Big data-based intelligent matching and application system of fiscal and tax policies

Abstract

The application discloses a big data-based intelligent matching and application system for financial and tax policies, and relates to the technical fields of big data processing and artificial intelligence.The federal transfer learning module is arranged, independent regional edge nodes are arranged in multiple regions, each node is independently trained based on local policy audit passed cases, the output probability distribution of the local deep learning teacher model of each regional edge node is regarded as dark knowledge and is transmitted to a lightweight student network, the lightweight student network learns the audit rules contained in the teacher models of all regions through a distillation loss function, knowledge extraction and compression of the audit experience of all regions are realized, the common features output by the lightweight student network represent the shared policy general knowledge across regions, and the regional specificity features reserved by each node depict the unique audit preferences and requirements of each region.Under the premise that the regional original data is not exposed, the knowledge distillation and safe migration of the audit experience across regions are completed.

Description

Technical Field

[0001] This invention relates to the fields of big data processing and artificial intelligence technology, specifically to a big data-based intelligent matching and application system for fiscal and tax policies. Background Technology

[0002] With the increasing prevalence of corporate group development and cross-regional operations, companies face significant differences in the implementation of tax policies when applying for them in different regions. These differences also manifest in the interpretation of policies, review standards, material format requirements, and approval rules. Consequently, the mature application experience accumulated by companies in a single region often becomes ineffective when expanding to new regions, leading to frequent rejections of application materials, extended approval cycles, and missed opportunities for preferential policies. This severely restricts the efficiency of companies' cross-regional development.

[0003] Existing big data-based tax policy matching systems primarily employ semantic matching technology for policy texts, recommending relevant policy items by calculating the similarity between enterprise characteristics and policy texts. However, this architecture creates "experience silos," with each regional system operating independently. Enterprises cannot reuse past successful experiences when operating across regions, requiring them to start from scratch to explore the review preferences and requirements of new regions each time they expand. Furthermore, the lack of data privacy protection mechanisms, particularly the centralized training of unified models using raw declaration data from various regions, involves sensitive business information and internal review rules of each region, posing serious risks of data security and privacy breaches, and is difficult to implement due to data compliance regulations. Finally, the static matching mechanism, based solely on the static semantics of policy texts, fails to dynamically learn and reflect the subtle changes in execution preferences, requirements, and interpretations developed over long-term practice in different regions, leading to significant discrepancies between matching results and actual declaration approval rates.

[0004] In summary, existing technologies cannot achieve intelligent and secure migration and reuse of cross-regional policy application experience while ensuring data privacy in each region, nor can they provide enterprises with operational localization adaptation solutions. Therefore, there is an urgent need for a system that can achieve cross-regional policy experience knowledge distillation and localization adaptation while protecting data privacy. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a big data-based intelligent matching and application system for fiscal and tax policies. This system utilizes a federated transfer learning module to deploy independent regional edge nodes across multiple regions. Each node is independently trained based solely on local policy review cases, with original application data always stored locally. Employing knowledge distillation technology, each regional edge node transmits the output probability distribution of its local deep learning teacher model as dark knowledge to a lightweight student network. Through distillation loss functions, the lightweight student network learns the review patterns inherent in the regional teacher models, achieving knowledge extraction and compression of regional review experience. This forms a lightweight knowledge model that can be transferred across domains. The common features output by the lightweight student network represent general policy knowledge shared across regions, while the regional-specific features retained by each node accurately depict the unique review preferences and requirements of each region. Without exposing the original regional data, this system completes the knowledge distillation and secure transfer of cross-regional policy review experience, enabling the cross-domain flow and sharing of regional review experience in the form of knowledge.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a big data-based intelligent matching and application system for fiscal and tax policies, the system comprising: The data acquisition and preprocessing module collects raw financial and tax policy data and historical declaration data from multiple heterogeneous data sources, and performs cleaning and standardization processing. The policy knowledge graph construction module constructs a policy knowledge graph containing policy entities, corporate entities, and their relationships based on cleaned and standardized data. The intelligent matching module is connected to the policy knowledge graph construction module and is configured to receive the feature information of the target enterprise and perform semantic calculation and matching based on the policy knowledge graph to generate an initial candidate policy list. The federated transfer learning module is a federated learning framework deployed across multiple regional nodes. Each regional node is independently trained based on local policy review and approval cases, and the common features and region-specific features of each region are extracted through knowledge distillation technology to form a lightweight knowledge model that can be transferred across domains. The localization adaptation module receives the application experience features accumulated in the first region when an enterprise expands from the first region to the second region. It uses feature decoupling technology to remove the region-specific parts of the application experience features accumulated in the first region and superimposes the region-specific features of the second region obtained from the federated transfer learning module to generate a localized adaptation version of the application scheme for the second region.

[0007] Furthermore, in the data acquisition and preprocessing module, the multi-source heterogeneous data sources include policy document databases, enterprise historical declaration archives, and authorized enterprise internal ERP systems. The data types collected include structured enterprise information and financial indicator data, semi-structured XML format policy document data, and unstructured policy text and review opinion data. The cleaning and standardization process includes cleaning the collected data to remove noise and handle missing values, performing entity identification to identify policy names, issuing agencies and company names, and standardizing the format and measurement units.

[0008] Furthermore, the process by which the policy knowledge graph construction module constructs the policy knowledge graph is as follows: Natural language processing (NLP) technology is used to extract entities and relationships from the data after it has been cleaned and standardized by the data collection and preprocessing module. The extracted entity types include policy entities, enterprise entities, industry entities, and regional entities. The extracted relationship types include published, applicable, previously applied for, and rejected. Based on the extracted entities and relationships, a knowledge graph structure containing nodes and edges is constructed, where nodes correspond to the entities and edges correspond to the relationships between the entities. The constructed knowledge graph is stored in a graph database to support subsequent semantic queries and path reasoning.

[0009] Furthermore, the intelligent matching module performs semantic calculations and matching based on the policy knowledge graph to generate an initial candidate policy list as follows: Receive the characteristic information of the target enterprise, which includes the target enterprise's industry category, revenue scale, R&D investment ratio and registered region; The characteristic information of the target enterprise is transformed into a feature vector. Extract the policy feature vectors corresponding to each candidate policy entity from the policy knowledge graph. ,in An index representing a candidate policy entity; Calculate the feature vector With each of the policy feature vectors semantic similarity score between The semantic similarity score ,in, Represents the cosine similarity function. Represents the feature vector With the policy feature vector dot product, Represents the feature vector The length of the mold, Represents the policy feature vector The modulus length; The candidate policy entities are sorted from highest to lowest according to their semantic similarity scores, and the top K candidate policy entities are selected to generate the initial candidate policy list.

[0010] Furthermore, the federated transfer learning module includes: Multiple regional edge nodes, each corresponding to a geographical region, each regional edge node stores local raw declaration data and runs local deep learning teacher models; The central coordination server communicates with the multiple regional edge nodes and is configured to coordinate the deep learning teacher model training process of each regional edge node, aggregate the common feature knowledge uploaded by each regional edge node, and distribute global model parameters.

[0011] Furthermore, the knowledge distillation technique is as follows: Each region's edge nodes use the output probability distribution of the deep learning teacher model as dark knowledge. This dark knowledge is then passed to a lightweight student network via a distillation loss function. The output of the lightweight student network represents both the common features and the region-specific features. The distillation loss function is expressed as follows: ,in, This represents the total number of samples of locally approved policy cases in the edge nodes of the region. For the deep learning teacher model to the first The probability distribution vector of each sample output. For the lightweight student network to the first The probability distribution vector of each sample output, KL Let KL divergence function be used to measure the difference between two probability distribution vectors. By minimizing the distillation loss function, the lightweight student network learns the auditing rules inherent in the deep learning teacher model.

[0012] Furthermore, the localization adaptation module includes: The feature decoupling unit performs matrix decomposition on the application experience features from the first region, decomposing them into a general feature matrix of the first region and a specific feature matrix of the first region. The feature recombination unit, connected to the feature decoupling unit, is configured to discard the first region-specific feature matrix and fuse and reconstruct the first region general feature matrix with the region-specific features of the second region obtained from the federated transfer learning module to generate the localized adapted application scheme.

[0013] Furthermore, the process by which the feature decoupling unit performs matrix decomposition on the declared experience features from the first region is as follows: The application experience characteristics of the first region are used as the experience feature matrix. ,in, This indicates the number of feature dimensions representing the application experience of the first region. This indicates the number of samples involved in the application experience of the first region; The feature decoupling unit uses a nonnegative matrix factorization algorithm to decompose the empirical feature matrix. Decomposed into the general feature matrix of the first region The specific feature matrix of the first region The product of, where For the hidden feature dimension, and ; The general feature matrix of the first region This represents general knowledge for policy review shared across regions, and the specific feature matrix of the first region. This represents the unique review preferences and requirements of the first region.

[0014] Furthermore, the process of fusion and reconstruction of the feature recombination units is as follows: Obtain the region-specific feature matrix of the second region from the federated transfer learning module. ,in, This indicates the number of samples involved in the application experience of the second region; The feature recombination unit combines the general feature matrix W of the first region with the region-specific feature matrix of the second region. Perform matrix multiplication to generate the fitting feature matrix. ; Based on the adaptation feature matrix The feature recombination unit reconstructs and generates a localized application scheme for the second region. The localized application scheme includes material organization logic, material list, filling guidelines, risk points and avoidance suggestions.

[0015] Compared with existing technologies, this big data-based intelligent matching and application system for fiscal and tax policies has the following advantages: I. This invention establishes a federated transfer learning module, deploying independent regional edge nodes across multiple regions. Each node is independently trained based solely on local policy review approval cases, with original application data always stored locally. Employing knowledge distillation technology, each regional edge node transfers the output probability distribution of its local deep learning teacher model as dark knowledge to a lightweight student network. Through distillation loss function, the lightweight student network learns the review patterns inherent in the teacher models of each region, achieving knowledge extraction and compression of review experience from each region. This forms a lightweight knowledge model that can be transferred across domains. The common features output by the lightweight student network represent general policy knowledge shared across regions, while the region-specific features retained by each node accurately depict the unique review preferences and requirements of each region. Without exposing the original regional data, this invention completes the knowledge distillation and secure transfer of cross-regional policy review experience, enabling the review experience of each region to flow and be shared across domains in the form of knowledge.

[0016] Second, this invention, by setting up a localization adaptation module and employing feature decoupling and recombination technology, allows enterprises to acquire the application experience features accumulated by the enterprise in the first region when expanding from the first region to the second region. The feature decoupling unit then decomposes the application experience feature matrix of the first region into a general feature matrix and a region-specific feature matrix. The general feature matrix represents general knowledge that can be transferred across regions, while the region-specific feature matrix represents region-specific requirements that cannot be transferred. After discarding the region-specific feature matrix, the feature recombination unit merges and reconstructs the general feature matrix with the region-specific features of the second region obtained from the federated transfer learning module, generating complete application features adapted to the second region. Based on these features, the system reconstructs and generates a localized application plan that includes material organization logic, a material list, filling guidelines, risk points, and avoidance suggestions. This enables enterprises to compliantly reuse past experience and obtain directly executable operational plans for the target region.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart illustrating the intelligent matching of fiscal and tax policies in a big data-based intelligent matching and application system. Figure 2 This is a block diagram of the module composition of the intelligent matching and application system for fiscal and tax policies based on big data in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the steps involved in generating an initial candidate policy list in an embodiment of the present invention. Detailed Implementation

[0020] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] To address the experience silos inherent in existing fiscal and tax policy matching systems when applied across regions, resolve the data privacy risks associated with centralized data training, and overcome the limitations of static matching mechanisms in adapting to dynamic changes in regional review preferences, this invention provides a big data-based intelligent matching and application system for fiscal and tax policies. This invention aims to achieve knowledge distillation and secure transfer of cross-regional policy review experience through a federated transfer learning framework, ensuring that original declaration data from each region remains within its local area. Furthermore, through a localization adaptation module, utilizing feature decoupling and recombination technologies, it generates a localized, adapted declaration plan that integrates general experience with regional characteristics when enterprises expand from one region to another. This addresses bottlenecks such as declaration failures and long approval cycles caused by regional implementation differences in cross-regional operations.

[0022] Specifically, such as Figure 2 As shown, the big data-based intelligent matching and application system for fiscal and tax policies includes: a data acquisition and preprocessing module, a policy knowledge graph construction module, an intelligent matching module, a federated transfer learning module, and a localization adaptation module, among which: The data acquisition and preprocessing module is used to collect raw financial and tax policy data and historical declaration data from multiple heterogeneous data sources, and to clean and standardize them. The policy knowledge graph construction module constructs a policy knowledge graph containing policy entities, corporate entities, and their relationships based on cleaned and standardized data. The intelligent matching module is connected to the policy knowledge graph construction module. It is configured to receive the feature information of the target enterprise and perform semantic calculation and matching based on the policy knowledge graph to generate an initial candidate policy list. The federated transfer learning module is a federated learning framework deployed across multiple regional nodes. Each regional node is independently trained based on local policy review and approval cases, and the common features and region-specific features of each region are extracted through knowledge distillation technology to form a lightweight knowledge model that can be transferred across domains. The localization adaptation module receives application experience features accumulated in the first region when an enterprise expands from the first region to the second region. It uses feature decoupling technology to remove the region-specific parts of the application experience features accumulated in the first region and overlays the region-specific features of the second region obtained from the federated transfer learning module to generate a localized adaptation version of the application plan for the second region.

[0023] In its implementation, the data acquisition and preprocessing module forms the data foundation for the system's knowledge construction and intelligent matching. This module collects raw data from multiple sources through standardized data interfaces. These sources include, but are not limited to: structured / semi-structured policy document databases storing the original policy documents; historical application archives of enterprises recording past application records and approval results; and internal ERP systems connected after authorization by the enterprise. The collected data types cover structured enterprise information and financial indicator data, semi-structured XML-formatted policy document data, and unstructured policy text and review opinion data. To ensure the quality of subsequent processing, the cleaning and standardization processes include data cleaning of the collected data, such as removing duplicate data, handling missing values, and correcting outliers; entity recognition, such as using a pre-trained named entity recognition model to identify key entities in the text, such as policy names, issuing agencies, enterprise names, and industry categories; and format unification and unit of measurement standardization.

[0024] Based on data preprocessing, the policy knowledge graph construction module is responsible for transforming fragmented and heterogeneous data into structured knowledge. This module utilizes natural language processing technology to extract entities and relationships from the cleaned and standardized data. Entity extraction types include policy entities, enterprise entities, industry entities, and regional entities. Relationship extraction types include: policy issued by a certain institution, policy applicable to a certain region, enterprise having previously applied for a certain policy, and application records being rejected. Based on the extracted entities and relationships, this module constructs a knowledge graph structure containing nodes and edges. Nodes correspond to the aforementioned types of entities, and edges correspond to various relationships between entities. The completed knowledge graph is stored in a graph database to support subsequent efficient semantic queries and path reasoning.

[0025] The intelligent matching module, based on the characteristics of the target enterprise, accurately filters out initially matching policies from the knowledge graph, such as... Figure 3 As shown, the process of generating an initial candidate policy list by performing semantic calculation and matching based on the policy knowledge graph is as follows: Receive the characteristic information of the target enterprise, which includes the target enterprise's industry category, revenue scale, R&D investment ratio and registered region; The characteristic information of the target enterprise is transformed into a feature vector. Extract the policy feature vectors corresponding to each candidate policy entity from the policy knowledge graph. ,in An index representing a candidate policy entity; Calculate the feature vector With each of the policy feature vectors semantic similarity score between The semantic similarity score ,in, Represents the cosine similarity function. Represents the feature vector With the policy feature vector dot product, Represents the feature vector The length of the mold, Represents the policy feature vector The modulus length; The candidate policy entities are sorted from highest to lowest according to their semantic similarity scores, and the top K candidate policy entities are selected to generate the initial candidate policy list.

[0026] The aforementioned federated transfer learning module deploys a federated learning framework consisting of a central coordination server and multiple regional edge nodes. Each regional edge node corresponds to a geographical region (Region 1, Region 2, etc.). Internally, each node stores its local raw application data, including company information, application materials, and review results, and runs a local deep learning teacher model. This teacher model is a deep neural network used to learn the review patterns of the reviewers in its region. The input is the application characteristics of the company, and the output is the probability of approval. The central coordination server communicates with all regional edge nodes, coordinating the training process of each node, aggregating common feature knowledge uploaded by each node, and distributing global model parameters, but it does not access any raw data.

[0027] In practical implementation, common and region-specific features of each region are extracted using knowledge distillation techniques to form a lightweight knowledge model that can be transferred across domains. The process is as follows: Each edge node in the region uses locally approved case data to train a deep learning teacher model. After training, in order to extract its knowledge and compress it into a lightweight model, each node uses knowledge distillation technology to pass the output probability distribution of the teacher model as dark knowledge to the lightweight student network. The loss function for training the student network includes two parts: one part is the difference between its prediction results and the true labels, and the other part is the distillation loss. This is used to measure the difference between the output probability distributions of the teacher model and the student network, specifically expressed as: ,in, This represents the total number of samples of local policy approval cases in the edge nodes of this region. For the deep learning teacher model to the first The probability distribution vector of each sample output. For the lightweight student network to the first The probability distribution vector of each sample output, KL The KL divergence function measures the difference between two probability distribution vectors. By minimizing the total loss function (including task loss and distillation loss), the lightweight student network learns the review patterns inherent in the deep learning teacher model. Through this process, the common features output by the student networks from different regions represent shared policy review knowledge across regions, while the parts retained by the distillation process that cannot be fitted by data from other regions form the region-specific features retained by each node, accurately depicting the unique review preferences and requirements of each region. Each node only uploads the parameters of the lightweight student network or the extracted common features to the central coordination server, achieving knowledge distillation and secure transfer of cross-regional policy review experience without exposing the original regional data.

[0028] The localization adaptation module is responsible for generating operational application plans when enterprises expand across regions. When an enterprise plans to expand from a first region (a region with accumulated successful experience) to a second region (a new target region), it receives application experience features accumulated by the enterprise in the first region, or personalized application features generated by nodes in the first region based on their historical successful cases. To effectively transfer these experiences to the second region, a feature decoupling unit is first used to perform matrix decomposition on the application experience features from the first region. In this embodiment, the feature decoupling unit uses a non-negative matrix decomposition algorithm. Specifically, the application experience features from the first region are represented as an experience feature matrix. ,in, This indicates the number of feature dimensions representing the application experience of the first region. This represents the number of samples (i.e., the number of historical successful cases) involved in the application experience of the first region. The feature decoupling unit decomposes the experience feature matrix using a non-negative matrix factorization algorithm. Decomposed into the general feature matrix of the first region The specific feature matrix of the first region The product of, where, For the hidden feature dimension, and After decomposition, the general feature matrix of the first region is... This represents general knowledge for policy review shared across regions, while the first region's specific feature matrix... This represents the unique review preferences and requirements of the first region.

[0029] After completing the decomposition, the feature decoupling unit discards the feature matrix representing the region-specific part. Only the feature matrix representing general knowledge is retained. Subsequently, the feature reorganization unit obtains the region-specific features of the second region from the federated transfer learning module, which are represented as a feature matrix. ,in, This matrix represents the number of samples involved in the application experience of the second region (the number of cases used to characterize the review preferences of the second region). In the federated transfer learning process, the feature recombining unit, independently distilled and securely shared by the second-region nodes, represents the unique review preferences and requirements of the second region and incorporates the general feature matrix. Region-specific feature matrix of the second region Perform matrix multiplication to generate the adaptation feature matrix for the second region. General application knowledge ( ) and the specific audit rules for the target region ( By combining these elements, a complete set of declaration features that meet the requirements of the second region can be reconstructed.

[0030] Finally, based on the adaptation feature matrix The feature reorganization unit uses a feature-to-text generation model to reconstruct and generate a localized application plan for the second region. This plan includes: material organization logic: guiding enterprises on how to organize the chapter order and focus of application materials according to the preferences of the second region; material list: listing the specific documents, certificates, forms and their format requirements required for review in the second region; filling guidance: providing detailed explanations and examples for the filling items, terminology and data standards unique to the application system in the second region; risk points and avoidance suggestions: analyzing common problems that may be encountered in the application process based on historical rejection cases in the second region, and providing specific avoidance strategies.

[0031] Through the collaborative work of each module, the system of this invention can realize the intelligent and secure migration and reuse of cross-regional policy application experience while ensuring data privacy in each region. It provides enterprises with full-chain intelligent support from policy matching to localized solution generation, thereby improving the efficiency and success rate of enterprises' cross-regional development.

[0032] In the specific implementation process, such as Figure 1 As shown, the specific process of intelligent matching of fiscal and tax policies in the big data-based fiscal and tax policy intelligent matching system provided by this invention is as follows: (1) Data acquisition and preprocessing: The system collects raw fiscal and tax policy data and enterprise historical data from multiple data sources.

[0033] The collected data is cleaned to remove noise and handle missing values.

[0034] Key entities such as policies and enterprises are identified and extracted, and the data is standardized in terms of format and measurement units to form standardized data that can be used for subsequent analysis.

[0035] (2) Constructing a policy knowledge graph Using natural language processing technology, entities such as enterprises, industries, and regions, as well as the relationships between these entities, are automatically extracted from normalized data.

[0036] Using the extracted entities as nodes and relationships as edges, a structured policy knowledge graph is constructed and stored in a graph database.

[0037] (3) Receiving target company information When a company uses the system, it inputs or has its own characteristic information retrieved by the system.

[0038] The system transforms the characteristic information of the target company into mathematical vectors.

[0039] In the policy knowledge graph, each candidate policy is also transformed into a corresponding feature vector.

[0040] Calculate the semantic similarity between the enterprise vector and each policy vector, and sort all candidate policies from high to low based on the similarity score.

[0041] The system automatically selects the policies with the highest semantic similarity ranking, generates an initial list of candidate policies, and recommends them to the target company.

[0042] (4) Federal learning and knowledge distillation A federated learning framework was deployed in each regional node. Each regional node independently trained a local deep learning model using local policy approval cases. The raw data was always stored locally and not shared with external parties.

[0043] Knowledge extraction: Each regional node uses knowledge distillation technology to compress and extract the complex auditing rules contained in the local large-scale teacher model into a lightweight student network. The student network produces two parts of knowledge: general features that can be shared across regions and specific features unique to this region.

[0044] (5) Trigger cross-region adaptation request When a business has already successfully applied for business in the first region and plans to expand to the second region, it submits a localization adaptation request to the system for the second region.

[0045] The system retrieves the characteristics of the company's successful application experience accumulated in the first region.

[0046] By using feature decoupling techniques, the empirical feature is decomposed into: a transferable general feature and a non-transferable first region-specific feature.

[0047] The system discards specific features that apply only to the first region.

[0048] The general features extracted from enterprise experience will be fused and reconstructed with the region-specific features of the second region obtained from the federated learning module.

[0049] Based on the new features after integration, the system automatically reconstructs and generates a localized application plan specifically tailored for the second region.

[0050] This plan is a ready-to-use guide that includes the material organization logic adapted to the requirements of the second region, a detailed list, filling guidelines, potential risk points, and avoidance suggestions.

[0051] In summary, this invention, by setting up a federated transfer learning module and employing knowledge distillation technology, enables the secure cross-domain flow and sharing of audit experience across different regions. By setting up a localization adaptation module and employing feature decoupling and recombination technology, it allows enterprises to compliantly reuse past experience and obtain directly executable operational solutions for target regions. This system design effectively solves the problems of data privacy risks and static matching mechanisms, providing efficient, secure, and accurate financial and tax policy application support for enterprises operating across regions.

[0052] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A big data-based intelligent matching and application system for fiscal and tax policies, characterized in that: The system includes: The data acquisition and preprocessing module collects raw financial and tax policy data and historical declaration data from multiple heterogeneous data sources, and performs cleaning and standardization processing. The policy knowledge graph construction module constructs a policy knowledge graph containing policy entities, corporate entities, and their relationships based on cleaned and standardized data. The intelligent matching module is connected to the policy knowledge graph construction module and is configured to receive the feature information of the target enterprise and perform semantic calculation and matching based on the policy knowledge graph to generate an initial candidate policy list. The federated transfer learning module is a federated learning framework deployed across multiple regional nodes. Each regional node is independently trained based on local policy review and approval cases, and the common features and region-specific features of each region are extracted through knowledge distillation technology to form a lightweight knowledge model that can be transferred across domains. The localization adaptation module receives the application experience features accumulated in the first region when an enterprise expands from the first region to the second region. It uses feature decoupling technology to remove the region-specific parts of the application experience features accumulated in the first region and superimposes the region-specific features of the second region obtained from the federated transfer learning module to generate a localized adaptation version of the application scheme for the second region.

2. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, In the data acquisition and preprocessing module, the multi-source heterogeneous data sources include policy document databases, enterprise historical declaration archives, and authorized enterprise internal ERP systems. The data types collected include structured enterprise information and financial indicator data, semi-structured XML format policy document data, and unstructured policy text and review opinion data. The cleaning and standardization process includes cleaning the collected data to remove noise and handle missing values, performing entity identification to identify policy names, issuing agencies and company names, and standardizing the format and measurement units.

3. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, The process by which the policy knowledge graph construction module constructs the policy knowledge graph is as follows: Natural language processing technology is used to extract entities and relationships from the data after it has been cleaned and standardized by the data collection and preprocessing module. The extracted entity types include policy entities, enterprise entities, industry entities, and regional entities. Based on the extracted entities and relationships, a knowledge graph structure containing nodes and edges is constructed, where nodes correspond to the entities and edges correspond to the relationships between the entities. The constructed knowledge graph is then stored in a graph database.

4. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, The intelligent matching module performs semantic calculations and matching based on the policy knowledge graph to generate an initial candidate policy list as follows: Receive the characteristic information of the target enterprise, which includes the target enterprise's industry category, revenue scale, R&D investment ratio and registered region; The characteristic information of the target enterprise is transformed into a feature vector. Extract the policy feature vectors corresponding to each candidate policy entity from the policy knowledge graph. ,in An index representing a candidate policy entity; Calculate the feature vector With each of the policy feature vectors semantic similarity score between The semantic similarity score ,in, Represents the cosine similarity function. Represents the feature vector With the policy feature vector dot product, Represents the feature vector The length of the mold, Represents the policy feature vector The modulus length; The candidate policy entities are sorted from highest to lowest according to their semantic similarity scores, and the top K candidate policy entities are selected to generate the initial candidate policy list.

5. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, The federated transfer learning module includes: Multiple regional edge nodes, each corresponding to a geographical region, each regional edge node stores local raw declaration data and runs local deep learning teacher models; The central coordination server communicates with the multiple regional edge nodes and is configured to coordinate the deep learning teacher model training process of each regional edge node, aggregate the common feature knowledge uploaded by each regional edge node, and distribute global model parameters.

6. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, The knowledge distillation technique is as follows: Each region's edge nodes use the output probability distribution of the deep learning teacher model as dark knowledge. This dark knowledge is then passed to a lightweight student network via a distillation loss function. The output of the lightweight student network represents both the common features and the region-specific features. The distillation loss function is expressed as follows: ,in, This represents the total number of samples of locally approved policy cases in the edge nodes of the region. For the deep learning teacher model to the first The probability distribution vector of each sample output. For the lightweight student network to the first The probability distribution vector of each sample output, KL The KL divergence function is used to measure the difference between two probability distribution vectors. By minimizing the distillation loss function, the lightweight student network learns the auditing rules inherent in the deep learning teacher model.

7. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 1, characterized in that, The localization adaptation module includes: The feature decoupling unit performs matrix decomposition on the application experience features from the first region, decomposing them into a general feature matrix of the first region and a specific feature matrix of the first region. The feature recombination unit, connected to the feature decoupling unit, is configured to discard the first region-specific feature matrix and fuse and reconstruct the first region general feature matrix with the region-specific features of the second region obtained from the federated transfer learning module to generate the localized adapted application scheme.

8. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 7, characterized in that, The process by which the feature decoupling unit performs matrix decomposition on the declared experience features from the first region is as follows: The application experience characteristics of the first region are used as the experience feature matrix. ,in, This indicates the number of feature dimensions representing the application experience of the first region. This indicates the number of samples involved in the application experience of the first region; The feature decoupling unit uses a nonnegative matrix factorization algorithm to decompose the empirical feature matrix. Decomposed into the general feature matrix of the first region The specific feature matrix of the first region The product of, where For the hidden feature dimension, and ; The general feature matrix of the first region This represents general knowledge for policy review shared across regions, and the specific feature matrix of the first region. This represents the unique review preferences and requirements of the first region.

9. The intelligent matching and application system for fiscal and tax policies based on big data as described in claim 7, characterized in that, The process of fusion and reconstruction of the feature recombination units is as follows: Obtain the region-specific feature matrix of the second region from the federated transfer learning module. ,in, This indicates the number of samples involved in the application experience of the second region; The feature recombination unit combines the general feature matrix W of the first region with the region-specific feature matrix of the second region. Perform matrix multiplication to generate the fitting feature matrix. ; Based on the adaptation feature matrix The feature reorganization unit reconstructs and generates a localized application scheme for the second region. The localized application scheme includes material organization logic, material list, filling guidelines, risk points and avoidance suggestions.

Images