally train the large model and update the model parameters.

[0034] Step 109: Results Output and Display. The final classification results, decision basis, and risk warnings regarding the company's environmental penalty records are displayed to the account manager through the user interface. The results of the "Loan Investment Classification Report" can be exported.

[0035] Beneficial Effects: Improved Classification Accuracy: Through deep semantic understanding of multi-source texts using a large model, combined with a dynamically updated policy knowledge base, biases caused by subjective human judgment are eliminated, increasing classification accuracy to over 95%; Increased Processing Efficiency: Automated text parsing, policy matching, and indicator verification reduce the time required for single loan classification to minutes, improving efficiency by 80%; Enhanced Policy Adaptability: Real-time synchronization of the latest policy provisions avoids human oversights and ensures classification results meet regulatory requirements; Reduced Operational Complexity: No need for manual memorization of massive policy details; the system automatically outputs classification criteria (such as matching policy provisions, keywords, and quantitative indicators) to assist in approval decisions; Support for Continuous Optimization: Through a manual review and feedback mechanism, labeled data is fed back to the model training set, enabling monthly adjustments and optimizations of the large model to continuously improve discrimination accuracy and adapt to changes in policies and business scenarios. Attached Figure Description

[0036] Figure 1 This is the system architecture diagram of the present invention.

[0037] Figure 2This is a flowchart of the data preprocessing process of the present invention.

[0038] Figure 3 This is a structural diagram of the policy knowledge base of this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Technical Solution / Principle: This invention proposes a large-scale model-based intelligent loan allocation identification system and method, applied to the automatic classification of loan allocation in the "five major areas" of finance. The system includes modules for multi-source data collection, data preprocessing, policy knowledge base, large-scale model analysis, rule engine, result fusion and decision-making, manual review and feedback, knowledge base and model updates, and user interaction. The method involves preprocessing multi-source data, loading policy knowledge, initial identification by the large-scale model, verification by the rule engine, fusion decision-making, manual review and feedback, updating the knowledge base and model, and finally outputting the displayed results. Through multi-module collaboration, the entire process of loan allocation, from data collection to classification decision-making and continuous optimization, is automated and intelligent.

[0041] The system first acquires structured and unstructured enterprise data using a multi-source data acquisition module. This data is then cleaned, standardized, and entity information is extracted by a data preprocessing module. The policy knowledge base module provides support with relevant policies, indicators, keywords, and case studies related to the "Five Major Articles." The large-scale model analysis module fine-tunes the large-scale model based on the financial sector, performs semantic understanding on the preprocessed data, matches policy knowledge base keywords with case studies, and outputs preliminary classifications and confidence levels. The rules engine module calls upon quantitative indicators from the policy knowledge base to verify the compliance of the large-scale model's preliminary results. The results fusion and decision-making module integrates the results from both modules to determine the final classification or trigger manual review. The manual review and feedback module processes items awaiting review, and the feedback results are used to optimize the policy knowledge base and the large-scale model. The knowledge base and model update module updates policies through policy crawlers and fine-tunes the large-scale model using new case studies and manual feedback. The user interaction module provides a visual interface to display the results and supporting evidence. Overall, by combining the large-scale model's semantic understanding capabilities with the rules engine's indicator verification, the system achieves intelligent loan allocation identification, while continuously optimizing system performance through manual feedback and automatic update mechanisms.

[0042] like Figure 1The system architecture diagram illustrates the overall architecture of the intelligent loan allocation judgment system based on a large-scale model. The system workflow is as follows: A multi-source data acquisition module collects multi-dimensional data from enterprises and then passes it to a data preprocessing module for processing. Simultaneously, a policy knowledge base module provides policy-related knowledge support to the large-scale model analysis module and the rule engine module. The large-scale model analysis module and the rule engine module perform analysis and rule verification based on the preprocessed data and policy knowledge, respectively. The results from both are input to a result fusion and decision-making module to generate preliminary judgment results. If manual intervention is required, the results are transferred to a manual review and feedback module. Feedback from this module is then passed to the knowledge base and model update module to update the policy knowledge base and optimize the large-scale model. Finally, a user interaction module enables interaction with users, displaying judgment results and other information.

[0043] like Figure 2 The diagram shows the data preprocessing flowchart, outlining the specific steps: First, raw data input is performed, importing various types of raw data collected from the enterprise; next, data cleaning and standardization operations are carried out, including noise reduction and data completion, and the data format is standardized; then, text extraction, recognition, and word segmentation are performed, extracting effective content from unstructured text and performing word segmentation; next, entity recognition is performed, identifying entity information such as project names; finally, structured indicators and unstructured keywords are generated, providing usable data for subsequent large-scale model analysis and other stages.

[0044] like Figure 3 The diagram shows the structure of the policy knowledge base, which comprises four parts: a policy rule base, a quantitative indicator base, a keyword base, and a case study base. The policy rule base stores policy documents and rules related to the "Five Major Articles" (a set of five key policy documents); the quantitative indicator base records the hard quantitative indicators for determining loan allocation in various fields; the keyword base contains core keywords for each field; and the case study base stores historical loan allocation classification cases. These four sub-bases collectively provide policy knowledge support for intelligent loan allocation determination.

[0045] Example:

[0046] A loan allocation intelligent identification system based on a large model includes the following modules:

[0047] Multi-source data acquisition module: Directly connects with the bank's credit system, enterprise credit reporting platform, business registration database, and patent database to collect structured and unstructured data from enterprises. Structured data includes registered capital, R&D investment ratio, and environmental certification numbers, while unstructured data includes annual reports, project feasibility reports, and environmental impact assessment documents. It supports two data acquisition modes: batch import and real-time capture. Batch import is used to process historical loan project data, while real-time capture is used to obtain the latest information of newly added enterprises, ensuring that the data covers the three core dimensions of basic enterprise information, operating data, and project details.

[0048] Data preprocessing module: Connected to the multi-source data acquisition module, it undertakes data cleaning and standardization functions. The processing flow is as follows: raw data input, data cleaning and standardization, text extraction, recognition and word segmentation, entity recognition, generation of structured indicators and unstructured keywords; For structured data: it verifies the integrity of fields. If the "environmental protection expenditure" field is missing, it will be automatically marked and a reminder will be sent to the account manager to complete it; it unifies the data format, converting all data related to "R&D investment ratio" into percentage form, such as "R&D investment of 5 million, revenue of 96 million" into "R&D investment ratio of 5.2%"; For unstructured data: it recognizes image-type text such as scanned environmental impact assessment reports using OCR technology, and uses the jieba word segmentation tool to split the text content, such as extracting entity information such as "patent number ZL202XXXXXXX" and "photovoltaic power station project", and clarifying the field attributes of the entities (such as the patent number corresponding to the technology field, and the photovoltaic project corresponding to the green field).

[0049] The policy knowledge base module consists of a policy clause library, a quantitative indicator library, a keyword library, and a case library. These four sub-libraries are interconnected and work together to support classification and judgment. The policy clause library is categorized into five major areas: science and technology finance, green finance, inclusive finance, pension finance, and digital finance. It stores the original text of relevant policy documents and their core clauses after breakdown, such as the "Special Action Plan for Financing of Science and Technology-based SMEs" and the "Green Credit Guidelines." The quantitative indicator library records the hard criteria for each area, such as "cumulative R&D investment ≥ 10 million yuan in the past 3 years" and "number of patents ≥ 3" for science and technology finance, and "number of employees ≤ 300" and "annual revenue ≤ 50 million yuan" for inclusive finance. The database includes: a "green finance" database (e.g., "environmental protection expenditure ≥ 3%)", a keyword database (labeled by industry, with green finance corresponding to "carbon emission reduction", "wind power projects", "ISO14001 certification"); elderly care finance corresponding to "long-term care insurance", "senior apartments", "elderly care industry fund"; and digital finance corresponding to "blockchain", "artificial intelligence", "digital investment", etc.); and a case database (stores historical categorized cases, each containing company data, categorization results, and judgment criteria, such as "a new energy company was judged as green finance due to 'photovoltaic projects' and 'ISO14001 certification'", and "a micro-enterprise was judged as inclusive finance due to '50 employees and annual revenue of 8 million yuan'".

[0050] Large-scale model analysis module: Connected to the data preprocessing module and the policy knowledge base module, it adopts a large-scale model fine-tuned based on financial corpus (such as a financial-specific model fine-tuned based on Llama2-70B); it receives preprocessed enterprise feature data (including structured indicators and unstructured keywords), extracts the core investment direction of projects through semantic understanding technology, such as identifying "technology field investment direction" from "core business is artificial intelligence chip R&D, with R&D investment of 12 million yuan in the past 3 years"; it matches the keyword library and case library of the policy knowledge base, and after comprehensively calculating the keyword matching degree and case similarity, it outputs preliminary classification results (such as "technology finance" and "green finance") and confidence levels (such as 92% and 89%).

[0051] The rules engine module connects to the large model analysis module and the policy knowledge base module. Its core function is to verify the compliance of the preliminary results of the large model, ensuring that they meet the quantitative requirements of the policy. For the preliminary classification results output by the large model, it automatically calls the quantitative indicator library of the corresponding field in the policy knowledge base for verification. For example, when the preliminary result is "science and technology finance", it verifies three indicators: "R&D investment ratio ≥ 5%", "Number of patents ≥ 3", and "Cumulative R&D investment in the past 3 years ≥ 10 million yuan". Verification result processing: If all indicators meet the standards, the credibility of the preliminary classification results is strengthened; if there are any items that do not meet the standards (such as "R&D investment ratio is only 3.2%)", the item is marked as "pending verification", and the non-compliant items and specific numerical differences are recorded in detail.

[0052] The Result Fusion and Decision Module, connected to the large model analysis module and the rule engine module, is the core of the system's decision-making process. It integrates two types of output results according to preset logic. The first type of decision logic is as follows: When the confidence level of the large model's preliminary classification result is ≥80%, and the rule engine verifies that all indicators meet the standards, the final classification result (e.g., "green finance") is directly generated, and the matching policy clause number (e.g., Article 5 of the "Green Credit Guidelines"), keywords (e.g., "wind power project"), and quantitative indicators (e.g., "environmental protection expenditure accounts for 4.5%) are associated. The second type of decision logic is as follows: When the confidence level of the large model is <80%, or the rule engine verifies that there are unmet items, the manual review process is triggered, the project is pushed to the expert workbench, and the analysis of doubts is output (e.g., "the text mentions 'blockchain technology application' but there is no proof of digital investment, and the proportion of digital investment does not meet the standard").

[0053] The manual review and feedback module, connected to the results fusion and decision-making module, is used to process "pending review" projects, achieving dual protection through "machine judgment + manual review"; Expert operation function: Displays full project data (basic enterprise information, financial data, project materials), preliminary results of the large model, and rule engine verification of doubts to experts, supporting experts to view original policy clauses; Experts manually determine the final category based on their professional judgment and enter the reasons for modification (e.g., "Although the digital investment ratio is 7.8%, slightly lower than 8%, the enterprise is a national-level digital economy pilot enterprise, which meets the special identification standards for digital finance, and is judged as digital finance"); Feedback mechanism: The final results and reasons marked by experts are automatically synchronized to the case library of the policy knowledge base, and are also stored as marked data in the large model training set for subsequent model optimization.

[0054] Module Descriptions (The following is an introduction to each module of this application and should not be construed as a limitation on any module. In actual use, you can choose according to your specific needs):

[0055] Multi-source data acquisition module: Based on computer network interfaces and data interfaces, this module allows for the selection of software code such as web crawlers and database query programs, depending on implementation needs. It collects structured and unstructured data from various data sources, including internal credit systems, enterprise credit reporting platforms, business registration databases, and patent databases. Data transmission and retrieval are achieved using network communication protocols (such as HTTP and TCP / IP), and database access technologies (such as JDBC and ODBC) are used to connect to various databases for data scraping.

[0056] Data preprocessing module: Using the computer's central processing unit (CPU) as the computing core, it temporarily stores data in memory and performs operations such as cleaning, format conversion, text extraction, and entity recognition on the collected data by writing data cleaning, standardization, and text processing algorithms (such as using Python's Pandas library for structured data processing, and using the jieba word segmentation library and OCR recognition software toolkit for unstructured text processing).

[0057] Policy knowledge base module: It relies on computer storage devices (such as hard drives and solid-state drives) to store data such as policy rules, quantitative indicators, keywords, and cases. It uses a database management system (such as MySQL and MongoDB) to build the data storage structure and uses data indexing and query algorithms to achieve efficient storage and retrieval of policy knowledge.

[0058] Large Model Analysis Module: Leveraging the high-performance CPUs and GPUs (Graphics Processing Units, used to accelerate large model calculations), this module utilizes memory and storage devices to support parameter storage and data processing for large models. Based on fine-tuned large models in the financial field (such as FinBERT), it loads model parameters and writes semantic matching and classification algorithms to perform semantic understanding and preliminary classification on preprocessed data. Rule Engine Module: Using the CPU as the computing unit and combining memory for rule matching calculations, this module writes rule parsing and indicator verification algorithms and calls quantitative indicator data from the policy knowledge base to perform compliance verification on the preliminary results output by the large model.

[0059] Results Fusion and Decision Module: Relying on the computer's CPU to integrate and calculate results, and using memory to store intermediate results, it can select decision logic according to actual needs, integrate the output results of the large model and the rule engine, and generate the final judgment result or trigger the manual review process according to the preset decision rules.

[0060] The manual review and feedback module enables interaction with human experts using computer input / output devices (such as monitors, keyboards, and mice). The interface display and data feedback methods can be designed according to actual needs. The module displays the items to be reviewed to the experts, collects the experts' annotation results, and synchronizes them to the policy knowledge base and the large model training set.

[0061] Knowledge base and model update module: It uses the computer's network interface to implement policy crawling function, which can crawl policy updates according to the selected web crawling program and update the policy knowledge base with the help of storage devices; it uses CPU and GPU as the computing core, and uses memory and storage devices to support fine-tuning training of large models. Through model training algorithms, it optimizes large models based on new cases and human feedback data.

[0062] User interaction module: Relying on computer input and output devices such as monitor, keyboard, and mouse, it uses a visual interface program (such as using Web front-end technology or desktop application development framework) to display classification results, decision-making basis, and other information to users, and supports user interaction (such as experts modifying classification results).

[0063] Finally, it should be noted that the present invention is not limited to the above embodiments, and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A loan allocation intelligent identification system based on a large model, characterized in that, include: The multi-source data acquisition module can connect with the bank's credit system, enterprise credit platform, business registration database, and patent database to collect structured and unstructured data of enterprises. It supports both batch import and real-time capture modes. The structured data includes registered capital, R&D investment ratio, and environmental certification number. The unstructured data includes annual reports, project feasibility reports, and environmental impact assessment documents. The data preprocessing module, connected to the multi-source data acquisition module, is used to clean and standardize the acquired data, verify the integrity of fields and unify the format of structured data, and use OCR to recognize image-type text and use word segmentation tools to split text and extract entity information for unstructured data. The processing flow is as follows: raw data input, data cleaning and standardization, text extraction, recognition and word segmentation, entity recognition, and generation of structured indicators and unstructured topic words. The policy knowledge base module includes a policy rule library, a quantitative indicator library, a keyword library, and a case library, which are used to store relevant policy documents, hard indicators in various fields, core terms, and historical categorized cases. The large model analysis module is connected to the data preprocessing module and the policy knowledge base module. It adopts a large model finely tuned in the financial field, receives preprocessed enterprise data, extracts the core investment direction of the project through semantic understanding, and matches it with the keyword library and case library in the policy knowledge base module to output preliminary classification results and confidence scores. The rules engine module, connected to the large model analysis module and the policy knowledge base module, is used to call the quantitative indicator library to perform hard indicator verification on the preliminary classification results output by the large model. If the indicators meet the standards, the classification results are strengthened; if they do not meet the standards, they are marked as "to be reviewed" and the non-compliant items are output. The results fusion and decision module is connected to the large model analysis module and the rule engine module. It is used to combine the outputs of the two to generate the final judgment result. When the confidence of the large model is ≥80% and the rule engine verification is passed, the classification result and basis are directly output. When the confidence is <80% or the rule engine verification is not passed, the manual review process is triggered and the doubt analysis is output. The manual review and feedback module is connected to the result fusion and decision-making module. It is used to show project data, preliminary model results and doubts to experts, receive final classification and annotations from experts, and automatically synchronize the expert annotation results to the case library and large model training set of the policy knowledge base module. The knowledge base and model update module is connected to the policy knowledge base module and the large model analysis module, and trains the large model based on new cases and human feedback data; The user interaction module, connected to the result fusion and decision-making module and the manual review and feedback module, is used to provide a visual interface to show account managers the final classification results, decision basis, and risk warnings.

2. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, The data preprocessing module automatically marks and completes the missing "environmental protection expenditure" in structured data, and converts "R&D investment ratio" into a percentage format. For unstructured data, it uses the jieba word segmentation tool to split the text and extract entity information including "patent number" and "photovoltaic power station project name".

3. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, The policy knowledge base module stores relevant policy documents in the policy rule base according to the fields of science and technology finance, green finance, inclusive finance, pension finance, and digital finance. The quantitative indicator library records: 1) Technology finance requirements: cumulative R&D investment ≥ 10 million yuan in the past 3 years; 2) Inclusive finance requirements: hard indicators for each field for companies with ≤ 300 employees; The keyword library is labeled by field with core terms such as "carbon emission reduction" and "wind power projects" for green finance, and "long-term care insurance" and "senior apartments" for elderly care finance; The case library stores historical categorized cases, including cases of new energy companies being judged as green finance due to photovoltaic projects and ISO14001 certification.

4. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, The preliminary classification results output by the large model analysis module include "financial technology", "green finance", "inclusive finance", "pension finance" and "digital finance", with a confidence level range of 0-100%. These results are used to assist the result fusion and decision-making module in determining whether to trigger the manual review process.

5. The intelligent loan allocation discrimination system based on a large model according to claim 4, characterized in that, When the rule engine module performs hard indicator verification on the preliminary classification results, if the large model determines it to be "technology finance", it will automatically verify the indicators "R&D investment ratio ≥ 5%" and "number of patents ≥ 3"; if the large model determines it to be "digital finance", it will automatically verify the indicator "digital investment ratio ≥ 8%".

6. The intelligent loan allocation discrimination system based on a large model according to claim 5, characterized in that, The manual review and feedback module synchronizes the expert annotation results to the case library and the large model training set, including the final category and reason for the expert judgment, which is used to optimize the verification criteria of the model parameters and rule engine.

7. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, The knowledge base and model update module crawls policy updates once a week via a policy crawler and trains the large model once a month to update the model's semantic matching parameters and classification weights.

8. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, The risk warnings displayed to account managers by the user interaction module include information about the company's environmental penalty records. The classification criteria output include the matching policy clause number, keywords, and specific values ​​of quantitative indicators.

9. The intelligent loan allocation discrimination system based on a large model according to claim 1, characterized in that, When the multi-source data acquisition module collects data, it automatically sends a prompt to the account manager to complete the data for companies that have not provided an environmental impact assessment report, ensuring that unstructured data covers environmental protection-related information of the project.

10. A method for intelligently identifying loan allocation based on a large model, characterized in that, Includes the following steps: Step 101: Multi-source data collection. Automatically capture structured and unstructured data such as business registration information, financial statements, project contracts, and patent certificates of enterprises through interfaces. Automatically send a prompt to the account manager to complete the data if an environmental impact assessment report is not provided. Step 102: Data preprocessing. The structured data is formatted and standardized, converting "R&D expenses of 5 million" into "R&D investment ratio of 5.2%". Unstructured data is extracted and entity recognition is performed. "Core business is artificial intelligence chip R&D" is extracted from the annual report, and the technology field to which the patent corresponding to the "patent number" belongs is marked. The processed enterprise feature data is output. The processing flow is as follows: raw data input, data cleaning and standardization, text extraction, recognition and word segmentation, entity recognition, generation of structured indicators and unstructured keywords. Step 103: Load policy knowledge. Load the latest policy provisions, quantitative indicators, keywords and historical cases from the policy knowledge base. If a policy update is detected that green finance has added a new "hydrogen energy project" classification standard, the loaded content will be updated automatically. Step 104: Initial identification by the large model. Input the enterprise feature data into the large model, identify the core investment direction of the project through semantic matching, compare it with the keyword library and case library of the policy knowledge base, and output the preliminary classification results and the percentage of confidence. The preliminary classification results include "digital finance". Step 105: Rule engine verification. Based on the preliminary classification results, the corresponding quantitative indicator library is called for verification. "Digital finance" must meet the requirement of "digital investment ratio ≥ 8%". If all indicators meet the requirements, proceed to step 106. If there are any non-compliant items, mark them as "pending review" and record the reasons for non-compliance. Step 106: Results fusion and decision-making. When the confidence level of the preliminary classification result is ≥80% and the rule engine verification is passed, the final classification result is generated. The final classification result includes "inclusive finance" and the associated quantitative indicators. When the confidence level is <80% or the verification fails, the manual review process is triggered. Step 107: Manual review and feedback. Experts review the project data, preliminary model results and doubts, manually determine the final category and enter the reasons. The expert's determination results and reasons are synchronized to the case library and used as training data. Step 108: Knowledge base and model update. The official policy website is automatically crawled weekly to parse new policy content and update the knowledge base. New cases and human feedback data are used monthly to incrementally train the large model and update the model parameters. Step 109: Results Output and Display. The final classification results, decision basis, and risk warnings regarding the company's environmental penalty records are displayed to the account manager through the user interface. The results can be exported.

Images