An industry policy intelligent matching and declaration assistance method and system based on a large model

Abstract

The application discloses an industry policy intelligent matching and declaration auxiliary method and system based on a large model, and belongs to the field of artificial intelligence and big data technology. The method automatically collects and deeply analyzes multi-source policy files through an NLP large model, extracts structured elements, and constructs a policy knowledge graph and a dynamic knowledge base; multi-source data is integrated by enterprise authorization, and a multi-dimensional digital portrait of the enterprise is constructed and updated in real time; a three-level progressive engine of label preliminary screening, semantic depth matching, compliance verification and conflict detection is used to realize accurate matching of the policy and the enterprise; an individualized policy list and a whole-process declaration navigation are generated, and declaration materials are intelligently assisted to be generated; the declaration progress and policy dynamics are monitored and early warned in real time, and review and model iteration optimization are completed based on declaration data. The system adopts a four-layer architecture of a data layer, a model layer, a core function layer and an application layer. The policy matching accuracy and the declaration efficiency are significantly improved, and the system is suitable for intelligent analysis, accurate matching and whole-process declaration auxiliary scenes of enterprise-friendly policies.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and big data technology, specifically relating to a method and system for intelligent matching and application assistance of industrial policies based on a large model. Background Technology

[0002] Currently, governments at all levels have introduced numerous pro-business industrial policies focusing on technological innovation, industrial support, employment stabilization, and tax reductions, which have become core supports for enterprise development and industrial upgrading. However, the existing policy service model still relies mainly on manual release, enterprise self-inquiry, and offline guidance, supplemented only by simple keyword tag matching tools. This model has significant shortcomings in both technical implementation and practical service, failing to meet enterprises' needs for efficient policy access, accurate policy matching, and smooth policy application. It also makes it difficult to accurately assess the effectiveness of government policy implementation. Specific technical deficiencies and pain points are as follows: 1. Policy information is scattered and the efficiency of obtaining it is extremely low: Policy releases cover multiple administrative levels such as national, provincial, municipal, district and industrial park, and are scattered across dozens of government functional departments such as development and reform, industry and information technology, science and technology, human resources and social security. Enterprises need to manually search across platforms and channels, making it difficult to obtain comprehensive and timely effective policy information, and are very likely to miss the policy application window.

[0003] 2. High threshold for policy interpretation and large errors in manual interpretation: Policy documents are mostly unstructured professional texts with complex clauses and dense professional terminology. Small and medium-sized enterprises lack professional policy interpretation personnel and cannot accurately judge their own suitability for policies. They generally have the problem of "not being able to understand or grasp" the policies. At the same time, manual interpretation of policies is inefficient and lacks standardization, which cannot adapt to the dynamic update needs of massive policies and is prone to interpretation errors.

[0004] 3. Insufficient policy matching accuracy and serious mismatches and omissions: Traditional policy matching tools only perform simple surface matching based on keywords and fixed tags. They lack the ability to deeply understand the semantic logic of policy clauses and the multidimensional business conditions of enterprises. The matching granularity is coarse and the flexibility is poor. There are common problems of missing highly applicable policies and mismatches that do not meet policy requirements. The actual matching accuracy is less than 30%.

[0005] 4. Lack of closed-loop support throughout the application process, resulting in high compliance risks: Enterprise policy application processes are complex, with numerous material requirements and strict node control. Existing technologies lack intelligent navigation and material support capabilities throughout the application process, making it easy for enterprises to encounter problems such as missing materials, filling errors, and node delays. At the same time, the lack of automated policy mutual exclusion detection and compliance verification mechanisms makes it easy for duplicate applications and non-compliant applications to occur, leading to application failures.

[0006] 5. Lack of dynamism in services and insufficient exploitation of data value: Existing tools cannot achieve real-time dynamic updates of enterprise digital profiles and policy databases, nor can they continuously optimize policy matching models based on historical data submitted by enterprises. The value of the data has not been effectively exploited, and it cannot provide data support for evaluating the effectiveness of policy implementation. The level of intelligence in policy services is seriously insufficient.

[0007] Therefore, there is an urgent need for an intelligent matching and application assistance method and system for industrial policies, in order to achieve intelligent and structured parsing of policy documents, dynamic construction of multi-dimensional digital profiles of enterprises, multi-level accurate matching of policies and enterprises, and full-process application assistance and progress monitoring, so as to completely break through the "last mile" of the implementation of policies that benefit enterprises. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for intelligent matching and application assistance of industrial policies based on a large model, which at least solves the problems in the background technology and improves the intelligence level of policy services, matching accuracy and enterprises' sense of policy gain.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for intelligent matching and application assistance of industrial policies based on a large model includes the following steps: S1. Deep analysis of policy documents and construction of structured knowledge base based on NLP big model: Automated collection and preprocessing of multi-source policy documents, extraction of core structured elements of policies through domain-fine-tuned NLP big model, and construction of policy knowledge graph and dynamically updated policy knowledge base; S2. Dynamic Construction of Multi-Dimensional Digital Profiles of Enterprises: Under the premise of enterprise compliance and authorization, integrate multi-source enterprise data, construct a multi-level enterprise tag system, and generate and dynamically update enterprise digital profiles. S3, Multi-level Matching and Inference Engine: Precise matching of policies and enterprises: Through a three-level progressive matching process, precise matching of policies and enterprises is achieved, and compliant and highly adaptable policy results are output. S4. Personalized Policy Matching List and Application Navigation Path Generation: Based on the output of step S3, a policy matching list exclusive to the enterprise is generated, the entire application process is broken down and a visual navigation path is generated, and application materials are intelligently assisted in being generated and optimized. S5. Dynamic monitoring and follow-up of the entire application process: Real-time monitoring and early warning of the progress of enterprise application projects and policy dynamics, and completion of review analysis and model optimization based on enterprise application data and results.

[0010] Preferably, step S1 specifically includes: By connecting with official channels and using distributed crawlers, we can achieve full automated collection of policy documents and preprocess the collected documents to form a standardized original policy text library. The preprocessing includes format cleaning, text extraction, deduplication and noise reduction, and version verification. Based on the policy domain annotated corpus, a general NLP model is supervised and fine-tuned to construct a dedicated model for extracting policy elements. This model performs end-to-end deep analysis of unstructured policy texts and automatically extracts the core structured elements of the policy. These core structured elements include at least the policy basic information, applicable objects, application conditions, support standards, process rules, and restrictive clauses. The extracted core structured elements of the policy are linked by entities, semantically normalized, and relationally modeled to construct a policy knowledge graph. The policy knowledge graph defines at least the entity types of "policy-issuing department-industry-qualification-intellectual property" and the association relationships of "issuance, application, requirement, reward, and mutual exclusion". Establish a dynamic knowledge base update mechanism to synchronize the release, revision, extension, and expiration status of policies in real time.

[0011] Preferably, step S2 specifically includes: Under the premise of enterprise compliance authorization, collect basic business registration information, operating and financial data, R&D and innovation data, qualification and honor information, credit compliance information and historical policy application information of enterprises; By using NLP large-scale models to perform semantic understanding and standardization processing on the collected data, a multi-level enterprise tag system is constructed. The multi-level enterprise tag system at least covers basic attribute tags, business capability tags, R&D innovation tags, qualification and honor tags, credit compliance tags, and application history tags. Set a data synchronization cycle for enterprises to synchronize dynamic information in real time, and realize the real-time dynamic update of the enterprise's digital profile. The dynamic information includes at least changes in enterprise registration, addition of intellectual property rights, updates to operating data, and acquisition of qualifications.

[0012] Preferably, the three-level progressive matching process in step S3 specifically includes: Level 1, Rapid Tag Initial Screening and Matching: Extract hard access tags that are vetoed by policies, accurately match them with enterprise profile tags, filter out policies that do not meet the requirements, and retain the set of policies that pass the initial screening; Level 2, Large-Scale Semantic Deep Matching and Fit Scoring: The policy clauses and enterprise profile information are encoded into policy semantic vectors and enterprise semantic vectors respectively through the NLP large-scale model. The cosine similarity between the policy semantic vectors and enterprise semantic vectors is calculated. At the same time, the zero-shot reasoning capability of the NLP large-scale model is used to judge the fit of each policy clause. The overall fit score is calculated based on the weighted weights set according to the importance of the clauses, and the policy set with high fit is selected. Level 3, Compliance Verification and Conflict Detection: Checks whether the enterprise has any circumstances that are explicitly prohibited from being declared by policy documents. Based on the policy knowledge graph, it completes mutual exclusion detection between policies to be declared and duplicate declaration conflict detection between policies to be declared and previously approved policies, and outputs the final accurate matching result of compliance.

[0013] Preferably, step S4 specifically includes: Based on the output of step S3, a personalized policy matching list is generated, which includes basic policy information, suitability score, core reward standards, enterprise compliance items, items to be improved, and application priority suggestions. For the policies in the list, the application process nodes are broken down based on the NLP big data model, and a visual application navigation path is generated, which clarifies the operation requirements, material list and precautions for each node. The nodes include material preparation, online application, acceptance and review, expert review, public announcement and fund disbursement. For core application materials, the system utilizes a large NLP model combined with enterprise profiles and policy application requirements to automatically generate initial drafts of the application materials and provides suggestions for content optimization, logical improvement, and compliance verification. The core application materials include at least the project application form, feasibility study report, and a list of qualification certificates.

[0014] Preferably, step S5 specifically includes: Connect to the policy application and acceptance platform to track the progress of the application projects submitted by enterprises throughout the entire process, synchronize the node status in real time, and push progress reminders and node warnings. The node status includes at least acceptance, review, evaluation, public announcement and disbursement. Real-time monitoring of new policies, revision announcements, application extensions, and window period adjustments in the policy database; automatic matching of appropriate policy dynamics based on enterprise profiles; and push of early warning information to enterprises. Based on the historical data and results of enterprise applications, a retrospective analysis is conducted using a large NLP model to output an analysis of application success rate, the core reasons for non-approval, and suggestions for subsequent application optimization. Furthermore, the weight parameters and semantic encoding capabilities of the matching model are continuously optimized based on the application result data.

[0015] Accordingly, the present invention also provides an intelligent matching and application assistance system for industrial policies based on a large model, including a data layer, a model layer, a core function layer and an application layer; The data layer is used to store, manage, and retrieve all system data. The data layer includes a policy text library, a multi-source enterprise database, a policy knowledge graph, a vector database, and a historical application database. The model layer is used to provide the core AI capabilities of the system. The model layer includes a large model for policy element extraction, a semantic encoding model, a multi-level matching reasoning model, and a large model for material generation. The core functional layer is used to implement the entire process of the above methods. The core functional layer includes a policy collection and analysis module, an enterprise profile building module, a multi-level matching and inference engine, an application assistance and navigation module, and an application progress monitoring module. The application layer is used to provide visual interactive interfaces for different user groups. The application layer includes enterprise user terminals, government management terminals, and operation management terminals.

[0016] Preferably, in the core functional layer: The policy collection and parsing module is used for the automated collection and preprocessing of policy documents from multiple sources, extraction of policy elements, construction of policy knowledge graphs, and dynamic updating and maintenance of policy knowledge bases. The enterprise profile building module is used for compliant collection and standardized processing of enterprise authorized data, systematic management of enterprise tags, construction and dynamic updating of multi-dimensional digital profiles of enterprises, and enterprise data security and access management. The multi-level matching and inference engine has built-in tag screening unit, semantic deep matching unit, and compliance verification and conflict detection unit to fully execute the three-level progressive matching process; The application assistance and navigation module is used to generate personalized policy matching lists, break down and visualize the entire application process navigation path, intelligently generate and optimize application materials, provide application milestone reminders, and answer frequently asked questions. The application progress monitoring module is used for tracking the entire process of enterprise application projects, real-time monitoring and early warning of policy dynamics, review and analysis of application data, iterative optimization of matching models, and statistical analysis of policy implementation effects.

[0017] Compared with the prior art, the present invention has the following technical effects: 1. Fully automated policy analysis significantly lowers the barrier to interpretation: Through domain-fine-tuned NLP models, fully automated deep analysis and structured processing of unstructured policy documents are achieved, solving the problems of low efficiency, large errors, and high professional threshold of manual interpretation. The accuracy rate of policy element extraction can reach over 95%. At the same time, the policy knowledge graph enables full-linkage management of policy elements, ensuring the comprehensiveness and timeliness of policy information and alleviating the problems of scattered policy information and difficulty in obtaining it.

[0018] 2. Innovative three-level matching architecture, balancing efficiency and accuracy: Breaking through the technical bottlenecks of traditional keyword matching, an innovative three-level progressive matching engine is designed: "initial tag screening - deep semantic matching - compliance verification". The first level of initial screening achieves millisecond-level filtering, greatly improving matching efficiency; the second level of large-scale model semantic matching achieves clause-level fine-grained adaptation, solving the problems of mismatch and omission; the third level of compliance verification avoids application risks. Finally, the policy matching accuracy rate can reach more than 90%, far exceeding the 30% accuracy rate of traditional tools.

[0019] 3. A closed-loop service for the entire application process to improve the success rate of enterprise applications: From precise policy matching, application navigation, and intelligent material generation to progress monitoring, dynamic early warning, and post-event optimization, a closed-loop service for preferential policies for enterprises has been formed. This completely solves the core pain points of enterprises "not being able to find, not understanding, and not being able to submit well". It can shorten the preparation time of enterprise application materials by 80% and increase the success rate of application by more than 60%. At the same time, through compliance verification and conflict detection, it effectively avoids compliance risks such as duplicate applications and non-compliant applications.

[0020] 4. Dynamic service capabilities enable continuous system optimization and iteration: Real-time dynamic updates of the policy knowledge base and enterprise digital profiles, real-time monitoring and precise push of policy dynamics, and continuous iteration and optimization of the matching model based on application data solve the problem of lack of dynamism in existing tools and services, and achieve continuous improvement of system capabilities.

[0021] 5. Enhance the effectiveness of policy implementation and achieve mutual empowerment between government and enterprises: For enterprises, it enables precise delivery of preferential policies, significantly reducing the time and manpower costs of policy acquisition, interpretation, and application, and enhancing enterprises' sense of policy gain; for the government, intelligent and precise matching greatly improves the efficiency of policy fund allocation and policy implementation; at the same time, the accumulated data of the entire application process can provide data support for government departments in policy formulation, effect evaluation, and optimization, achieving mutual empowerment between government and enterprises, promoting the upgrading of industrial policy services from "manual and fragmented" to "intelligent, standardized, and closed-loop", and improving the service level of the entire industry. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of intelligent matching and application assistance for industrial policies based on a large model. Figure 2 This is an architecture diagram of an intelligent matching and application assistance system for industrial policies based on a large model. Figure 3 This is a diagram illustrating the working principle of a multi-level matching and inference engine. Detailed Implementation

[0023] The following detailed description illustrates the specific implementation method: Example

[0024] like Figure 1 As shown, a method for intelligent matching and application assistance of industrial policies based on a large model includes five core execution steps, forming a data closed loop throughout the entire process, as detailed below: Step S1: Deep analysis of policy documents and construction of a structured knowledge base based on NLP large-scale models This step is the foundation of this method and one of the core innovations of this invention. It is used to achieve automated and high-precision extraction of all policy elements, as well as the standardization and structuring of unstructured policy texts. It constructs an interconnected policy knowledge graph and a dynamically updated policy knowledge base, solving the technical problems of scattered policy texts, difficult interpretation, and low standardization. Specifically, it includes the following sub-steps: S11. Automated collection and preprocessing of multi-source policy documents: Connect with official channels including government websites at all levels, policy release platforms of functional departments, management committees of industrial parks, and industry associations. Use distributed crawlers to achieve full-volume automated collection of policy documents, covering at least policy notices, implementation rules, supplementary announcements, and application guidelines. Clean the collected policy documents, extract the text, remove duplicates and noise, and verify the version to form a standardized original policy text library.

[0025] S12. Construction of a Domain-Specific Policy Element Extraction Model: Based on a massive labeled policy domain corpus, a supervised fine-tuning of a general NLP model is performed to optimize the model's specific capabilities in policy text entity recognition, relation extraction, and clause comprehension, constructing a dedicated policy element extraction model. Through this fine-tuned model, end-to-end deep parsing of unstructured policy texts is conducted to automatically extract core structured policy elements. These core structured policy elements include at least the following: Basic policy information: policy name, issuing authority, issuance date, implementation period, policy level, and policy category; Applicable to: at least including enterprise type, industry sector, size classification, registered location requirements, and years of establishment; Application requirements: hard indicators (revenue, R&D investment ratio, number of intellectual property rights, qualification certification, staff size), soft requirements (industry sector, project type, credit status), and veto items; Support criteria should include at least the amount of financial rewards, tax reduction and exemption ratios, talent subsidies, land use preferences, and qualification certification; Process rules: application deadlines, accepting departments, list of materials, processing procedures, review stages, and rules for public announcement and disbursement; Restrictive clauses: These should include at least the circumstances under which no application is permitted, rules on mutually exclusive policies, and restrictions on duplicate applications.

[0026] S13. Construction of Policy Knowledge Graph and Dynamic Knowledge Base: The extracted core structured policy elements are linked by entities, semantically normalized, and relationally modeled to construct a policy knowledge graph that defines at least the entity types of "policy-issuing department-industry-qualification-intellectual property" and the relationships of "issuance, application, requirement, reward, and mutual exclusion." A dynamic update mechanism for the knowledge base is established to monitor the issuance, revision, extension, and expiration status of policies in real time, and to update element information synchronously to ensure the timeliness and accuracy of the knowledge base.

[0027] Step S2: Dynamic Construction of Enterprise Multidimensional Digital Profile This step provides enterprise-side data support for the matching process, enabling the standardization and tagging of enterprise information across all dimensions. Specifically, it includes the following sub-steps: S21. Enterprise Authorization Multi-Source Data Integration: Under the premise of enterprise compliant authorization, it connects with the industrial and commercial information platform, the State Intellectual Property Office, the enterprise credit information publicity system, the tax system and the enterprise's self-developed operation and management system to collect enterprise full-dimensional data in six categories, including basic industrial and commercial information, operating and financial data, R&D and innovation data, qualification and honor information, credit compliance information and historical policy application information.

[0028] S22. Multi-level Enterprise Tagging System and Dynamic Profile Construction: Based on the collected full-dimensional enterprise data, semantic understanding and standardization are performed through NLP large-scale models to construct a multi-level enterprise tagging system that covers at least basic attribute tags, operational capability tags, R&D innovation tags, qualification and honor tags, credit compliance tags, and application history tags; an enterprise data synchronization cycle is set to synchronize dynamic information such as enterprise business registration changes, new intellectual property rights, updated operational data, and qualification acquisition in real time, so as to realize the real-time iteration of the enterprise digital profile and ensure that the profile is completely matched with the actual situation of the enterprise.

[0029] Step S3, Multi-level Matching and Inference Engine Policy - Precise Enterprise Matching This step is the core innovation of this invention, which designs a three-level progressive matching and reasoning architecture. Through the architecture of "initial screening with hard labels - deep semantic matching - compliance conflict detection", it balances the efficiency and accuracy of policy matching, breaks through the technical bottleneck of coarse granularity and low accuracy of traditional label matching, and realizes the fine-grained adaptation of policy provisions to the multi-dimensional situation of enterprises. Specifically, it includes the following sub-steps: S31, Level 1: Rapid Tag Initial Screening and Matching Extract the policy's veto-type hard access tags and accurately match them with enterprise profile tags. Quickly filter out policies that do not meet the access requirements at all, and only retain policies that meet all hard thresholds to enter the next level of matching. The veto-type hard access tags should at least include the place of registration, enterprise type, years of establishment, industry sector, and required qualifications. This step can filter out more than 90% of invalid policies, greatly reducing the amount of subsequent calculations and improving the overall matching efficiency.

[0030] S32, Level Two: Large Model Semantic Deep Similarity Matching and Fit Scoring For policies that pass the initial screening, the deep semantic understanding capabilities of NLP models are used to achieve a refined matching between policy provisions and enterprise conditions. The specific process is as follows: All application clauses, applicable requirements, and support scenarios of the policy are encoded into fixed-dimensional policy semantic vectors through a large model and stored in a vector database. The core information of an enterprise is encoded into an enterprise semantic vector. The core information includes at least the enterprise's full-dimensional profile information, operating status, R&D capabilities, and projects under development. The cosine similarity between the policy semantic vector and the enterprise semantic vector is calculated. At the same time, the zero-shot reasoning capability of the large model is used to make a judgment on the suitability of each application clause of the policy, and output the compliance score of each clause, explanation of non-compliance, and optimization and improvement items. Based on the importance of the clauses, a weighted average is set, and the overall fit score between the policy and the enterprise is calculated. The policies are then sorted in descending order of the total score, and those with a fit score higher than the preset threshold are selected to proceed to the next level of verification.

[0031] S33, Level 3: Compliance Verification and Conflict Detection For policies with high compatibility, the final compliance verification and risk assessment are completed by combining policy restrictions, historical enterprise application data, and policy incompatibilities rules. The specific process is as follows: Compliance verification: Check whether the company has any circumstances that are explicitly prohibited from filing under the policy, and eliminate policies that do not meet the compliance requirements. Prohibited circumstances include records of being a dishonest person subject to enforcement, major administrative penalties, and past records of falsifying applications; Conflict detection: Based on the mutual exclusion relationship in the policy knowledge graph, it detects whether there are mutually exclusive application restrictions between policies to be applied for in the same batch or in the same funding pool. At the same time, it checks whether there are conflicts between the policies that the enterprise has previously approved and the policies to be applied for, such as duplicate applications or duplicate enjoyment of support, and generates conflict risk warnings. The output will ultimately be a policy matching result that passes compliance verification and has no core conflict risk.

[0032] Step S4: Generation of Personalized Policy Matching List and Application Navigation Path This step is the value output stage of the present invention, realizing the visualization and practical transformation of the matching results, and specifically includes the following sub-steps: S41. Personalized Policy Matching List Generation: Based on the output of the multi-level matching engine, a unique policy matching list is generated for enterprises. The list includes the policy name, issuing unit, application deadline, suitability score, core reward criteria, enterprise compliance items, items to be improved, and application priority suggestions. It supports one-click filtering and detailed viewing.

[0033] S42. Generation of a full-process application navigation path: For each policy in the list, the application process is broken down based on an NLP big data model to generate a visual application navigation path. The entire application process is broken down into 6 core nodes: material preparation, online application, acceptance and review, expert review, public announcement and fund disbursement. The processing time limit, operation steps, material requirements, precautions and solutions to common problems for each node are clearly defined, providing enterprises with a "step-by-step" application guide.

[0034] S43. Intelligent Assisted Generation of Application Materials: For core application materials, including at least a project application form, feasibility study report, and a list of qualification certificates, the system uses a large NLP model combined with enterprise profile information and policy application requirements to automatically generate draft materials and provide suggestions for content optimization, logical improvement, and compliance verification, significantly reducing the threshold for enterprises to write application materials.

[0035] Step S5: Dynamic monitoring and follow-up of the entire application process This step is used to achieve closed-loop management of the entire application service, and specifically includes the following sub-steps: S51. Real-time monitoring of application progress: Connects to the policy application acceptance platform to track the progress of the application projects submitted by enterprises throughout the entire process, and synchronizes the status of at least the nodes including acceptance, review, evaluation, public announcement and disbursement in real time, and pushes progress reminders and node warnings to enterprises as soon as possible.

[0036] S52. Real-time policy updates: The system monitors new policies, revision announcements, application extensions, and window period adjustments in the policy database 24 / 7. Based on enterprise profiles, it automatically matches appropriate policy updates and pushes early warning information to enterprises in a timely manner to prevent them from missing application opportunities.

[0037] S53. Review and Optimization of Application Data: Based on the historical data and results of enterprise applications, a review analysis is conducted using a large NLP model to output an analysis of application success rate, the core reasons for non-approval, and suggestions for subsequent application optimization. At the same time, based on the application result data, the weight parameters and semantic encoding capabilities of the matching model are continuously optimized to improve matching accuracy.

[0038] The method in this embodiment constructs a closed-loop service system covering the entire process of policy matching, application navigation, material assistance, progress monitoring, and review and optimization. Based on a large model, it realizes the visual decomposition of the application path and the intelligent generation of application materials, forming a complete intelligent service solution for preferential policies for enterprises, filling the gap in the existing technology for full-process application assistance capabilities.

[0039] To implement the above methods, this invention also provides an intelligent matching and application assistance system for industrial policies based on a large model. This system adopts a layered architecture design (e.g., Figure 2As shown), the various modules work together to achieve the complete workflow of the above methods. This system includes a data layer, a model layer, a core function layer, and an application layer. The data layer is used to store, manage, and access all system data. It includes a policy text library, a multi-source enterprise database, a policy knowledge graph, a vector database, and a historical application database.

[0040] The model layer is based on a finely tuned NLP large model, which provides the system with core AI capabilities including parsing, encoding, matching, reasoning and generation. The model layer includes a large model for policy element extraction, a semantic encoding model, a multi-level matching reasoning model and a large model for material generation.

[0041] The core functional layer, as the core functional unit of this system, is used to implement the five core steps of the above-mentioned methods. The core functional layer includes a policy collection and parsing module, an enterprise profile construction module, a multi-level matching and inference engine, an application assistance and navigation module, and an application progress monitoring module. These modules work together to complete the entire process service of the above methods. Specifically, the policy collection and parsing module is used for multi-source automated collection of policy documents, text cleaning and preprocessing, as well as policy element extraction based on NLP large-scale models, policy knowledge graph construction, and dynamic updating and maintenance of the policy knowledge base. The enterprise profile construction module is used for compliant collection of enterprise authorized data, data cleaning and standardization processing, systematic management of enterprise tags, construction and dynamic updating of multi-dimensional digital profiles of enterprises, and enterprise data security and access control. The multi-level matching and inference engine is the core module of this system, and its working principle diagram is shown below. Figure 3 As shown, the multi-level matching and inference engine has built-in tag initial screening unit, semantic deep matching unit, and compliance verification and conflict detection unit to fully execute the three-level progressive matching process and output accurate policy matching results; the application assistance and navigation module is used for personalized policy matching list generation, full-process application navigation path decomposition and visualization, intelligent generation and optimization of application materials, application node reminders, and answers to frequently asked questions; the application progress monitoring module is used for full-process progress tracking of enterprise application projects, real-time monitoring and early warning of policy dynamics, review and analysis of application data, iterative optimization of matching models, and statistical analysis of policy implementation effect data.

[0042] The application layer provides a visual interactive interface for different user groups, offering corresponding operation functions and service entry points. The application layer includes enterprise user terminals, government management terminals, and operation management terminals.

[0043] Taking Chongqing-based specialized and innovative SME A as an example, the company's core information is as follows: Registered in Chongqing High-tech Zone, established in 2018, belonging to the software and information technology services industry, it is a high-tech enterprise and a Chongqing specialized and innovative SME, with 85 employees, revenue of 62 million yuan in 2025, R&D investment accounting for 8.5%, possessing 3 invention patents and 12 software copyrights, with no record of dishonesty or administrative penalties, and 3 successful applications for science and technology policies in the past 3 years. Based on the method and system of this embodiment, the specific implementation process is as follows: Step 1: Construction of Policy Knowledge Base After the system is launched, the policy collection and analysis module connects to official channels of at least the National Development and Reform Commission, the Ministry of Science and Technology, the Chongqing Municipal Commission of Economy and Information Technology, and the Chongqing High-tech Zone Management Committee through distributed crawlers, automatically collecting at least 1,286 preferential policies for enterprises in the fields of science and technology, industry and information technology, and services. The collected policy documents are preprocessed to form a standardized original policy text library. Through a large-scale model for extracting policy elements with domain fine-tuning, all policy documents are analyzed and core structured elements are extracted. Based on the extracted elements, a policy knowledge graph is constructed, and a dynamic update mechanism for the policy knowledge base is established to form a dynamically updated policy knowledge base, which monitors the effectiveness / invalidation status of policies in real time.

[0044] Step 2: Building a Digital Profile for the Enterprise After Company A completes compliance authorization, the company profile building module automatically connects to the industrial and commercial information platform, the State Intellectual Property Office, the Chongqing Enterprise Credit Information Disclosure System, and the company's self-developed business management system. It collects the company's basic industrial and commercial information (registered capital, establishment date, registered address), operating and financial data (revenue and profit for the past three years), R&D and innovation data (R&D investment ratio, number of R&D personnel), qualification and honor information (high-tech enterprise, specialized and innovative enterprise certificate), credit and compliance information (no violation records), and historical policy application information (3 successful applications). Through NLP large-scale modeling, the collected data undergoes semantic understanding and standardization processing to generate a multi-dimensional digital profile of Company A. For example, a multi-level enterprise tag system is constructed, covering 12 major categories and 86 sub-categories, such as: basic attribute tags (registered location: Chongqing High-tech Zone; years of establishment: 8 years; industry: software and information technology services), operational capability tags (revenue scale: 62 million; R&D investment ratio: 8.5%), R&D and innovation tags (invention patents: 3; software copyrights: 12), qualification and honor tags (high-tech enterprise, specialized and innovative enterprise), credit compliance tags (no dishonesty), and application history tags (100% success rate). A daily data synchronization cycle is set to realize real-time iteration of the digital profile of enterprise A, ensuring that the profile is completely matched with the actual situation of enterprise A.

[0045] Step 3: Execution of the three-level matching engine First-level tag initial screening: The multi-level matching and inference engine extracts all the hard access tags of the policies and performs precise matching with the profile tags of Company A. Policies that do not meet the requirements, such as those registered in Chongqing but not in the same industry, those that have not been established for a long time, and those that are not high-tech enterprises, are filtered out. 112 policies that pass the initial screening are selected from 1286 policies. The second level of semantic deep matching involves encoding 112 policies into policy semantic vectors using a semantic encoding model and storing them in a vector database. Company A's full-dimensional profile information is also encoded into a company semantic vector, and the cosine similarity between the two is calculated. Simultaneously, each policy's application clause is assessed for suitability. For example, regarding the clause "the previous year's R&D expenses should account for no less than 7% of sales revenue" in the "Chongqing Key Software Enterprise Recognition Policy," Company A's R&D investment ratio is 8.5%, meeting the requirement and earning a high score. Weighted scores are assigned based on the importance of the clauses (e.g., 0.5 for hard indicators, 0.3 for recommended indicators, and 0.2 for bonus indicators), and the overall suitability score is calculated. A suitability threshold of 60 points is set, and 38 highly suitable policies are selected.

[0046] Level 3 Compliance Verification and Conflict Detection: Compliance verification was performed on 38 policies to confirm that Company A had no circumstances prohibiting its application; at the same time, based on the mutual exclusion relationship in the policy knowledge graph, it was found that 2 policies had duplicate support restrictions in the same funding pool as the policies already approved for Company A, and these were removed, resulting in a final list of 36 precisely matched policies.

[0047] Step 4: Application Assistance and Navigation Generation: The system generates a personalized policy matching list for Company A, sorted from highest to lowest suitability. Each policy is marked with a deadline, core reward standards, eligible items for the company, items to be improved, and application priority. For the highest priority policy, "Chongqing Key Software Enterprise Recognition," the application assistance and navigation module calls the material generation model and automatically breaks down the application process into 6 core nodes: material preparation (e.g., business license, audit report, R&D ledger, intellectual property certificate, etc.), online submission, acceptance and review, expert review, public announcement, and fund disbursement. Each node provides specific operation guidance and material templates. At the same time, based on Company A's profile information, the system automatically generates a draft of the project application and proposes optimization suggestions such as "supplementing software product sales data for the past two years to enhance competitiveness."

[0048] Step 5: Full-Process Monitoring and Review: The system connects to the Chongqing Municipal Commission of Economy and Information Technology's policy application acceptance platform to monitor the entire process of the application submitted by Company A, providing real-time updates on the status of acceptance, review, and public announcement stages. Simultaneously, the system continuously monitors the policy database. When a new policy for the "Chongqing Digital Economy Industry Development Special Fund" is released, it automatically matches and pushes an early warning based on Company A's profile, prompting the company to apply promptly. After the application period ends, based on Company A's application results (assuming successful approval) and the entire process data, the review analysis module outputs the following: The key factors for this approval are the R&D investment ratio and the high-tech enterprise qualification. It is recommended to focus on accumulating intellectual property rights to match higher-level policy optimization suggestions. This review result is also used to update the application history tags in the company profile and fine-tune the weight of the R&D investment ratio indicator in the matching model to further improve subsequent matching accuracy.

[0049] In this embodiment, the method and system of the present invention provide intelligent matching and application assistance services for industrial policies to enterprise A, achieving an accuracy rate of 96.2% in policy element extraction and 93.5% in policy matching. The preparation time for enterprise application materials is shortened from the traditional 5 working days to 1 working day, and the efficiency of application preparation is improved by 80%. At the same time, through precise matching and full-process application assistance, enterprise A successfully obtained the "Chongqing Key Software Enterprise Certification" and successfully applied for the "Chongqing Digital Economy Industry Development Special Fund" policy. The success rate of application is 65% higher than that of the traditional model, which fully verifies the practicality and effectiveness of the present invention.

[0050] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for intelligent matching and application assistance of industrial policies based on a large model, characterized in that, Includes the following steps: S1. Deep analysis of policy documents and construction of structured knowledge base based on NLP big model: Automated collection and preprocessing of multi-source policy documents, extraction of core structured elements of policies through domain-fine-tuned NLP big model, and construction of policy knowledge graph and dynamically updated policy knowledge base; S2. Dynamic Construction of Multi-Dimensional Digital Profiles of Enterprises: Under the premise of enterprise compliance and authorization, integrate multi-source enterprise data, construct a multi-level enterprise tag system, and generate and dynamically update enterprise digital profiles. S3, Multi-level Matching and Inference Engine: Precise matching of policies and enterprises: Through a three-level progressive matching process, precise matching of policies and enterprises is achieved, and compliant and highly adaptable policy results are output. S4. Personalized Policy Matching List and Application Navigation Path Generation: Based on the output of step S3, a policy matching list exclusive to the enterprise is generated, the entire application process is broken down and a visual navigation path is generated, and application materials are intelligently assisted in being generated and optimized. S5. Dynamic monitoring and follow-up of the entire application process: Real-time monitoring and early warning of the progress of enterprise application projects and policy dynamics, and completion of review analysis and model optimization based on enterprise application data and results.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: By connecting with official channels and using distributed crawlers, we can achieve full automated collection of policy documents and preprocess the collected documents to form a standardized original policy text library. The preprocessing includes format cleaning, text extraction, deduplication and noise reduction, and version verification. Based on the policy domain annotated corpus, a general NLP model is supervised and fine-tuned to construct a dedicated model for extracting policy elements. This model performs end-to-end deep analysis of unstructured policy texts and automatically extracts the core structured elements of the policy. These core structured elements include at least the policy basic information, applicable objects, application conditions, support standards, process rules, and restrictive clauses. The extracted core structured elements of the policy are linked by entities, semantically normalized, and relationally modeled to construct a policy knowledge graph. The policy knowledge graph defines at least the entity types of "policy-issuing department-industry-qualification-intellectual property" and the association relationships of "issuance, application, requirement, reward, and mutual exclusion". Establish a dynamic knowledge base update mechanism to synchronize the release, revision, extension, and expiration status of policies in real time.

3. The method according to claim 1, characterized in that, Step S2 specifically includes: Under the premise of enterprise compliance authorization, collect basic business registration information, operating and financial data, R&D and innovation data, qualification and honor information, credit compliance information and historical policy application information of enterprises; By using NLP large-scale models to perform semantic understanding and standardization processing on the collected data, a multi-level enterprise tag system is constructed. The multi-level enterprise tag system at least covers basic attribute tags, business capability tags, R&D innovation tags, qualification and honor tags, credit compliance tags, and application history tags. Set a data synchronization cycle for enterprises to synchronize dynamic information in real time, and realize the real-time dynamic update of the enterprise's digital profile. The dynamic information includes at least changes in enterprise registration, addition of intellectual property rights, updates to operating data, and acquisition of qualifications.

4. The method according to claim 1, characterized in that, The three-level progressive matching process in step S3 specifically includes: Level 1, Rapid Tag Initial Screening and Matching: Extract hard access tags that are vetoed by policies, accurately match them with enterprise profile tags, filter out policies that do not meet the requirements, and retain the set of policies that pass the initial screening; Level 2, Large-Scale Semantic Deep Matching and Fit Scoring: The policy clauses and enterprise profile information are encoded into policy semantic vectors and enterprise semantic vectors respectively through the NLP large-scale model. The cosine similarity between the policy semantic vectors and enterprise semantic vectors is calculated. At the same time, the zero-shot reasoning capability of the NLP large-scale model is used to judge the fit of each policy clause. The overall fit score is calculated based on the weighted weights set according to the importance of the clauses, and the policy set with high fit is selected. Level 3, Compliance Verification and Conflict Detection: Checks whether the enterprise has any circumstances that are explicitly prohibited from being declared by policy documents. Based on the policy knowledge graph, it completes mutual exclusion detection between policies to be declared and duplicate declaration conflict detection between policies to be declared and previously approved policies, and outputs the final accurate matching result of compliance.

5. The method according to claim 1, characterized in that, Step S4 specifically includes: Based on the output of step S3, a personalized policy matching list is generated, which includes basic policy information, suitability score, core reward standards, enterprise compliance items, items to be improved, and application priority suggestions. For the policies in the list, the application process nodes are broken down based on the NLP big data model, and a visual application navigation path is generated, which clarifies the operation requirements, material list and precautions for each node. The nodes include material preparation, online application, acceptance and review, expert review, public announcement and fund disbursement. For core application materials, the system utilizes a large NLP model combined with enterprise profiles and policy application requirements to automatically generate initial drafts of the application materials and provides suggestions for content optimization, logical improvement, and compliance verification. The core application materials include at least the project application form, feasibility study report, and a list of qualification certificates.

6. The method according to claim 1, characterized in that, Step S5 specifically includes: Connect to the policy application and acceptance platform to track the progress of the application projects submitted by enterprises throughout the entire process, synchronize the node status in real time, and push progress reminders and node warnings. The node status includes at least acceptance, review, evaluation, public announcement and disbursement. Real-time monitoring of new policies, revision announcements, application extensions, and window period adjustments in the policy database; automatic matching of appropriate policy dynamics based on enterprise profiles; and push of early warning information to enterprises. Based on the historical data and results of enterprise applications, a retrospective analysis is conducted using a large NLP model to output an analysis of application success rate, the core reasons for non-approval, and suggestions for subsequent application optimization. Furthermore, the weight parameters and semantic encoding capabilities of the matching model are continuously optimized based on the application result data.

7. A large-scale model-based intelligent matching and application assistance system for industrial policies, characterized in that, It includes a data layer, a model layer, a core functionality layer, and an application layer; The data layer is used to store, manage, and retrieve all system data. The data layer includes a policy text library, a multi-source enterprise database, a policy knowledge graph, a vector database, and a historical application database. The model layer is used to provide the core AI capabilities of the system. The model layer includes a large model for policy element extraction, a semantic encoding model, a multi-level matching reasoning model, and a large model for material generation. The core functional layer is used to implement the entire process steps of the method described in any of claims 1-6. The core functional layer includes a policy collection and analysis module, an enterprise profile construction module, a multi-level matching and reasoning engine, an application assistance and navigation module, and an application progress monitoring module. The application layer is used to provide visual interactive interfaces for different user groups. The application layer includes enterprise user terminals, government management terminals, and operation management terminals.

8. The system according to claim 7, characterized in that, In the core functional layer: The policy collection and parsing module is used for the automated collection and preprocessing of policy documents from multiple sources, extraction of policy elements, construction of policy knowledge graphs, and dynamic updating and maintenance of policy knowledge bases. The enterprise profile building module is used for compliant collection and standardized processing of enterprise authorized data, systematic management of enterprise tags, construction and dynamic updating of multi-dimensional digital profiles of enterprises, and enterprise data security and access management. The multi-level matching and inference engine has built-in tag screening unit, semantic deep matching unit, and compliance verification and conflict detection unit to fully execute the three-level progressive matching process; The application assistance and navigation module is used to generate personalized policy matching lists, break down and visualize the entire application process navigation path, intelligently generate and optimize application materials, provide application milestone reminders, and answer frequently asked questions. The application progress monitoring module is used for tracking the entire process of enterprise application projects, real-time monitoring and early warning of policy dynamics, review and analysis of application data, iterative optimization of matching models, and statistical analysis of policy implementation effects.

Images