Enterprise and science and technology policy matching construction method based on large model algorithm

Abstract

The invention discloses an enterprise and science and technology policy matching construction method based on a large model algorithm, and relates to the technical field of policy resource intelligent matching. The method comprises the following steps: acquiring enterprise basic information and dynamic operation data, and generating an enterprise descriptive document; constructing an enterprise feature vector based on the enterprise descriptive document; performing semantic analysis on the policy text and extracting multi-level features to generate policy feature vectors; calculating a matching degree based on the enterprise feature vector and the policy feature vector, and generating a comprehensive matching score table; and according to the comprehensive matching score table, recommending policy resources adapted to enterprise demands. According to the method, the comprehensiveness and the accuracy of policy matching are remarkably improved by constructing the multi-level representation system of the enterprise feature vectors and the policy feature vectors; and meanwhile, by adopting a technical route of deep semantic embedding and explicit-hidden feature separation, intelligent understanding of policy texts is realized, and then the matching degree of enterprises and policy resources is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent matching technology for policy resources, and in particular to a method for constructing a matching system between enterprises and science and technology policies based on a large model algorithm. Background Technology

[0002] In the context of current information and digital transformation, the precise matching of enterprises with science and technology policies has become a crucial link in promoting industrial innovation and economic development. However, due to limited channels for policy information dissemination, high complexity of interpretation, and insufficient resources for SMEs, many enterprises struggle to efficiently access and utilize policy resources tailored to their needs. In recent years, the development of artificial intelligence technology, particularly large-scale model algorithms, has offered new possibilities for solving these problems. Large-scale model algorithms, through their powerful natural language processing and deep learning capabilities, can intelligently analyze policy texts and combine them with personalized enterprise characteristics to achieve precise matching, thereby significantly improving the efficiency and accuracy of policy matching.

[0003] A search revealed Chinese invention patent CN119226499B, which discloses a policy-enterprise matching method and apparatus based on a large model, published on March 4, 2025. This patent proposes a policy-enterprise matching method based on graph contrastive learning. It constructs a bipartite graph of policy and enterprise information and utilizes a recommendation model to learn and match the complex relationships between policies and enterprises. The advantage of this technical solution lies in capturing underlying semantic information through contrastive learning, improving the model's understanding of graph structures. However, this method relies heavily on graph structure modeling and training, failing to fully consider the implicit conditions in policy texts. Furthermore, its comprehensive analytical capabilities for multi-dimensional enterprise characteristics (such as industry type, size, and technological innovation needs) are limited, potentially leading to insufficient comprehensiveness and flexibility in the matching results. In addition, the solution lacks a mechanism for constructing and real-time updating a dynamic policy database, which may affect the timeliness and accuracy of the matching results.

[0004] A search revealed Chinese invention patent CN113870083B, which discloses a method, apparatus, system, electronic device, and readable storage medium for policy matching, published on October 18, 2024. This patent proposes a policy matching method based on policy application condition extraction and user parameter information matching. It extracts configuration rules from the application conditions of the target policy and matches them with enterprise application information to obtain the policy matching degree. The advantage of this technical solution is that it improves the efficiency and accuracy of policy matching. However, this method mainly focuses on the explicit matching of policy application conditions and lacks the ability to mine the deep semantic information of policy texts, especially the ability to identify implicit conditions in policies. Furthermore, this solution does not fully utilize the deep learning capabilities of large-scale model algorithms, which may limit its ability to handle complex policy texts and multi-dimensional enterprise needs. The personalization and applicability of the matching results need further improvement.

[0005] The aforementioned problems indicate that existing policy matching methods still have certain shortcomings in areas such as in-depth analysis of policy texts, comprehensive consideration of multi-dimensional enterprise characteristics, and the construction and real-time updating of dynamic policy databases. Therefore, this invention provides a method for constructing enterprise-technology policy matching based on large-scale model algorithms. It aims to achieve intelligent extraction of policy information, accurate characterization of enterprise needs, multi-level and multi-dimensional matching logic design, and intelligent management and recommendation of dynamic policy databases by integrating advanced large-scale model algorithms. This will comprehensively improve the efficiency, accuracy, and applicability of policy matching, meeting the policy needs of enterprises during their digital transformation. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for constructing a matching system between enterprises and science and technology policies based on a large model algorithm. By constructing a multi-level representation system of enterprise feature vectors and policy feature vectors, the comprehensiveness and accuracy of policy matching are significantly improved. At the same time, by adopting a technical approach of deep semantic embedding and separation of explicit and implicit features, intelligent understanding of policy texts is achieved, thereby improving the matching degree between enterprises and policy resources.

[0007] In a first aspect, the present invention provides a method for constructing a matching system between enterprises and science and technology policies based on a large model algorithm, comprising the following steps: S1. Obtain basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; S2. Construct enterprise feature vectors based on enterprise descriptive documents; S3. Perform semantic parsing on the policy text and extract explicit and implicit conditional features. Then, perform weighted fusion of the explicit and implicit conditional features to generate a policy feature vector. S4. Calculate the matching degree based on enterprise feature vectors and policy feature vectors, and generate a comprehensive matching score table; S5. Recommend policy resources that match the needs of enterprises based on the comprehensive matching score table, and establish a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and adjust the weight parameters of feature dimensions in the matching degree calculation based on the adoption status. The basic enterprise information includes enterprise registration information, industry classification, technological innovation capability assessment, and financial status indicators. The dynamic operational data includes the company's recent project progress, R&D investment ratio, and market performance data.

[0008] Preferably, the acquisition of basic enterprise information and dynamic operational data includes the following steps: Collect basic enterprise information and preprocess the unstructured data in the basic enterprise information. The preprocessing includes removing redundant data, standardizing data format, and extracting key fields based on natural language processing technology. Collect dynamic operational data and calibrate the time series in the dynamic operational data to ensure that the dynamic operational data is sorted by time; The company's basic information and dynamic operational data are integrated into a unified data framework to form a descriptive document for the company.

[0009] Preferably, the step of constructing an enterprise feature vector based on enterprise descriptive documents includes the following steps: Numerical processing is performed on the fields in the enterprise's descriptive documents. Discrete fields are converted into numerical vectors through one-hot encoding, and continuous fields are mapped to fixed intervals through normalization. The one-hot encoding is used to process classification data, converting it into binary vectors to avoid the model misjudging the numerical relationship between categories. Based on the assessment of technological innovation capabilities and financial status indicators in the enterprise's basic information, and combined with the R&D investment ratio and market performance data in the dynamic operation data, a joint feature vector representing the static attributes and dynamic behavior of the enterprise is constructed. Principal component analysis is used to reduce the dimensionality of the joint feature vector, which retains the main information while reducing redundant dimensions, and generates enterprise feature vectors; these enterprise feature vectors are used to accurately reflect the core characteristics of the enterprise.

[0010] Preferably, the step of semantically parsing the policy text to extract explicit and implicit conditional features, and then weightedly fusing the explicit and implicit conditional features to generate a policy feature vector, includes the following steps: A policy text library was generated by collecting policy release sections from government official websites, science and technology information portals, and third-party policy compilation databases through multiple sets of customized web crawlers. The policy text is segmented into words, and meaningless words are filtered out based on a stop word list to eliminate text noise and retain key semantic information. The BERT series model algorithm based on the Transformer architecture is used to perform deep semantic embedding on policy text, transforming natural language descriptions into high-dimensional semantic vectors and generating initial semantic vectors; Based on the initial semantic vector, the explicit and implicit conditional features of the policy text are analyzed through a hierarchical feature extraction module. The explicit conditional features mainly reflect the specific requirements for policy application, while the implicit conditional features focus on the deeper intent of the policy text. The feature fusion module performs weighted fusion of explicit and implicit conditional features to generate a policy feature vector that combines policy constraints and incentives. This policy feature vector is used to accurately characterize the objective requirements of the policy and reflect the strategic intent of policy formulation.

[0011] Preferably, the step of calculating the matching degree based on enterprise feature vectors and policy feature vectors to generate a comprehensive matching score table includes the following steps: Spatial alignment is performed on enterprise feature vectors and policy feature vectors to ensure that they are comparable on the same feature dimensions; A multi-level similarity calculation method is used to evaluate the degree of fit between enterprise feature vectors and policy feature vectors from different dimensions; For explicit conditional features, a rule-based matching algorithm is used to strictly compare the enterprise's qualifications with the policy's hard requirements to ensure that the enterprise meets the basic threshold for policy application; for implicit conditional features, semantic similarity calculation is used to assess the fit between the enterprise's technological development direction, market positioning and policy support goals. In the matching degree calculation process, the weight allocation of different features is comprehensively considered; for technology innovation policies, the weight of enterprise R&D investment and technology capability scores will be increased accordingly; while for industry support policies, more attention is paid to the industry attributes and market performance of enterprises. A comprehensive matching score table is generated based on the matching calculation results. The comprehensive matching score table is used to quantify the degree of fit between enterprises and policies.

[0012] Preferably, the step of recommending policy resources that match the needs of enterprises based on the comprehensive matching scoring table includes the following steps: Based on the comprehensive matching score table, policy resources with a matching score higher than the preset matching threshold are selected. The selected policy resources are sorted from high to low according to their matching scores, with particular emphasis on the correlation between the policy resources and the core characteristics of the enterprises. Based on the industry classification tags and technology innovation capability scores in the enterprise's descriptive documents, priority will be given to recommending policy resources that are highly relevant to the enterprise's industry, especially policy support programs that are targeted at specific industries or have a technology innovation orientation; A new dynamic feedback mechanism has been added, which is used to continuously track the actual adoption of recommended policy resources by enterprises. Based on the actual adoption of recommended policy resources by enterprises, the weight parameters of each feature dimension in the matching degree calculation formula are adjusted regularly to enable the recommendation strategy to adapt to changes in the policy environment and the evolution of enterprise needs.

[0013] Secondly, the present invention provides a system for constructing enterprise-technology policy matching based on a large model algorithm, which applies the enterprise-technology policy matching construction method based on the large model algorithm described above. The system includes a data acquisition module, a feature construction module, a policy parsing module, a matching calculation module, and a recommendation output module. The data acquisition module is used to acquire basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; The feature construction module is used to construct enterprise feature vectors based on enterprise descriptive documents; The policy parsing module is used to perform semantic parsing on the policy text and extract explicit condition features and implicit condition features. The explicit condition features and implicit condition features are then weighted and fused to generate a policy feature vector. The matching calculation module is used to calculate the matching degree based on the enterprise feature vector and the policy feature vector, and generate the matching result; The recommendation output module is used to recommend policy resources that match the needs of enterprises based on matching degree ranking, and to establish a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and to adjust the weight parameters of feature dimensions in matching degree calculation based on the adoption status.

[0014] Preferably, the recommendation output module includes a dynamic feedback submodule and a weight adjustment submodule; The dynamic feedback submodule is used to record the adoption of recommended policy resources by enterprises and generate feedback data; The weight adjustment submodule is used to adjust the weight parameters in the matching degree calculation formula based on feedback data to optimize the matching accuracy. If a company repeatedly adopts policy resources with a high degree of matching, the weight of the corresponding industry classification label will be increased; if a company repeatedly fails to adopt policy resources with a low degree of matching, the weight of the corresponding industry classification label will be decreased.

[0015] Preferably, the policy parsing module includes an explicit condition extraction submodule, an implicit condition mining submodule, and a feature fusion module: The explicit condition extraction submodule is used to extract the logical constraint relationships in the policy application conditions; The implicit condition mining submodule is used to focus on potential incentive objectives and applicable scenarios in policy texts; The outputs of the explicit condition extraction submodule and the implicit condition mining submodule are weighted and summed by the feature fusion module to generate the final policy feature vector.

[0016] Preferably, the feature construction module includes a data preprocessing submodule and a feature dimensionality reduction submodule; The data preprocessing submodule is used to construct a joint feature vector representing the static attributes and dynamic behaviors of an enterprise; The feature dimensionality reduction submodule is used to perform dimensionality reduction processing on the joint feature vector using principal component analysis to extract the main features and reduce computational complexity. The outputs of the data preprocessing submodule and the feature dimensionality reduction submodule are integrated by a joint feature vector generator to form the final enterprise feature vector.

[0017] Compared with the prior art, the present invention has the following advantages and beneficial effects: This invention significantly improves the comprehensiveness and accuracy of policy matching by constructing a multi-level representation system of enterprise feature vectors and policy feature vectors. At the feature construction level, it not only integrates static attributes of enterprises (such as structured data like industry classification and financial status) but also innovatively introduces dynamic operational data (such as time-series information like R&D investment and market performance), forming high-information-density feature representations through dimensionality reduction techniques such as principal component analysis. This dual-modal feature extraction mechanism can simultaneously capture the inherent characteristics and development trends of enterprises, providing a more complete feature foundation for matching calculations.

[0018] Secondly, in terms of policy analysis, a deep semantic embedding and explicit / implicit feature separation approach is adopted to achieve intelligent understanding of policy texts. After extracting initial semantic vectors through a pre-trained large model, the explicit conditional features (such as hard constraints like qualification requirements) and implicit conditional features (such as strategic intentions like industry guidance) of the policy are analyzed separately. Then, feature fusion is used to generate interpretable policy feature vectors. This hierarchical analysis method ensures a strict match between policy requirements and can also uncover the potential correlation between policies and enterprise needs, effectively solving the problem of insufficient semantic understanding in traditional matching methods.

[0019] This invention significantly reduces the cost for enterprises to obtain policy resources through an automated and intelligent matching process, while improving the efficiency of government policy implementation. Its comprehensive scoring system not only quantifies the matching degree but also assesses the potential impact of policies on enterprises, providing multi-dimensional references for decision-making. Overall, this patent achieves efficient matching of policy resources and enterprise needs through technological innovation, and has significant practical implications for optimizing the business environment and promoting technological innovation. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the construction method for matching enterprises with science and technology policies based on a large model algorithm, as described in this invention.

[0021] Figure 2 This is a schematic diagram of the structure of the enterprise and technology policy matching system based on the large model algorithm of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. It should be noted that relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0023] Example 1 Please see Figure 1 This invention provides a method for constructing a matching system between enterprises and science and technology policies based on a large model algorithm, including: S1. Obtain basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; S2. Construct enterprise feature vectors based on enterprise descriptive documents; S3. Perform semantic parsing on the policy text and extract explicit and implicit conditional features. Then, perform weighted fusion of the explicit and implicit conditional features to generate a policy feature vector. S4. Calculate the matching degree based on enterprise feature vectors and policy feature vectors, and generate a comprehensive matching score table; S5. Recommend policy resources that match the needs of enterprises based on the comprehensive matching score table, and establish a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and adjust the weight parameters of feature dimensions in the matching degree calculation based on the adoption status. The basic enterprise information includes enterprise registration information, industry classification, technological innovation capability assessment, and financial status indicators. The dynamic operational data includes the company's recent project progress, R&D investment ratio, and market performance data.

[0024] Specifically, the acquisition of basic enterprise information and dynamic operational data includes the following steps: Collect basic enterprise information and preprocess the unstructured data in the basic enterprise information. The preprocessing includes removing redundant data, standardizing data format, and extracting key fields based on natural language processing technology. Collect dynamic operational data and calibrate the time series in the dynamic operational data to ensure that the dynamic operational data is sorted by time; By integrating basic enterprise information and dynamic operational data into a unified data framework, a descriptive document for the enterprise can be formed. For example, the descriptive document can be expressed in the following format: ,in, A unique identifier for the enterprise. For enterprise type number, For the company's establishment timestamp, For the geographical coordinates of the enterprise, As an indicator of financial condition, Industry-specific labels for businesses. Assess enterprises' technological innovation capabilities. The proportion of enterprise R&D investment, This process, which assesses a company's market performance capabilities, involves collecting multi-dimensional characteristics of the enterprise to provide high-quality data input for subsequent feature construction module processing.

[0025] Specifically, the process of constructing enterprise feature vectors based on enterprise descriptive documents includes the following steps: The fields in the enterprise's descriptive documents are numerically processed. Discrete fields are converted into numerical vectors through one-hot encoding to eliminate spurious numerical correlations between categories. Continuous fields are normalized and mapped to fixed intervals to ensure fair weighting of features with different dimensions. The one-hot encoding is used to process classification data, converting it into binary vectors to avoid the model misjudging the numerical relationship between categories. Based on the assessment of technological innovation capabilities and financial status indicators in the enterprise's basic information, and combined with the R&D investment ratio and market performance data in the dynamic operation data, a joint feature vector representing the static attributes and dynamic behavior of the enterprise is constructed. The joint feature vector can reflect the enterprise's industry attributes, technological level and financial health status, and can also capture the enterprise's development trend through the time-series characteristics of the dynamic operation data, thereby forming a more comprehensive enterprise profile.

[0026] To improve the computational efficiency and representational capability of the joint feature vector, principal component analysis is used to reduce the dimensionality of the joint feature vector. While retaining the main information, redundant dimensions are reduced, and finally, enterprise feature vectors are generated. The enterprise feature vectors can accurately reflect the core characteristics of enterprises and adapt to the computational needs of large-scale policy matching, laying a data foundation for subsequent matching degree calculations.

[0027] Specifically, the step of semantically parsing the policy text library and extracting multi-level features to generate policy feature vectors includes the following steps: A policy text library was generated by collecting policy release sections from government official websites, science and technology information portals, and third-party policy compilation databases through multiple sets of customized web crawlers. To ensure the processability of policy texts, it is necessary to first segment the policy texts into words and then filter out meaningless words based on a stop word list in order to eliminate text noise and retain key semantic information. In the semantic parsing stage, the preprocessed policy text is first input into the BERT model. The model then uses the output vector corresponding to the final hidden layer classification label, or performs average pooling on all word vectors, to generate an initial semantic vector. This initial semantic vector captures the deep semantic information of the policy text. This step effectively captures the contextual relationships and semantic connotations of the policy text, providing a basic representation for subsequent feature extraction.

[0028] Based on the initial semantic vector, the explicit and implicit conditional features of the policy text are further analyzed through a hierarchical feature extraction module. Explicit conditional features mainly reflect the specific requirements for policy application, such as structured constraints like industry restrictions, qualification requirements, and application deadlines. The logical relationships within these constraints are extracted and transformed into computable standardized features through the explicit conditional extraction submodule. Implicit conditional features focus on the deeper intent of the policy text, such as potential goals like industry support guidance and technological innovation incentives. The implicit conditional mining submodule identifies the core concerns of the policy and its applicable scenarios, thereby uncovering the implicit relationship between the policy and enterprise needs. The feature fusion module weighted and fused explicit and implicit conditional features to generate a policy feature vector that combines policy constraints and incentives. This policy feature vector not only accurately represents the objective requirements of the policy but also reflects the strategic intent behind its formulation, thus providing a multi-dimensional representation of policy features for subsequent matching degree calculations. This process ensures the comprehensiveness and interpretability of the policy feature vector, laying an important foundation for accurately matching enterprise needs.

[0029] Specifically, the step of calculating the matching degree based on enterprise feature vectors and policy feature vectors and generating a comprehensive matching score table includes the following steps: First, spatial alignment is performed on the enterprise feature vector and the policy feature vector to ensure that the enterprise feature vector and the policy feature vector are comparable on the same feature dimension; Based on this, a multi-level similarity calculation method is adopted to evaluate the degree of fit between enterprise feature vectors and policy feature vectors from different dimensions; For explicit conditional features, rule-based matching algorithms are mainly used to strictly compare the enterprise's qualifications with the policy's hard requirements to ensure that the enterprise meets the basic threshold for policy application; for implicit conditional features, semantic similarity calculation is used to assess the fit between the enterprise's technological development direction, market positioning and policy support goals. In the matching degree calculation process, the weight allocation of different characteristics is comprehensively considered. For example, for technology innovation policies, the weight of enterprise R&D investment and technology capability scores will be increased accordingly; while for industry support policies, more attention will be paid to the industry attributes and market performance of enterprises. This dynamic weight adjustment mechanism can ensure that the matching results not only meet the hard constraints of the policies, but also reflect the core orientation of the policies; In addition, a comprehensive matching score table is generated based on the matching calculation results. The comprehensive matching score table can quantify the degree of fit between enterprises and policies.

[0030] The comprehensive matching score table is not generated through simple linear sorting, but rather by a multi-objective optimization algorithm that maximizes the fit between enterprise characteristics and implicit policy conditions while meeting basic policy requirements. This approach avoids matching bias caused by excessively high values ​​for a single indicator, ensuring that the recommended results are both policy compliant and truly meet the development needs of enterprises. The generated comprehensive matching score table will be output in a structured format, including key information such as matching score, policy applicability analysis, and recommendation priority. This process fully considers the complexity and multi-dimensional characteristics of policy matching, making the recommended results both objective and reflecting the deep-seated relationship between policies and enterprises.

[0031] Specifically, the step of recommending policy resources that match the needs of enterprises based on the comprehensive matching scoring table, establishing a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and adjusting the weight parameters of feature dimensions in the matching degree calculation based on the adoption status includes the following steps: Based on the comprehensive matching score table, policy resources with matching scores higher than the preset matching threshold are selected to ensure that the recommended policy resources have basic suitability. The selected policy resources are sorted from high to low according to their matching scores, with particular emphasis on the correlation between the policy resources and the core characteristics of the enterprises. Based on the industry classification tags and technology innovation capability scores in the enterprise's descriptive documents, priority will be given to recommending policy resources that are highly relevant to the enterprise's industry, especially those policy support programs that are industry-specific or technology innovation-oriented. To enhance the adaptability of the matching method, this invention employs a dynamic feedback mechanism. This mechanism continuously tracks the actual adoption of recommended policy resources by enterprises, including behavioral data such as viewing frequency and application intention. Based on this feedback, the weight parameters of each feature dimension in the matching degree calculation formula are periodically adjusted, enabling the recommendation strategy to adapt to changes in the policy environment and evolving enterprise needs. For example, if a certain type of policy resource is consistently favored by enterprises in a specific industry, the system will automatically increase the matching weight of relevant industry features; if certain policy conditions frequently become obstacles for enterprises in actual applications, the system will adjust the scoring criteria for those conditions accordingly. This closed-loop optimization mechanism ensures that the policy recommendation system can continuously improve, constantly enhancing the accuracy and practicality of the recommendation results.

[0032] Example 2 Please see Figure 2 This invention provides a system for constructing enterprise-technology policy matching based on a large model algorithm. The system applies the enterprise-technology policy matching construction method based on the large model algorithm described above. The system includes a data acquisition module, a feature construction module, a policy parsing module, a matching calculation module, and a recommendation output module. The data acquisition module is used to acquire basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; The feature construction module is used to construct enterprise feature vectors based on enterprise descriptive documents; The policy parsing module is used to perform semantic parsing on policy texts and extract multi-level features to generate policy feature vectors; The matching calculation module is used to calculate the matching degree based on the enterprise feature vector and the policy feature vector, and generate the matching result; The recommendation output module is used to recommend policy resources that match the needs of enterprises based on matching degree ranking, and to establish a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and to adjust the weight parameters of feature dimensions in matching degree calculation based on the adoption status.

[0033] Specifically, the recommendation output module includes a dynamic feedback submodule and a weight adjustment submodule; The dynamic feedback submodule is used to record the adoption of recommended policy resources by enterprises and generate feedback data; The weight adjustment submodule is used to adjust the weight parameters in the matching degree calculation formula based on feedback data to optimize the matching accuracy. If a company repeatedly adopts policy resources with a high degree of matching, the weight of the corresponding industry classification label will be increased; if a company repeatedly fails to adopt policy resources with a low degree of matching, the weight of the corresponding industry classification label will be decreased.

[0034] Specifically, the policy parsing module includes an explicit condition extraction submodule, an implicit condition mining submodule, and a feature fusion module; The explicit condition extraction submodule is used to extract the logical constraints in the policy application conditions; The implicit condition mining submodule is used to focus on potential incentive objectives and applicable scenarios in policy texts; The outputs of the explicit condition extraction submodule and the implicit condition mining submodule are weighted and summed by the feature fusion module to generate a policy feature vector.

[0035] Specifically, the feature construction module includes a data preprocessing submodule, a feature dimensionality reduction submodule, and a joint feature vector generator; The data preprocessing submodule is used to construct a joint feature vector representing the static attributes and dynamic behaviors of an enterprise; The feature dimensionality reduction submodule is used to reduce the dimensionality of the joint feature vector by principal component analysis and extract the main features to reduce computational complexity. The outputs of the data preprocessing submodule and the feature dimensionality reduction submodule are integrated by a joint feature vector generator to form the final enterprise feature vector.

[0036] Example 3 Please see Figure 1-2 In order to enable those skilled in the art to fully understand and implement the present invention, the specific implementation principle of the present invention is further supplemented below with a specific application scenario.

[0037] In a policy service platform of a science and technology park, the method and system of this invention are applied to the scenario of precise matching between enterprises and science and technology policies. The platform acquires basic information and dynamic operational data of enterprises within the park through a data acquisition module, and generates enterprise feature vectors using a feature construction module. Simultaneously, a policy parsing module performs semantic analysis on the latest science and technology innovation support policies and generates policy feature vectors. A matching calculation module calculates the matching degree based on the two feature vectors and generates a comprehensive matching score table. A recommendation output module filters policy resources suitable for enterprises based on the matching results, prioritizing support policies highly relevant to their technological innovation direction. Through a dynamic feedback mechanism, the platform records the enterprises' adoption of recommended policies and adjusts the weight parameters in the matching algorithm based on feedback data, thereby optimizing matching accuracy. The specific implementation principles and operation process of each step are described in detail below.

[0038] First, during the operation of the data acquisition module, the platform accesses basic information such as enterprise registration information, industry classification, technological innovation capability assessment, and financial status indicators through interfaces. For unstructured data, such as natural language fields in enterprise descriptive documents, the data acquisition module uses natural language processing technology to extract key fields, remove redundant data, and standardize data formats. For dynamic operational data, the platform collects data on the enterprise's recent project progress, R&D investment ratio, and market performance, and calibrates the time series to ensure that the data is sorted chronologically. Subsequently, the data acquisition module integrates the enterprise's basic information and dynamic operational data into a unified data framework, forming the enterprise's descriptive document. ,in, A unique identifier for the enterprise. For enterprise type number, For the company's establishment timestamp, For the geographical coordinates of the enterprise, As an indicator of financial condition, Industry-specific labels for businesses. Assess enterprises' technological innovation capabilities. The proportion of enterprise R&D investment, This demonstrates the company's market performance capabilities. Through the above operations, the data acquisition module has completed a comprehensive collection of the company's multi-dimensional characteristics, providing high-quality data input for subsequent modules.

[0039] Secondly, during the feature construction module's operation, the data preprocessing submodule quantifies the fields in the enterprise's descriptive documents. Discrete fields are converted into numerical vectors using one-hot encoding, while continuous fields are normalized and mapped to fixed intervals. Subsequently, the data preprocessing submodule combines the enterprise's basic information, including its technological innovation capability assessment and financial status indicators, with dynamic operational data such as R&D investment ratio and market performance data, to construct a joint feature vector representing the enterprise's static attributes and dynamic behaviors. The feature dimensionality reduction submodule uses principal component analysis to reduce the dimensionality of the joint feature vector, extracting key features to reduce computational complexity. Finally, the outputs of the data preprocessing and feature dimensionality reduction submodules are integrated by a joint feature vector generator to form the enterprise feature vector. This process achieves efficient extraction and representation of enterprise features, providing accurate input data for the subsequent matching calculation module.

[0040] Next, during the operation of the policy parsing module, the policy text is first segmented into words, and meaningless words are filtered out based on a stop word list. Then, the policy text is semantically embedded using BERT series algorithms based on the Transformer architecture, generating an initial semantic vector. The explicit condition extraction submodule extracts the logical constraints in the policy application conditions through the rule matching module, using them as explicit condition features. The implicit condition mining submodule focuses on the potential incentive targets and applicable scenarios in the policy text through an attention mechanism module, extracting implicit condition features. The outputs of the explicit condition extraction submodule and the implicit condition mining submodule are weighted and fused through the feature fusion module to generate the final policy feature vector. Through multi-level parsing of the policy text, the policy parsing module achieves a deep understanding of the policy content, providing high-quality input data for the subsequent matching calculation module.

[0041] Then, during the matching calculation module's operation, the matching degree is calculated based on the enterprise feature vector and policy feature vector, generating matching results. First, the enterprise and policy feature vectors are spatially aligned to ensure comparability across the same feature dimensions. Based on this, a multi-level similarity calculation method is used to evaluate the degree of fit between the enterprise and policy feature vectors from different dimensions. For explicit conditional features, a rule-based matching algorithm is primarily used to rigorously compare the enterprise's qualifications with the policy's hard requirements, ensuring the enterprise meets the basic threshold for policy application. For implicit conditional features, semantic similarity calculation is used to assess the fit between the enterprise's technological development direction, market positioning, and policy support objectives. During the matching degree calculation, the weight allocation of different features is comprehensively considered. For example, for technology innovation policies, the weight of enterprise R&D investment and technological capability scores is increased accordingly; while for industry support policies, more attention is paid to the enterprise's industry attributes and market performance. This dynamic weight adjustment mechanism ensures that the matching results both conform to the policy's hard constraints and reflect its core orientation.

[0042] In addition, a comprehensive matching score table is generated based on the matching calculation results. The comprehensive matching score table can quantify the degree of fit between enterprises and policies.

[0043] Finally, policy resources tailored to the company's needs are recommended based on the comprehensive matching score table. First, policy resources with matching scores higher than a preset matching threshold are selected based on the comprehensive matching score table to ensure basic suitability. The selected policy resources are then sorted from highest to lowest matching score, with particular emphasis on the correlation between the policy resources and the company's core characteristics. Combining the industry classification tags and technology innovation capability scores in the company's descriptive documents, policy resources highly relevant to the company's industry are prioritized, especially those policy support schemes targeting specific industries or with a technology innovation orientation. To enhance the adaptability of the matching method, this invention employs a dynamic feedback mechanism. This mechanism continuously tracks the actual adoption of recommended policy resources by enterprises, including behavioral data such as viewing frequency and application intention. Based on this feedback, the weight parameters of each feature dimension in the matching degree calculation formula are periodically adjusted, enabling the recommendation strategy to adapt to changes in the policy environment and evolving enterprise needs. For example, if a certain type of policy resource is consistently favored by enterprises in a specific industry, the system will automatically increase the matching weight of relevant industry features; if certain policy conditions frequently become obstacles for enterprises in actual applications, the system will adjust the scoring criteria for those conditions accordingly. This closed-loop optimization mechanism ensures that the policy recommendation system can continuously improve, constantly enhancing the accuracy and practicality of the recommendation results.

[0044] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between modules may be electrical, mechanical, or other forms.

[0045] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for constructing a matching system between enterprises and science and technology policies based on a large-scale model algorithm, characterized in that, Including the following steps: S1. Obtain basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; S2. Construct enterprise feature vectors based on enterprise descriptive documents; S3. Perform semantic parsing on the policy text and extract explicit and implicit conditional features. Then, perform weighted fusion of the explicit and implicit conditional features to generate a policy feature vector. S4. Calculate the matching degree based on enterprise feature vectors and policy feature vectors, and generate a comprehensive matching score table; S5. Recommend policy resources that match the needs of enterprises based on the comprehensive matching score table, and establish a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and adjust the weight parameters of feature dimensions in the matching degree calculation based on the adoption status. The basic enterprise information includes enterprise registration information, industry classification, technological innovation capability assessment, and financial status indicators. The dynamic operational data includes the company's recent project progress, R&D investment ratio, and market performance data.

2. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 1, characterized in that: The process of obtaining basic enterprise information and dynamic operational data includes the following steps: Collect basic enterprise information and preprocess the unstructured data in the basic enterprise information. The preprocessing includes removing redundant data, standardizing data format, and extracting key fields based on natural language processing technology. Collect dynamic operational data and calibrate the time series in the dynamic operational data to ensure that the dynamic operational data is sorted by time; The company's basic information and dynamic operational data are integrated into a unified data framework to form a descriptive document for the company.

3. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 2, characterized in that: The process of constructing enterprise feature vectors based on enterprise descriptive documents includes the following steps: Numerical processing is performed on the fields in the enterprise's descriptive documents. Discrete fields are converted into numerical vectors through one-hot encoding, and continuous fields are mapped to fixed intervals through normalization. The one-hot encoding is used to process classification data, converting it into binary vectors to avoid the model misjudging the numerical relationship between categories. Based on the assessment of technological innovation capabilities and financial status indicators in the enterprise's basic information, and combined with the R&D investment ratio and market performance data in the dynamic operation data, a joint feature vector representing the static attributes and dynamic behavior of the enterprise is constructed. Principal component analysis is used to reduce the dimensionality of the joint feature vector, which retains the main information while reducing redundant dimensions, and generates enterprise feature vectors; these enterprise feature vectors are used to accurately reflect the core characteristics of the enterprise.

4. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 1, characterized in that: The process involves semantically parsing the policy text to extract explicit and implicit conditional features, then weighting and fusing these features to generate a policy feature vector. This includes the following steps: A policy text library was generated by collecting policy release sections from government official websites, science and technology information portals, and third-party policy compilation databases through multiple sets of customized web crawlers. The policy text is segmented into words, and meaningless words are filtered out based on a stop word list to eliminate text noise and retain key semantic information. The BERT series model algorithm based on the Transformer architecture is used to perform deep semantic embedding on policy text, transforming natural language descriptions into high-dimensional semantic vectors and generating initial semantic vectors; Based on the initial semantic vector, the explicit and implicit conditional features of the policy text are analyzed through a hierarchical feature extraction module. The explicit conditional features mainly reflect the specific requirements for policy application, while the implicit conditional features focus on the deeper intent of the policy text. The feature fusion module performs weighted fusion of explicit and implicit conditional features to generate a policy feature vector that combines policy constraints and incentives. This policy feature vector is used to accurately characterize the objective requirements of the policy and reflect the strategic intent of policy formulation.

5. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 4, characterized in that: The process of calculating the matching degree based on enterprise feature vectors and policy feature vectors, and generating a comprehensive matching score table, includes the following steps: Spatial alignment is performed on enterprise feature vectors and policy feature vectors to ensure that they are comparable on the same feature dimensions; A multi-level similarity calculation method is used to evaluate the degree of fit between enterprise feature vectors and policy feature vectors from different dimensions; For explicit conditional features, a rule-based matching algorithm is used to strictly compare the enterprise's qualifications with the policy's hard requirements to ensure that the enterprise meets the basic threshold for policy application; for implicit conditional features, semantic similarity calculation is used to assess the fit between the enterprise's technological development direction, market positioning and policy support goals. In the matching degree calculation process, the weight allocation of different features is comprehensively considered; for technology innovation policies, the weight of enterprise R&D investment and technology capability scores will be increased accordingly; while for industry support policies, more attention is paid to the industry attributes and market performance of enterprises. A comprehensive matching score table is generated based on the matching calculation results. The comprehensive matching score table is used to quantify the degree of fit between enterprises and policies.

6. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 5, characterized in that: The process of recommending policy resources that match the needs of enterprises based on the comprehensive matching scoring table, establishing a dynamic feedback mechanism to track the actual adoption of recommended policy resources by enterprises, and adjusting the weight parameters of feature dimensions in the matching degree calculation based on the adoption status includes the following steps: Based on the comprehensive matching score table, policy resources with a matching score higher than the preset matching threshold are selected. The selected policy resources are sorted from high to low according to their matching scores, with particular emphasis on the correlation between the policy resources and the core characteristics of the enterprises. Based on the industry classification tags and technology innovation capability scores in the enterprise's descriptive documents, priority will be given to recommending policy resources that are highly relevant to the enterprise's industry, especially policy support programs that are targeted at specific industries or have a technology innovation orientation; A new dynamic feedback mechanism has been added, which is used to continuously track the actual adoption of recommended policy resources by enterprises. Based on the actual adoption of recommended policy resources by enterprises, the weight parameters of each feature dimension in the matching degree calculation formula are adjusted regularly to enable the recommendation strategy to adapt to changes in the policy environment and the evolution of enterprise needs.

7. A system for constructing enterprise-technology policy matching based on a large model algorithm, which applies the enterprise-technology policy matching construction method based on the large model algorithm described in claims 1-6 above, wherein the system includes a data acquisition module, a feature construction module, a policy parsing module, a matching calculation module, and a recommendation output module; The data acquisition module is used to acquire basic enterprise information and dynamic operational data, and generate enterprise descriptive documents; The feature construction module is used to construct enterprise feature vectors based on enterprise descriptive documents; The policy parsing module is used to perform semantic parsing on the policy text and extract explicit condition features and implicit condition features. The explicit condition features and implicit condition features are then weighted and fused to generate a policy feature vector. The matching calculation module is used to calculate the matching degree based on the enterprise feature vector and the policy feature vector, and generate the matching result; The recommendation output module is used to recommend policy resources that are suitable for the needs of enterprises based on the matching degree.

8. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 7, characterized in that: The recommendation output module includes a dynamic feedback submodule and a weight adjustment submodule; The dynamic feedback submodule is used to record the adoption of recommended policy resources by enterprises and generate feedback data; The weight adjustment submodule is used to adjust the weight parameters in the matching degree calculation formula based on feedback data to optimize the matching accuracy. If a company repeatedly adopts policy resources with a high degree of matching, the weight of the corresponding industry classification label will be increased. If a company repeatedly fails to adopt policy resources with low matching degree, the weight of the corresponding industry classification label will be reduced.

9. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 7, characterized in that: The policy analysis module includes an explicit condition extraction submodule, an implicit condition mining submodule, and a feature fusion module: The explicit condition extraction submodule is used to extract the logical constraint relationships in the policy application conditions; The implicit condition mining submodule is used to focus on potential incentive objectives and applicable scenarios in policy texts; The outputs of the explicit condition extraction submodule and the implicit condition mining submodule are weighted and fused by the feature fusion module to generate the final policy feature vector.

10. The method for constructing enterprise-technology policy matching based on large model algorithm according to claim 7, characterized in that: The feature construction module includes a data preprocessing submodule and a feature dimensionality reduction submodule; The data preprocessing submodule is used to construct a joint feature vector representing the static attributes and dynamic behaviors of an enterprise; The feature dimensionality reduction submodule is used to perform dimensionality reduction processing on the joint feature vector using principal component analysis to extract the main features and reduce computational complexity. The outputs of the data preprocessing submodule and the feature dimensionality reduction submodule are integrated by a joint feature vector generator to form the final enterprise feature vector.

Images