System for intelligently matching government benefit policy for enterprise, matching method and medium

Abstract

The invention discloses a system for intelligently matching a government benefit policy for an enterprise, a matching method and a medium, which can comprehensively and efficiently obtain policy information, screen policies and calculate the matching degree based on the similarity between a policy semantic vector formed by the policy information and an enterprise portrait vector formed by the enterprise information, and improve the matching degree of the policy. And finally realizing accurate policy matching to enterprises.

Description

Technical Field

[0001] This invention belongs to the technical field of computer data artificial intelligence processing, specifically relating to a system, matching method, and medium for intelligently matching government preferential policies for enterprises. Background Technology

[0002] Governments at all levels typically release preferential policies for businesses through official websites, government information disclosure platforms, and WeChat official accounts. Businesses can access policy texts through keyword searches and categorized browsing. Some regions have established integrated government service platforms, providing centralized policy release and search services. Businesses can browse policy announcements, download policy documents, and perform application procedures to some extent on the same platform. Besides government platforms and information aggregation systems, many businesses also obtain policy information through accounting firms, legal service agencies, and park service centers, receiving policy interpretation and application guidance from professionals. Although existing government service platforms, policy information aggregation websites, and manual consultation services have alleviated the difficulty for businesses to obtain policy information to some extent, the following technical problems and shortcomings still exist: 1. Information is scattered and difficult to cover comprehensively. Policies are released through various channels at different levels and departments, and information is updated frequently. Enterprises need to search repeatedly on multiple websites or systems, which can easily lead to the omission of policies that are relevant to them.

[0003] 2. Policy texts are unstructured, making them difficult to understand. Most policy documents are presented in the form of announcements and official documents, with numerous clauses and complex expressions. There is a lack of a unified method for extracting structured information, making it difficult for enterprises to quickly understand and apply them.

[0004] 3. Lack of intelligent personalized matching: Existing platforms mainly rely on keyword search or category indexing, making it difficult to combine information such as the company's industry, size, qualifications, and development stage to conduct accurate policy matching and recommendations.

[0005] 4. Manual interpretation is costly and inefficient. Enterprises, especially small and medium-sized enterprises, often rely on third-party consulting agencies for policy interpretation and screening, which increases additional costs and reduces the efficiency and reach of policy benefits.

[0006] Therefore, based on the aforementioned problems existing in the process of matching enterprises with policies in the existing technology, this invention discloses a system, matching method, and medium for intelligent matching of enterprises with government preferential policies for enterprises. Summary of the Invention

[0007] This invention discloses a system, matching method, and medium for intelligent matching of government preferential policies to enterprises. It can comprehensively and efficiently acquire policy information, and based on the similarity between the policy semantic vector formed by the policy information and the enterprise profile vector formed by the enterprise information, it filters policies and calculates the matching degree, ultimately achieving accurate matching of policies to enterprises.

[0008] This invention is achieved through the following technical solution: A system for intelligently matching government preferential policies to enterprises includes a data crawling module, a policy filtering module, a large language model parsing module, an enterprise profile construction module, a similarity retrieval module, and an AI discriminant analysis module. The data crawling module collects policy information; the policy filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags; the large language model parsing module introduces a conditional action structure extraction mechanism to decompose candidate policy documents into structured policy content; the enterprise profile construction module forms enterprise profile vectors based on enterprise user information; the similarity retrieval module converts the structured policy content into policy semantic vectors suitable for semantic retrieval and calculates the similarity between the policy semantic vectors and the enterprise profile vectors to obtain a set of candidate policies; the AI ​​discriminant analysis module analyzes and compares each structured policy in the candidate policy set with the enterprise profile input into the large language model to determine the matching degree between the current policy and the enterprise profile, and pushes the corresponding policy to the enterprise based on the matching degree.

[0009] To better realize the present invention, the similarity retrieval module further includes a vectorization construction module, a vector retrieval module, and a candidate policy generation module. The vectorization construction module converts structured policy content into policy semantic vectors suitable for semantic detection and stores the policy semantic vectors in a vector database. The vector detection module is connected to the enterprise profile construction module and the vectorization construction module respectively. The vector detection module performs similarity retrieval on the policy semantic vectors and the enterprise profile vectors. The candidate policy generation module summarizes policies with similarity higher than the similarity threshold to form a candidate policy set.

[0010] To better realize this invention, the AI ​​discriminant analysis module further includes a matching module, a cross-validation module, and a logical consistency detection module. The matching module is used to compare the structured policy content of each policy in the candidate policy set with the attribute values ​​in the enterprise profile into a large language model, and output the matching degree between the current policy and the enterprise profile based on the comparison results. The cross-validation module is used to simultaneously mobilize at least two different large language models to participate in the comparison process between the structured policy content and the attribute values ​​in the enterprise profile. The logical consistency detection module verifies the policy-related condition information output by the large language model to obtain the final policy pushed to the enterprise.

[0011] To better realize the present invention, the AI ​​discrimination and analysis module further includes a risk analysis module, which is used to identify whether there are compliance risks between the conditional information involved in the push policy and the enterprise's historical data.

[0012] To better realize the present invention, a visualization output display module is further included. The visualization output display module is used to structure and organize the matching policies pushed to the enterprise by the AI ​​discrimination and analysis module, and to visually output the policy recommendation list and detailed report to the enterprise.

[0013] To better realize the present invention, the policy screening and filtering module further includes a keyword module and a rule module. The keyword module is used to identify and extract keywords in the policy information; the rule module filters the policy information based on syntactic rules and determines whether the filtered policy information is related to the theme of benefiting enterprises.

[0014] To better realize the present invention, the large language model parsing module further includes a condition module and an action module. The condition module is used to extract and describe the condition information in the candidate policy documents, and the action module is used to extract the enterprise support information related to the current policy.

[0015] To better realize the present invention, the large language model parsing module further includes a lifecycle management module, which is used to inject time-based relevant lifecycle fields into each candidate policy document.

[0016] A method for intelligently matching enterprises with government preferential policies includes the following steps: Step 1: Use the data crawling module to crawl publicly available policy information to form the original policy dataset; Step 2: The policy screening and filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags; Step 3: Input the candidate policy documents into the big language model, which will perform semantic parsing and automatic summarization of the text of the candidate policy documents to generate structured policy content that includes at least the policy name, applicable conditions, support methods, application process, and validity period. Step 4: Based on the conditional information, support methods for businesses, and relevant lifecycle fields in the structured policy content, the structured policy content is transformed into a policy semantic vector suitable for semantic retrieval through the similarity retrieval module; Step 5: Input the enterprise data into the enterprise profile building module. The enterprise profile building module performs semantic parsing and vectorization on the enterprise data to obtain the enterprise profile vector. Step 6: Calculate the similarity between the policy semantic vector and the enterprise profile vector, sort the policies from high to low similarity, and extract the top n policies to form a candidate policy set; Step 7: Input the candidate policy set and enterprise profile into the big language model. The big language model extracts the applicable conditions from the candidate policy set and the enterprise features from the enterprise profile, and compares and matches the applicable conditions with the enterprise features. Step 8: Sort the policies from high to low based on their matching degree, and extract the top m policies to push to enterprises.

[0017] A medium for intelligently matching enterprises with government preferential policies for enterprises, wherein the medium stores a computer program, and when the computer program is executed, it implements a matching method for intelligently matching enterprises with government preferential policies for enterprises.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) By introducing large language model parsing, context enhancement, structured extraction and vector retrieval technology, this invention realizes the automated acquisition, intelligent parsing and precise matching of government policies that benefit enterprises, effectively overcoming the defects of information dispersion, difficulty in understanding and lack of personalized matching in the existing technology; (2) The present invention can cover policy information at multiple levels from the State Council to the county-level government through the data crawling module, avoiding omissions caused by manual retrieval; (3) This invention uses a large language model to perform semantic parsing on unstructured policy texts, transforming them into unified fixed sentence patterns and structured data, which facilitates retrieval and subsequent analysis; it adopts vectorized storage and similarity retrieval, which greatly improves retrieval efficiency and matching accuracy; and combined with enterprise profiles, the large language model judges the applicability of policies, which can provide personalized recommendations for enterprises of different sizes, industries and stages. Attached Figure Description

[0019] Figure 1 A schematic diagram of the system architecture for intelligently matching government preferential policies for enterprises; Figure 2 A flowchart illustrating the process of intelligently matching enterprises with government preferential policies. Detailed Implementation

[0020] Example 1: This embodiment discloses a system for intelligently matching enterprises with government preferential policies, such as... Figure 1As shown, the system includes a data crawling module, a policy filtering module, a large language model parsing module, an enterprise profile building module, a similarity retrieval module, and an AI discriminant analysis module. The data crawling module collects policy information. The policy filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags. The large language model parsing module introduces a conditional action structure extraction mechanism to decompose candidate policy documents into structured policy content. The enterprise profile building module forms enterprise profile vectors based on enterprise user information. The similarity retrieval module converts the structured policy content into policy semantic vectors suitable for semantic retrieval and calculates the similarity between the policy semantic vectors and the enterprise profile vectors to obtain a set of candidate policies. The AI ​​discriminant analysis module analyzes and compares each structured policy in the candidate policy set with the enterprise profile input into the large language model to determine the matching degree between the current policy and the enterprise profile, and pushes the corresponding policy to the enterprise based on the matching degree.

[0021] The data crawling module is responsible for automatically collecting policy information released by official websites and government information disclosure platforms at all levels, including the State Council, provincial, municipal, and county governments. The collected content includes policy announcements, notices, official documents, interpretive articles, and policy Q&As. The system uses scheduled tasks to periodically access a pre-defined list of sites. For static pages, HTTP requests are prioritized for crawling; for pages with dynamic rendering or complex interactions, browser automation components simulate user access. All crawled webpage content is uniformly transcoded and cleaned before being stored as raw policy data in the policy raw dataset. To address the issue of diverse policy document sources and inconsistent formats, the data crawling module further integrates multi-source heterogeneous policy data fusion technology. For HTML text-based policies, the system directly extracts the main text and attachment links. For policy announcements in PDF, official document format, scanned copies, and image formats, the system first uses an optical character recognition engine to recognize text in the image area, then uses a layout analysis model to identify the title structure, paragraph levels, headers, footers, and table areas in the document, mapping the recognition results into a structured document representation. For policy documents containing tables, the system uses a table-based structured model to reconstruct the row and column relationships of the tables, breaking down information such as subsidy amounts, ratios, times, and industry categories into data fields that can be directly used by subsequent modules. To ensure the reliability and integrity of the crawled data, the data crawling module also implements policy document version tracking and text quality assessment functions. The system establishes a unique identifier for each policy. When a new version of a document with the same identifier is detected, a paragraph-level difference comparison algorithm is used to calculate the changes between the old and new documents, marking newly added, modified, and repealed clauses, and generating a corresponding policy change summary. The text quality assessment model performs garbled character detection, duplicate detection, OCR recognition error rate assessment, and text integrity judgment. Documents suspected of being incomplete or of poor quality are marked or removed, improving the accuracy of subsequent parsing and matching from the source. The data crawling module can also be implemented through government open application interfaces, RSS subscription interfaces, batch downloading of government open compressed packages, and synchronization with third-party policy databases to meet the deployment needs of different regions and system environments.

[0022] Furthermore, the system includes a visualization output module. This module is used to structure and organize the matching policies pushed to enterprises by the AI-based discriminative analysis module, and visually output a policy recommendation list and detailed report to the enterprises. The visualization output module structures the policy matching conclusions generated by the AI-based discriminative analysis module and outputs a policy recommendation list or detailed report to enterprise users. The list includes basic information for each recommended policy, a policy summary, applicable conditions, support methods, application path, recommendation level, reasons for inapplicability, and compliance risk warnings. The system can present this information in the form of web pages, downloadable documents, or data returned from APIs, according to enterprise preferences. For scenarios requiring in-depth analysis, the visualization output module can also graphically display the relationship between enterprises and matching policies, constructing a visual network of enterprise-policy-conditions-support methods, helping enterprise managers understand the policy space applicable to them from a global perspective. In scenarios serving government departments, the visualization output module can generate statistical reports based on enterprise query and matching records, used to analyze the demand and matching status of various preferential policies for enterprises in different regions, industries, and sizes, providing a basis for government decision-making in subsequent policy design and resource allocation optimization.

[0023] Through the collaborative work of the above modules, this invention forms a complete technical closed loop, from automatic collection of multi-source government data, multimodal parsing, structured extraction and vectorized database construction, to enterprise profile construction, vector retrieval, AI intelligent judgment and risk analysis, and finally to result output and visualization. It can efficiently and accurately match enterprises with applicable government preferential policies and realize a low-cost, scalable and implementable intelligent policy service solution.

[0024] A method for intelligently matching enterprises with government policies benefiting businesses, such as... Figure 2 As shown, it includes the following steps: Step 1: Use the data crawling module to crawl publicly available policy information to form the original policy dataset; Step 2: The policy screening and filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags; Step 3: Input the candidate policy documents into the big language model, which will perform semantic parsing and automatic summarization of the text of the candidate policy documents to generate structured policy content that includes at least the policy name, applicable conditions, support methods, application process, and validity period. Step 4: Based on the conditional information, support methods for businesses, and relevant lifecycle fields in the structured policy content, the structured policy content is transformed into a policy semantic vector suitable for semantic retrieval through the similarity retrieval module; Step 5: Input the enterprise data into the enterprise profile building module. The enterprise profile building module performs semantic parsing and vectorization on the enterprise data to obtain the enterprise profile vector. Step 6: Calculate the similarity between the policy semantic vector and the enterprise profile vector, sort the policies from high to low similarity, and extract the top n policies to form a candidate policy set; Step 7: Input the candidate policy set and enterprise profile into the big language model. The big language model extracts the applicable conditions from the candidate policy set and the enterprise features from the enterprise profile, and compares and matches the applicable conditions with the enterprise features. Step 8: Sort the policies from high to low based on their matching degree, and extract the top m policies to push to enterprises.

[0025] A medium for intelligently matching enterprises with government preferential policies for enterprises, wherein the medium stores a computer program, and when the computer program is executed, it implements a matching method for intelligently matching enterprises with government preferential policies for enterprises.

[0026] Example 2: This embodiment discloses a system for intelligently matching government preferential policies for enterprises, which is an improvement on Embodiment 1. The similarity retrieval module includes a vectorization construction module, a vector retrieval module, and a candidate policy generation module. The vectorization construction module converts structured policy content into policy semantic vectors suitable for semantic detection and stores the policy semantic vectors in a vector database. The vector detection module is connected to the enterprise profile construction module and the vectorization construction module respectively. The vector detection module performs similarity retrieval on the policy semantic vectors and the enterprise profile vectors. The candidate policy generation module summarizes policies with similarity higher than the similarity threshold to form a candidate policy set.

[0027] Furthermore, the vectorization construction module transforms structured policy content into vector representations suitable for semantic retrieval and stores them in a vector database. Based on the application requirements of the policy knowledge base, the system selects an appropriate Chinese text vectorization model, encoding fields such as policy name, policy summary, key descriptions in the Condition-Action structure, support methods, and applicable conditions to generate high-dimensional semantic vectors. Each policy record is associated with its corresponding basic fields and lifecycle information in the vector database, forming a complete policy knowledge entry.

[0028] To improve retrieval accuracy, this invention further introduces a hybrid vector strategy based on semantic vectors, fusing semantic features with structural features. Structural features include policy release level (national, provincial, municipal, district), industry tags associated with the policy, support method coding, funding scale segmentation coding, and key numerical thresholds. The system concatenates or weights the semantic vectors and structural feature vectors to form a hybrid vector representation, enabling it to not only focus on semantic similarity during similarity retrieval but also identify policy entries that better match enterprise conditions at the structural level.

[0029] The vectorization building module uses Milvus, FAISS, or other engines that support approximate nearest neighbor search as its underlying implementation and builds multi-level indexes to improve retrieval performance. To improve the overall system response speed, this module also implements a multi-level caching mechanism. For frequently accessed hot policy vectors, common enterprise profile vectors, and query vectors that appear repeatedly in a short period of time, the system will cache the retrieval results or intermediate feature vectors to reduce redundant calculations and model calls.

[0030] Furthermore, after receiving the enterprise profile vector, the vector retrieval module performs a similarity search in the policy vector knowledge base. The system uses metrics such as Euclidean distance, inner product, or cosine similarity, combined with a vector index structure, to quickly retrieve several policy vectors most closely matching the enterprise profile within a high-dimensional vector space, forming a candidate policy set. During the search process, the system comprehensively considers the policy lifecycle field and the release level, filtering out expired, repealed, or replaced policies, and prioritizing policies that are currently valid and relevant to the enterprise's region.

[0031] The vector retrieval module employs a two-stage retrieval architecture. The first stage relies on a vector database to achieve high-speed candidate recall, resulting in a moderately sized set of candidate policies. The second stage inputs both the enterprise profile and the candidate policy texts into a large language model. This model then performs fine-grained scoring and ranking of the semantic relevance between the enterprise's questions and the policy texts, thereby re-ranking the candidate set and forming a more accurate list of policy candidates. The vector retrieval module can also weight and adjust the ranking results based on the enterprise's preferences, such as a preference for financial subsidies, tax incentives, or loan support, to meet differentiated needs.

[0032] The rest of this embodiment is the same as that of Embodiment 1, so it will not be described again.

[0033] Example 3: This embodiment discloses a matching system for intelligently matching government preferential policies for enterprises, which is an improvement on embodiment 1 or 2. The AI ​​discriminant analysis module includes a matching module, a cross-validation module, and a logical consistency detection module. The matching module is used to compare the structured policy content of each policy in the candidate policy set with the attribute values ​​in the enterprise profile into a large language model, and output the matching degree between the current policy and the enterprise profile based on the comparison results. The cross-validation module is used to simultaneously mobilize at least two different large language models to participate in the comparison process between the structured policy content and the attribute values ​​in the enterprise profile. The logical consistency detection module verifies the policy-related condition information output by the large language model to obtain the policy finally pushed to the enterprise. The AI ​​discriminant analysis module includes a risk analysis module, which is used to identify whether there are compliance risks between the condition information involved in the pushed policy and the enterprise's historical data.

[0034] After acquiring the candidate policy set, the matching module inputs the structured content of each candidate policy along with the enterprise profile into the large language model for item-by-item analysis. The large language model compares the policy's applicable conditions and restrictive clauses with various attribute values ​​in the enterprise profile to determine whether the policy is "fully applicable," "partially applicable," or "inapplicable." For fully applicable policies, the model provides a recommendation level such as "recommended" or "strongly recommended"; for partially applicable policies, the model indicates the necessary supplementary materials or unmet marginal conditions; for inapplicable policies, the model provides a clear explanation of the reasons for inapplicability.

[0035] To mitigate the risk of erroneous judgments from large language models in complex policy scenarios, this invention introduces a multi-model cross-validation mechanism in the cross-validation module and a logical consistency detection mechanism in the logical consistency detection module. The cross-validation module can simultaneously call two large language models with different architectures or training focuses to provide evaluation results. When the two models show significant inconsistencies in key conclusions, the policy is marked as "requiring manual review," preventing the system from directly outputting potentially controversial conclusions to enterprises. Simultaneously, the logical consistency detection module performs program-level verification of numerical thresholds, time conditions, and eligibility criteria involved in the large language model output. For example, it verifies whether the enterprise's operating revenue falls within a certain specified range based on its previous year's financial data, or calculates whether the enterprise has been "established for at least two years" based on its establishment date, thereby preventing erroneous recommendations due to misunderstandings.

[0036] The risk analysis module includes a policy compliance risk analysis function. Based on policy restrictions and the company's past application records, this function automatically identifies situations such as duplicate applications, overlapping receipt of similar funds, non-compliance with fund usage restrictions, or being within a policy-mandated application ban period. It provides risk warnings in the output results, helping companies conduct compliance self-checks before applying. Furthermore, this module includes original citations or summaries of key clauses in its output conclusions, allowing companies to clearly understand the basis of the system's judgment and improving the transparency and explainability of the intelligent judgment process.

[0037] The rest of this embodiment is the same as that of embodiment 1 or 2, so it will not be described again.

[0038] Example 4: This embodiment discloses a matching system for intelligent matching of government preferential policies for enterprises, which is an improvement on any one of embodiments 1-3. The policy screening and filtering module includes a keyword module and a rule module. The keyword module is used to identify and extract keywords in the policy information. The rule module filters the policy information based on syntactic rules and determines whether the filtered policy information is related to the theme of preferential policies for enterprises.

[0039] After acquiring the original policy dataset, the policy filtering module performs initial filtering of the text content using a keyword module. Based on a pre-set keyword library and rule set, the system scans the policy titles, body text, and attachments. By identifying words or phrases such as "enterprise," "subsidy," "tax incentives," "financing support," "SMEs," "special funds," and "technological innovation," the rule module combines syntactic rules to determine whether the policy is relevant to the theme of benefiting enterprises, thus filtering out a set of candidate policies. During this process, the policy filtering module also considers document quality assessment results, excluding documents that are clearly missing text, contain serious garbled characters, or have incomplete content from the candidate set.

[0040] To improve the efficiency of subsequent retrieval and judgment, the policy screening and filtering module will also assign priority tags to each candidate policy based on information such as policy release level, policy type, release time, and whether it is still valid. Policies with higher priority will be processed first in the vectorization and retrieval stages, realizing the priority modeling and matching of high-value policies benefiting enterprises.

[0041] The rest of this embodiment is the same as any one of embodiments 1-3, so it will not be described again.

[0042] Example 5: This embodiment discloses a matching system for intelligent matching of government preferential policies for enterprises, which is an improvement on any one of embodiments 1-4. The large language model parsing module includes a condition module and an action module. The condition module is used to extract and describe the condition information in the candidate policy documents, and the action module is used to extract the preferential support information related to the current policy.

[0043] The large language model parsing module is used to transform candidate policy documents into structured policy records in a unified format. The system inputs the screened policy text into a pre-trained large language model that has been fine-tuned for the policy text. The large language model performs semantic understanding and segmentation analysis on the policy content, and generates structured results based on a preset output template. These results include fields such as policy name, issuing entity, applicable objects, applicable conditions, support methods, funding scale, application process, required materials, policy validity period, and applicable regions. This transforms the originally unstructured, lengthy official documents into computable and searchable data objects.

[0044] Building upon this foundation, this module introduces a Condition-Action structure extraction mechanism. The condition and action modules further break down policy clauses into condition and action components. The condition component describes the various thresholds that enterprises need to meet, such as industry category, enterprise size, registered location, establishment date, annual operating revenue, asset size, R&D expenditure ratio, and whether it is a high-tech enterprise. The action component describes the preferential or support measures provided by the government, such as financial subsidies, loan interest subsidies, tax reductions, site subsidies, and talent awards. The system uses named entity recognition and relation extraction models to identify the numerical, temporal, and eligibility conditions involved in the clauses and stores them in a structured form, thus providing a basis for subsequent rule-based and model-based automatic judgments.

[0045] The large language model parsing module includes a lifecycle management module, which injects time-based lifecycle fields into each candidate policy document. To address the issue of frequent policy adjustments, the lifecycle management module also incorporates policy lifecycle management technology, adding relevant lifecycle fields such as release date, effective date, implementation period, whether it has been repealed, and whether it has been replaced by a higher-level policy to each policy record in the structured results. In the subsequent retrieval and intelligent judgment stages, the system can automatically filter out expired or replaced policies based on these fields, ensuring the timeliness and effectiveness of the matching results returned to enterprises. The large language model parsing module can employ a BERT+CRF structured extraction model, a ChatGLM-like Chinese large language model combined with template generation, or a rule engine for field filling to complete structured parsing. The specific implementation can be replaced according to computing power and data scale.

[0046] The rest of this embodiment is the same as any one of embodiments 1-4, so it will not be described again.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A system for intelligently matching enterprises with government preferential policies, characterized in that, It includes a data crawling module, a policy filtering module, a large language model parsing module, an enterprise profile building module, a similarity retrieval module, and an AI discriminant analysis module. The data crawling module is used to collect policy information, and the policy filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags. The large language model parsing module introduces a conditional action structure extraction mechanism to decompose candidate policy documents and obtain structured policy content. The enterprise profile building module generates an enterprise profile vector based on enterprise user information; The similarity retrieval module is used to convert structured policy content into policy semantic vectors suitable for semantic retrieval, and calculate the similarity between the policy semantic vectors and the enterprise profile vectors to obtain a set of candidate policies. The AI ​​discriminant analysis module analyzes and compares each structured policy content in the candidate policy set with the enterprise profile input big language model to determine the matching degree between the current policy and the enterprise profile, and pushes the corresponding policy to the enterprise based on the matching degree.

2. The system for intelligently matching government preferential policies for enterprises according to claim 1, characterized in that, The similarity retrieval module includes a vectorization construction module, a vector retrieval module, and a candidate policy generation module. The vectorization construction module converts structured policy content into policy semantic vectors suitable for semantic detection and stores the policy semantic vectors in a vector database. The vector detection module is connected to the enterprise profile construction module and the vectorization construction module, respectively. The vector detection module performs similarity retrieval on the policy semantic vectors and the enterprise profile vectors. The candidate policy generation module summarizes policies with similarity higher than the similarity threshold to form a candidate policy set.

3. The system for intelligently matching government preferential policies for enterprises according to claim 1, characterized in that, The AI ​​discriminant analysis module includes a matching module, a cross-validation module, and a logical consistency detection module. The matching module compares the structured policy content of each policy in the candidate policy set with the attribute values ​​in the enterprise profile into a large language model, and outputs the matching degree between the current policy and the enterprise profile based on the comparison results. The cross-validation module simultaneously mobilizes at least two different large language models to participate in the comparison process between the structured policy content and the attribute values ​​in the enterprise profile. The logical consistency detection module verifies the policy-related condition information output by the large language model to obtain the final policy pushed to the enterprise.

4. The system for intelligently matching government preferential policies for enterprises according to claim 3, characterized in that, The AI ​​discrimination and analysis module includes a risk analysis module, which is used to identify whether there are compliance risks between the conditional information involved in the push policy and the company's historical data.

5. The system for intelligently matching government preferential policies for enterprises according to claim 1, characterized in that, It also includes a visualization output display module, which is used to structure and organize the matching policies pushed to enterprises by the AI ​​discrimination and analysis module, and to visually output a policy recommendation list and detailed report to enterprises.

6. The system for intelligently matching government preferential policies for enterprises according to claim 1, characterized in that, The policy screening and filtering module includes a keyword module and a rule module. The keyword module is used to identify and extract keywords from policy information. The rule module filters policy information based on syntactic rules and determines whether the filtered policy information is related to the theme of benefiting enterprises.

7. The system for intelligently matching government preferential policies for enterprises according to claim 1, characterized in that, The large language model parsing module includes a condition module and an action module. The condition module is used to extract and describe the condition information in the candidate policy documents, and the action module is used to extract information on preferential support methods for enterprises related to the current policy.

8. The system for intelligently matching government preferential policies for enterprises according to claim 7, characterized in that, The large language model parsing module includes a lifecycle management module, which is used to inject time-based relevant lifecycle fields into each candidate policy document.

9. A method for intelligently matching enterprises with government preferential policies, implemented based on the system described in any one of claims 1-8, characterized in that, Includes the following steps: Step 1: Use the data crawling module to crawl publicly available policy information to form the original policy dataset; Step 2: The policy screening and filtering module filters policy information based on keywords and filtering rules to obtain candidate policy documents with priority tags; Step 3: Input the candidate policy documents into the big language model, which will perform semantic parsing and automatic summarization of the text of the candidate policy documents to generate structured policy content that includes at least the policy name, applicable conditions, support methods, application process, and validity period. Step 4: Based on the conditional information, support methods for businesses, and relevant lifecycle fields in the structured policy content, the structured policy content is transformed into a policy semantic vector suitable for semantic retrieval through the similarity retrieval module; Step 5: Input the enterprise data into the enterprise profile building module. The enterprise profile building module performs semantic parsing and vectorization on the enterprise data to obtain the enterprise profile vector. Step 6: Calculate the similarity between the policy semantic vector and the enterprise profile vector, sort the policies from high to low similarity, and extract the top n policies to form a candidate policy set; Step 7: Input the candidate policy set and enterprise profile into the big language model. The big language model extracts the applicable conditions from the candidate policy set and the enterprise features from the enterprise profile, and compares and matches the applicable conditions with the enterprise features. Step 8: Sort the policies from high to low based on their matching degree, and extract the top m policies to push to enterprises.

10. A medium for intelligently matching enterprises with government preferential policies for businesses, characterized in that, The medium stores a computer program, which, when executed, implements the method of claim 9.

Images