A method and system for intelligent matching of policy resources

Abstract

This application relates to the field of intelligent matching technology for policy resources, and in particular to a method and system for intelligent matching of policy resources. The method includes data collection, matching strategy identification, visual hierarchical feedback, and strategy adjustment and optimization. Data collection is used to collect attribute data of applicants and text data of policy clauses. Matching strategy identification is used to generate optimal target matching strategy data for the attribute data of applicants and text data of policy clauses. This invention achieves efficient and accurate matching between applicants and policy resources by introducing technologies such as BERT embedding model, graph neural network algorithm module, particle swarm search algorithm module, and natural language processing interface. It effectively solves the contradiction between the dynamism of policies and the static nature of systems, the contradiction between semantic complexity and the simplicity of human-computer interaction, and the contradiction between the accuracy of resource supply and service efficiency, significantly improving the scientificity and practicality of intelligent matching of policy resources.

Description

Technical Field

[0001] This application relates to the field of intelligent matching technology for policy resources, and in particular to a method and system for intelligent matching of policy resources. Background Technology

[0002] With the increasing complexity and diversification of policy resource management, achieving intelligent matching between applicants and policy resources has become a key issue in improving government efficiency and service quality. Currently, intelligent policy resource matching systems mainly rely on the following three technical solutions:

[0003] The first type of solution uses a static matching method based on a rule engine, which compares the policy terms with the conditions of the applicant by manually constructing IF-THEN logical rules. This method has obvious drawbacks: First, policy terms are updated frequently (such as tax incentive policies with a high annual revision rate), and the lag in manually maintaining the rule base leads to poor timeliness of matching results. Second, in scenarios with multiple combinations of conditions (such as simultaneously meeting "high-tech enterprise qualification + R&D expense deduction"), logical conflicts between rules are easy to increase, which seriously affects the accuracy of matching. Third, cross-regional policy differences (such as the subsidy standard gradient between the eastern coastal areas and the western regions) require repeated development of localized rules, which results in high maintenance costs.

[0004] The second type of approach is based on keyword matching and text similarity calculation, using algorithms such as TF-IDF and BM25 to achieve rapid retrieval of policy texts and demand descriptions. The core problem with this type of approach is that the lack of semantic understanding leads to misjudgment of synonyms, such as treating "technology-based SMEs" and "innovative micro-enterprises" as equivalent. At the same time, it ignores the logical connection between policy elements, affecting the substantive consistency between the matching results and the original policy text.

[0005] The third approach attempts to introduce a single machine learning model, using algorithms such as decision trees and random forests for classification and prediction. While this method improves automation, it suffers from three major technical bottlenecks: First, it relies heavily on feature engineering, requiring manual labeling of policy tags, which leads to a surge in labeling costs when dealing with a policy database of millions of entries. Second, it suffers from a significant cold start problem, requiring 3-6 months of training data accumulation after a new policy is released to achieve effective matching accuracy. Third, the model has poor interpretability, making it impossible to explain the matching criteria to users in government information disclosure scenarios.

[0006] The aforementioned shortcomings lead to three contradictions in the practical application of existing systems: the contradiction between the dynamism of policies and the static nature of the system, the contradiction between semantic complexity and the simplicity of human-computer interaction, and the contradiction between the accuracy of resource supply and service efficiency. Furthermore, the existing system lacks a dynamic feedback mechanism, making it unable to optimize matching strategies based on user submissions, resulting in a long-term decline in matching accuracy.

[0007] Therefore, there is an urgent need to propose a method and system for intelligent matching of policy resources that can effectively address the above-mentioned shortcomings. Summary of the Invention

[0008] To address the shortcomings of existing intelligent policy resource matching systems in terms of policy dynamism, semantic complexity, and accuracy of resource supply, this invention provides an intelligent policy resource matching method and system. By introducing technologies such as the BERT embedding model, graph neural network algorithm module, particle swarm search algorithm module, and natural language processing interface, it achieves efficient and accurate matching between applicants and policy resources. This effectively resolves the contradictions between policy dynamism and system statics, semantic complexity and human-computer interaction simplicity, and accuracy of resource supply and service efficiency, significantly improving the scientific rigor and practicality of intelligent policy resource matching.

[0009] The objective of this invention can be achieved through the following technical solutions:

[0010] In a first aspect, the present invention provides a method for intelligent matching of policy resources, comprising the following steps:

[0011] S1. Data Collection: Collect data on the attributes of the applicant and the text of policy clauses;

[0012] S2. Matching strategy identification: The particle swarm search algorithm is used to identify specific matching strategies based on the fitness score and matching probability, and generate target matching strategy data.

[0013] The BERT embedding model is used to calculate the semantic similarity between the applicant's attribute data and the policy constraint keywords to generate a suitability score; the graph neural network algorithm is used to match the applicant's attribute data and the policy clause text data to obtain the matching probability.

[0014] S3. Visualized Layered Feedback: Based on the applicant's attribute data, policy clause text data, and target matching strategy data, the matching results are executed and displayed in a layered visual interface.

[0015] S4. Strategy Adjustment and Optimization: Collect user feedback and generate strategy adjustment judgment data based on the matching strategy adjustment instructions. If the user feedback indicates that no adjustment is needed, the current policy matching operation will end directly. If the user feedback indicates that adjustment is needed, perform a strategy optimization operation on the target matching strategy data and generate optimized matching strategy data.

[0016] As a preferred embodiment of the present invention, the data acquisition includes the following steps:

[0017] Enterprise attribute information can be entered online through the government service platform to generate declaration entity attribute data; the declaration entity attribute data includes enterprise type, industry sector, registered address, registered capital, number of employees, qualification certificates, R&D investment ratio and credit rating;

[0018] The policy clause text information is entered online through the government service platform and policy clause text data is generated; the policy clause text information includes the policy name, scope of application, and support direction.

[0019] As a preferred embodiment of the present invention, the generation of the adaptability score includes the following steps:

[0020] A set of policy constraint keywords is constructed based on the attribute data of the applicant; the policy constraint keywords include blacklisted enterprises, dishonest persons subject to enforcement, parties involved in major tax violations, environmental protection violation records, work safety accident records, and major legal proceedings;

[0021] The BERT embedding model is used to calculate the semantic similarity between the applicant's attribute data and the policy constraint keywords, and an adaptation score is generated.

[0022] As a preferred embodiment of the present invention, the matching probability includes the following steps:

[0023] Construct a policy classification tag data set; the policy classification tag data includes categories such as technological innovation, energy conservation and emission reduction, rural revitalization, employment and entrepreneurship, and financial services;

[0024] The system performs attribute feature matching based on the applicant's attribute data and the policy classification label data set. It uses a graph neural network algorithm to model the multidimensional features of the applicant's attribute data and combines it with the policy clause text data to obtain the matching probability between the applicant's attribute data and the policy classification label data.

[0025] As a preferred embodiment of the present invention, the particle swarm search algorithm includes:

[0026] Parameter settings: Set the number of particles, number of iterations, and initialize the inertia weight. acceleration constant and ;

[0027] Particle swarm initialization: Each particle represents a complete target matching strategy data set. A feasible region for the target matching strategy data is defined. The initial position of each particle is randomly generated within this feasible region, and its initial velocity is 0 or a small range of random values. Individual optimality is achieved. The initial position of the particle is the global optimum. The position of the particle with the highest fitness;

[0028] Fitness function and iteration function design: The fitness function, particle velocity iteration function, and position iteration function are defined by combining the fitness score and matching probability. If the fitness of the current particle position is better than the historical value, then the fitness function is updated. If the current particle position is better than all other particle positions, then update. ;

[0029] The fitness function and iteration function include:

[0030] Fitness function calculation formula:

[0031] ;

[0032] Formula for calculating particle velocity iteration function:

[0033] ;

[0034] Formula for calculating particle position iteration function:

[0035] ;

[0036] in, and These are the weighting coefficients, and =1, For target matching strategy, For the suitability score, if the semantic similarity between the applicant's attribute data and the policy constraint keywords exceeds a preset threshold, then... If the value is 0, and it does not exceed the preset threshold, then... =1, This represents the matching probability between the applicant attribute data output by the graph neural network and the policy classification label data. and for Random numbers between these ranges increase the randomness of the search. and The first The particle in the first Second and third The speed of the next iteration and The first The particle in the first Second and third The position of the next iteration;

[0037] Termination condition definition: The maximum number of iterations is reached or the global optimum is less than the threshold for N consecutive iterations;

[0038] Output target matching strategy data;

[0039] The particle swarm search algorithm is divided into three parts: a global exploration phase, a local optimization phase, and a convergence phase.

[0040] The global exploration phase is used to quickly cover the search space and avoid getting stuck in local optima; the local optimization phase is used to focus on high-potential areas and improve the accuracy of target matching strategy data; the convergence phase is used to stabilize the optimal target matching strategy data and prevent oscillations.

[0041] As a preferred embodiment of the present invention, the visualized hierarchical feedback includes the following steps:

[0042] The data on the attributes of the applicant, the text of the policy clauses, and the target matching strategy are transmitted to the terminal device through a cloud computing network, and are displayed in an orderly and hierarchical manner through a visual interface.

[0043] A matching strategy adjustment selection button module is generated synchronously on the visual interface; the matching strategy adjustment selection button module includes a matching strategy adjustment selection button module and a matching strategy no-adjustment selection button module.

[0044] As a preferred embodiment of the present invention, the strategy adjustment and optimization includes the following steps:

[0045] The matching strategy adjustment selection button module in the visual interface collects user feedback on the target matching strategy data online and generates user feedback data, which includes instructions that need adjustment and instructions that do not need adjustment.

[0046] The system processes user feedback data by identifying keywords related to matching strategy adjustment instructions and generates strategy adjustment judgment data. When user feedback indicates an instruction requiring adjustment, it means the user needs to adjust the matching rules and operation steps in the target matching strategy data, and the output strategy adjustment judgment data indicates adjustment is required. When user feedback indicates an instruction not requiring adjustment, it means the user does not need to adjust the matching rules and operation steps in the target matching strategy data, and the output strategy adjustment judgment data indicates no adjustment is required. In this case, the current policy matching operation ends directly.

[0047] As a preferred embodiment of the present invention, the strategy adjustment and optimization further includes the following steps:

[0048] When the strategy adjustment judgment data indicates that adjustment is needed, the user can re-enter the matching strategy adjustment text information for the target matching strategy data online through the natural language processing interface, and generate optimized matching strategy data.

[0049] Secondly, the present invention also provides a policy resource intelligent matching system that executes the policy resource intelligent matching method described above. The system includes a declaration subject attribute collection module, a matching strategy generation module, a matching result feedback module, and a strategy optimization module.

[0050] The application entity attribute collection module is used to input enterprise attribute information and policy clause text information online to generate application entity attribute data; the application entity attribute collection module includes an application entity information input unit, a policy clause information input unit, a policy constraint storage unit, and a policy suitability judgment unit;

[0051] The matching strategy generation module is used to identify specific matching strategies based on the applicant's attribute data and policy clause text data, and generate target matching strategy data after data identification.

[0052] The matching result feedback module executes the matching results based on the applicant's attribute data, policy clause text data, and target matching strategy data, and displays the results in a layered visual interface. The matching result feedback module includes a matching result feedback unit, a user feedback collection unit, a strategy adjustment judgment unit, and a matching strategy optimization unit.

[0053] The strategy optimization module is used to perform strategy optimization on the target matching strategy data and generate optimized matching strategy data when user feedback indicates that adjustments are needed.

[0054] As a preferred embodiment of the present invention, the applicant attribute collection module includes an applicant information input unit, a policy clause information input unit, a policy constraint storage unit, and a policy suitability judgment unit; the matching result feedback module includes a matching result feedback unit, a user feedback collection unit, a strategy adjustment judgment unit, and a matching strategy optimization unit.

[0055] This invention provides a method and system for intelligent matching of policy resources, which has the following beneficial effects:

[0056] 1. Collect applicant attribute information and policy clause text information online through the government service platform to provide data support for accurate matching of target policies; use the BERT embedding model to calculate the semantic similarity between applicant attribute data and policy constraint keywords, use the graph neural network algorithm to match applicant attribute data and policy clause text data, and use the particle swarm search algorithm to obtain target matching strategy data to achieve accurate matching of applicant attributes and policy categories, thereby improving the applicability and efficiency of policy matching;

[0057] 2. The system provides intuitive feedback on the matching results of the applicant's attribute information, policy category information, and target matching strategy information through a visual interface; it also collects user feedback based on the visual interface and accurately identifies user opinions on adjusting the matching strategy, thereby improving the rationality and scientific nature of the generated matching strategy; at the same time, users can re-enter the text information on adjusting the matching strategy data for the target matching strategy online through a natural language processing interface, thereby improving the quality and convenience of policy matching. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 A flowchart illustrating a policy resource intelligent matching method provided in an embodiment of the present invention;

[0060] Figure 2 A schematic diagram illustrating the process of generating and optimizing target matching strategy data according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of the structure of a policy resource intelligent matching system provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of this application.

[0063] The preferred embodiments of the present invention will be further described below with reference to the accompanying drawings;

[0064] Example 1

[0065] Please see Figure 1 This embodiment provides a method for intelligent matching of policy resources, including:

[0066] In step S1, the applicant enters enterprise attribute information online through the government service platform to generate applicant attribute data, and simultaneously enters policy clause text information to generate policy clause text data;

[0067] The applicant's attribute information includes specific details such as enterprise type, industry sector, registered address, registered capital, number of employees, qualification certificates, R&D investment ratio, and credit rating; the policy clause text information includes policy name, scope of application, and support direction; the government service platform, as the core entry point for data collection, can ensure the integrity and timeliness of the data.

[0068] Step S2 employs a particle swarm search algorithm to identify specific matching strategies based on the applicant's attribute data and policy clause text data, generating target matching strategy data; including:

[0069] The particle swarm search algorithm includes:

[0070] Parameter settings: Set the number of particles, number of iterations, and initialize the inertia weight. acceleration constant and ;

[0071] Particle swarm initialization: Each particle represents a complete target matching strategy data set. A feasible region for the target matching strategy data is defined. The initial position of each particle is randomly generated within this feasible region, and its initial velocity is 0 or a small range of random values. Individual optimality is achieved. The initial position of the particle is the global optimum. The position of the particle with the highest fitness;

[0072] Based on the attribute data of the applicant, a set of policy constraint keywords is constructed. The policy constraint keywords include keywords such as blacklisted enterprises, dishonest persons subject to enforcement, parties involved in major tax violations, environmental protection violation records, work safety accident records, and major legal proceedings. Then, the BERT embedding model is used to calculate the semantic similarity between the attribute data of the applicant and the policy constraint keywords in the set of policy constraint keywords to generate a suitability score.

[0073] A policy classification label dataset is constructed; the policy classification label dataset includes categories such as technological innovation, energy conservation and emission reduction, rural revitalization, employment and entrepreneurship, and financial services; attribute feature matching is performed based on the applicant's attribute data and the policy classification label dataset, a graph neural network algorithm is used to model the multidimensional features of the applicant's attribute data, and the matching probability between the applicant's attribute data and the policy classification label data is obtained by combining the topological structure of the policy classification label dataset;

[0074] Fitness function and iteration function design: The fitness function, particle velocity iteration function, and position iteration function are defined by combining the fitness score and matching probability. If the fitness of the current particle position is better than the historical value, then the fitness function is updated. If the current particle position is better than all other particle positions, then update. ;

[0075] The fitness function and iteration function include:

[0076] Fitness function calculation formula:

[0077] ;

[0078] Formula for calculating particle velocity iteration function:

[0079] ;

[0080] Formula for calculating particle position iteration function:

[0081] ;

[0082] in, and These are the weighting coefficients, and =1, For target matching strategy, For the suitability score, if the semantic similarity between the applicant's attribute data and the policy constraint keywords exceeds a preset threshold, then... If the value is 0, and it does not exceed the preset threshold, then... =1, This represents the matching probability between the applicant attribute data output by the graph neural network and the policy classification label data. and for Random numbers between these ranges increase the randomness of the search. and The first The particle in the first Second and third The speed of the next iteration and The first The particle in the first Second and third The position of the next iteration;

[0083] Termination condition definition: The maximum number of iterations is reached or the global optimum is less than the threshold for N consecutive iterations;

[0084] Output target matching strategy data;

[0085] The particle swarm search algorithm is divided into three parts: a global exploration phase, a local optimization phase, and a convergence phase.

[0086] During the global exploration phase, the inertia weight is initialized to 0.9, and the acceleration constant is... and All values ​​are set to 2 to quickly cover the search space and avoid getting trapped in local optima; during the local optimization phase, the inertia weight is reduced to 0.4, which is the speedup constant. Set it to 2.5. The inertia weight is set to 1.5 to focus on high-potential areas and improve the accuracy of target matching strategy data. During the convergence phase, the inertia weight is fixed at 0.3, allowing only small perturbations. If the fitness does not improve after 10 consecutive iterations, the search is terminated early. This phase is used to stabilize the optimal target matching strategy data and prevent oscillations.

[0087] Step S3 performs a matching result feedback operation based on the applicant attribute data, policy clause text data, and target matching strategy data; including:

[0088] The data on the applicant's attributes, policy clauses, and target matching strategies are transmitted to the terminal device via a cloud computing network and displayed in a visual interface in an orderly, layered manner from left to right. The applicant's attribute data is located on the left side of the interface, the policy clauses are centered, and the target matching strategy data is on the right side. In addition, matching strategy adjustment selection button modules are set at the bottom left and right sides of the visual interface. These modules include buttons for selecting whether the matching strategy needs to be adjusted and buttons for selecting whether the matching strategy does not need to be adjusted. Users can intuitively view the matching results through this interface and decide whether to adjust the matching strategy according to their own needs.

[0089] Step S4 collects user feedback data and performs matching strategy adjustment judgment processing to generate strategy adjustment judgment data; including:

[0090] The matching strategy adjustment selection button module in the visual interface collects user feedback on the matching results online; when a user clicks the matching strategy adjustment selection button module, user feedback data is generated as an instruction to adjust; when a user clicks the matching strategy no longer needs to be adjusted selection button module, user feedback data is generated as an instruction not to adjust.

[0091] If the user feedback indicates a need for adjustment, the output strategy adjustment judgment data will show that adjustment is required. Simultaneously, the user can re-enter the matching strategy adjustment text information online via a natural language processing interface, generating optimized matching strategy data. This natural language processing interface supports either GPT-4 or PaLM natural language processing technology, accurately understanding the user's adjustment intentions and generating corresponding optimized matching strategy data. If the user feedback indicates a need for adjustment, the output strategy adjustment judgment data will show that adjustment is not required, and the current policy matching operation will end directly.

[0092] Example 2

[0093] Please see Figure 3 This embodiment provides a policy resource intelligent matching system that executes the policy resource intelligent matching method described above. The system includes an application subject attribute collection module, a matching strategy generation module, a matching result feedback module, and a strategy optimization module.

[0094] The application entity attribute collection module is used to input enterprise attribute information and policy clause text information online to generate application entity attribute data; the application entity attribute collection module includes an application entity information input unit, a policy clause information input unit, a policy constraint storage unit, and a policy suitability judgment unit;

[0095] The matching strategy generation module is used to identify specific matching strategies based on the applicant's attribute data and policy clause text data, and generate target matching strategy data after data identification.

[0096] The matching result feedback module executes the matching results based on the applicant's attribute data, policy clause text data, and target matching strategy data, and displays the results in a layered visual interface. The matching result feedback module includes a matching result feedback unit, a user feedback collection unit, a strategy adjustment judgment unit, and a matching strategy optimization unit.

[0097] The strategy optimization module is used to perform strategy optimization on the target matching strategy data and generate optimized matching strategy data when user feedback indicates that adjustments are needed.

[0098] Example 3

[0099] Please see Figure 1-2 To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario:

[0100] During the deployment of a policy resource intelligent matching system in a certain province, the government service platform served as the core data collection portal for applicants and policy clauses. First, a high-tech enterprise entered its attribute information online through the platform, including its enterprise type as "high-tech enterprise," industry sector as "information technology," registered location as "City A of a certain province," employee size as "500 people," and R&D investment ratio as "8%." Simultaneously, the enterprise also uploaded relevant policy clause texts, such as a certain R&D expense subsidy policy for high-tech enterprises. This data was stored separately through the applicant information entry unit and the policy clause information entry unit of the government service platform and then transmitted to the subsequent processing module.

[0101] Subsequently, in the matching strategy identification stage, the policy suitability judgment unit constructed a set of policy constraint keywords, including keywords such as "blacklisted enterprises," "dishonest judgment debtors," and "parties involved in major tax evasion cases." Using a BERT embedding model, semantic similarity was calculated between the applicant's attribute data and the keywords in the policy constraint keyword set, generating a suitability score. The policy category label storage unit pre-stored policy category label data for categories such as technological innovation, energy conservation and emission reduction, and rural revitalization. The policy category analysis unit, by calling the graph neural network algorithm module, performed multi-dimensional feature modeling and topological structure matching between the applicant's attribute data and the label data in the policy category label data set to obtain the matching probability.

[0102] Next, based on the fitness score and matching probability, the fitness function, particle velocity iteration function, and position iteration function of the particle swarm search algorithm are defined, and target matching strategy data is generated. In the global exploration phase, the particle swarm search algorithm performs a global fast coverage search to avoid getting trapped in local optima. In the local optimization phase, it focuses on high-potential regions to improve the search accuracy of the target matching strategy data. In the convergence phase, it is used to stabilize the optimal target matching strategy data and prevent oscillations. Thus, the globally optimal target matching strategy data is accurately generated.

[0103] During the visualized hierarchical feedback phase, the matching result feedback unit transmits the applicant's attribute data, policy clause text data, and target matching strategy data to the terminal device via a cloud computing network and displays them in a visual interface. For example, the left side of the interface displays the company's attribute information, the middle displays the "science and technology innovation" policy category it matches, and the right side displays the specific operational details of the R&D expense subsidy policy. Users can decide whether to adjust the matching strategy through the matching strategy adjustment selection button module at the bottom of the interface. This hierarchical display method is intuitive and clear, making it easy for users to quickly understand the matching results.

[0104] During the strategy adjustment and optimization phase, the user feedback collection unit collects user feedback online through the matching strategy adjustment selection button module on the visual interface. For example, if a user believes that certain policy clauses in the matching results do not meet actual needs, they can click the "Matching strategy needs adjustment" button, generating user feedback data as an instruction to adjust. If the user is satisfied with the matching results, they can click the "Matching strategy does not need adjustment" button, generating user feedback data as an instruction not to adjust. The strategy adjustment judgment unit generates strategy adjustment judgment data based on the user feedback data. If no adjustment is needed, the matching operation ends directly.

[0105] When the strategy adjustment judgment data indicates that adjustments are needed, the user re-enters the adjustment text information for the target matching strategy data online through the natural language processing interface; for example, the user may propose "increasing the additional subsidy policy for enterprises' R&D investment ratio to more than 10%", the natural language processing interface uses technologies such as GPT-4 or PaLM to accurately understand the user's adjustment intention and generate optimized matching strategy data; the strategy optimization unit receives the optimized matching strategy data and updates it to the system for use in subsequent matching operations.

[0106] Through the implementation of the above specific application scenarios, this invention effectively solves the contradictions between policy dynamism and system staticity, semantic complexity and human-computer interaction simplicity, and the contradiction between the accuracy of resource supply and service efficiency, significantly improving the scientificity and practicality of intelligent matching of policy resources.

[0107] The above are merely preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent matching of policy resources, characterized in that: Includes the following steps: S1. Data Collection: Collect data on the attributes of the applicant and the text of policy clauses; S2. Matching strategy identification: The particle swarm search algorithm is used to identify specific matching strategies based on the fitness score and matching probability, and generate target matching strategy data. The BERT embedding model is used to calculate the semantic similarity between the applicant's attribute data and the policy constraint keywords to generate a suitability score; the graph neural network algorithm is used to match the applicant's attribute data and the policy clause text data to obtain the matching probability. S3. Visualized Layered Feedback: Based on the applicant's attribute data, policy clause text data, and target matching strategy data, the matching results are executed and displayed in a layered visual interface. S4. Strategy Adjustment and Optimization: Collect user feedback and generate strategy adjustment judgment data based on the matching strategy adjustment instructions. If the user feedback indicates that no adjustment is needed, the current policy matching operation will end directly. If the user feedback indicates that adjustment is needed, perform a strategy optimization operation on the target matching strategy data and generate optimized matching strategy data. The method of matching the applicant's attribute data and policy clause text data using a graph neural network algorithm includes the following steps: Construct a policy classification tag data set; the policy classification tag data includes categories such as technological innovation, energy conservation and emission reduction, rural revitalization, employment and entrepreneurship, and financial services; The attribute features of the applicant are matched with the policy classification label data. A graph neural network algorithm is used to model the multidimensional features of the applicant attribute data. Combined with the policy clause text data, the matching probability between the applicant attribute data and the policy classification label data is obtained. The particle swarm search algorithm includes: Parameter settings: Set the number of particles, number of iterations, and initialize the inertia weight. acceleration constant and ; Particle swarm initialization: Each particle represents a complete target matching strategy data set. A feasible region for the target matching strategy data is defined, and the initial positions of the particles are randomly generated within this feasible region, with individual optimality considered. The initial position of the particle is the global optimum. The position of the particle with the highest fitness; Fitness function and iteration function design: The fitness function, particle velocity iteration function, and position iteration function are defined by combining the fitness score and matching probability. If the fitness of the current particle position is better than the historical value, then the fitness function is updated. If the current particle position is better than all other particle positions, then update. ; The fitness function and iteration function include: Fitness function calculation formula: ； Formula for calculating particle velocity iteration function: ； Formula for calculating particle position iteration function: ； in, and These are the weighting coefficients, and , For target matching strategy, For the suitability score, if the semantic similarity between the applicant's attribute data and the policy constraint keywords exceeds a preset threshold, then... for If it does not exceed the preset threshold, then for , This represents the matching probability between the applicant attribute data output by the graph neural network and the policy classification label data. and for Random numbers between and The first The particle in the first Second and third The speed of the next iteration and The first The particle in the first Second and third The position of the next iteration; Termination condition definition: The maximum number of iterations is reached or the global optimum is less than the threshold for N consecutive iterations; Output target matching strategy data; The particle swarm search algorithm is divided into three parts: a global exploration phase, a local optimization phase, and a convergence phase. The global exploration phase is used to quickly cover the search space and avoid getting stuck in local optima; the local optimization phase is used to focus on high-potential areas and improve the accuracy of target matching strategy data; the convergence phase is used to stabilize the optimal target matching strategy data and prevent oscillations. The strategy adjustment and optimization also includes the following steps: When the strategy adjustment judgment data indicates that adjustment is needed, the user can re-enter the matching strategy adjustment text information for the target matching strategy data online through the natural language processing interface, and generate optimized matching strategy data. The data collection includes inputting enterprise attribute information online through the government service platform and generating application entity attribute data; the application entity attribute data includes enterprise type, industry sector, registered address, registered capital, employee size, qualification certificates, R&D investment ratio, and credit rating; and inputting policy clause text information online through the government service platform and generating policy clause text data; the policy clause text information includes policy name, scope of application, and support direction. The process of generating a suitability score involves constructing a set of policy constraint keywords based on the applicant's attribute data. These policy constraint keywords include blacklisted enterprises, dishonest judgment debtors, parties involved in major tax evasion cases, environmental violation records, work safety accident records, and major legal proceedings. A BERT embedding model is used to calculate the semantic similarity between the applicant's attribute data and the policy constraint keywords to generate a suitability score.

2. The policy resource intelligent matching method according to claim 1, characterized in that: The visualized hierarchical feedback includes the following steps: The data on the attributes of the applicant, the text of the policy clauses, and the target matching strategy are transmitted to the terminal device through a cloud computing network, and are displayed in an orderly and hierarchical manner through a visual interface. A matching strategy adjustment selection button module is generated synchronously on the visual interface; the matching strategy adjustment selection button module includes a matching strategy adjustment selection button module and a matching strategy no-adjustment selection button module.

3. The policy resource intelligent matching method according to claim 1, characterized in that: The strategy adjustment and optimization includes the following steps: The matching strategy adjustment selection button module in the visual interface collects user feedback on the target matching strategy data online and generates user feedback data, which includes instructions that need adjustment and instructions that do not need adjustment. The system processes user feedback data by identifying keywords related to matching strategy adjustment instructions and generates strategy adjustment judgment data. When user feedback indicates an instruction requiring adjustment, it means the user needs to adjust the matching rules and operation steps in the target matching strategy data, and the output strategy adjustment judgment data indicates adjustment is required. When user feedback indicates an instruction not requiring adjustment, it means the user does not need to adjust the matching rules and operation steps in the target matching strategy data, and the output strategy adjustment judgment data indicates no adjustment is required. In this case, the current policy matching operation ends directly.

4. A policy resource intelligent matching system, executing the policy resource intelligent matching method as described in any one of claims 1 to 3, characterized in that: The system includes a module for collecting the attributes of the applicant, a module for generating matching strategies, a module for providing matching results, and a module for optimizing strategies. The reporting entity attribute collection module is used to input enterprise attribute information and policy clause text information online and generate reporting entity attribute data; The matching strategy generation module is used to identify specific matching strategies based on the applicant's attribute data and policy clause text data, and generate target matching strategy data after data identification. The matching result feedback module performs matching based on the applicant's attribute data, policy clause text data, and target matching strategy data, and displays the results in a layered visual interface. The strategy optimization module, when user feedback indicates that adjustments are needed, performs strategy optimization on the target matching strategy data and generates optimized matching strategy data. The application subject attribute collection module includes an application subject information input unit, a policy clause information input unit, a policy constraint storage unit, and a policy suitability judgment unit; the matching result feedback module includes a matching result feedback unit, a user feedback collection unit, a strategy adjustment judgment unit, and a matching strategy optimization unit.

Images