Taxation preferential policy automatic matching method and system based on training model, and medium

Abstract

The invention discloses a tax preferential policy automatic matching method and system based on a training model, and a medium, mainly relates to the technical field of policy matching, and is used for solving the problems of single and fixed rule matching relation and difficulty in rule updating of an existing tax matching scheme. Comprising the following steps: determining a window assignment size corresponding to each policy representation sequence; determining the window size of an attention mechanism corresponding to the policy representation sequence according to the size of the window assignment, and updating the policy representation sequence table through the attention mechanism; acquiring policy interaction data in office software related to the current policy matching organization; inputting a plurality of pieces of label-free data into a self-supervised learning ability model, and pre-training the self-supervised learning ability model to obtain a pre-trained model; and inputting the updated policy representation sequence table and the policy interaction data into a pre-training model to finish fine adjustment of the pre-training model so as to obtain a trained training model.

Description

Technical Field

[0001] This application relates to the technical field of automatic matching of tax preferential policies, and particularly relates to a method, system and medium for automatic matching of tax preferential policies based on a training model. Background Art

[0002] Since tax preferential policies usually have many preconditions, cover a wide range of aspects, and are highly professional, only some professional staff in enterprises can obtain and understand relevant information. However, in the current enterprise information management mode (especially in the sharing mode), tax professional positions and sales business positions are usually separated. Business positions are not familiar with tax policies, and tax positions do not understand the actual business situation. In order to apply tax preferential policies reasonably and legally, only a large number of offline communications can be used to assist in judgment, which is very prone to errors and omissions. In particular, the current tax preference application adopts the mode of "self-judgment, self-declaration, and post-event supervision". Although it provides greater flexibility to enterprises, it also puts higher requirements on the tax management and risk monitoring of enterprises. Enterprises need to independently identify available preferential policies and avoid policy abuse.

[0003] The existing tax matching solutions mainly include: implementing based on enterprise-defined matching rules. First, map rules for policies and tax-related business types are set according to the enterprise's understanding of the policies. Subsequently, mapping rules for tax-related business types and tax-related basic data are set and persisted in the database. In actual business, fixed tax-related business types are found based on business data, and then the corresponding preferential policies are matched.

[0004] However, since the custom rules are fixed rules with a single matching relationship, if maintained too meticulously, it is necessary to exhaust business types or tax-related data. Therefore, the fixed-stored and matched custom rules cannot enable enterprises to accurately apply preferential policies. In addition, in order to reduce the tax burden on enterprises, the tax bureau continuously issues new tax preferential policies and also adjusts or extends the original preferential policies. Each time there is an update, the original custom rules need to be updated and approved. The entire adjustment process has a high labor cost and may also be updated in a timely manner, resulting in the failure to match tax preferential policies during a certain period. Summary of the Invention

[0005] In view of the above deficiencies of the prior art, this application provides a method, system and medium for automatic matching of tax preferential policies based on a training model to solve the problems of single and fixed matching relationships of existing tax matching solutions and difficult rule updates.

[0006] In a first aspect, this application provides a method for automatic matching of tax preferential policies based on a training model, the method including: Obtain tax preference policies and interpretation documents, and extract the policy characterization sequence list; among them, the policy characterization sequence list is composed of several policy characterization sequences; Based on whether the preset high-weight keywords exist in the current policy characterization sequence and the weight interval to which the policy keywords existing in the current policy characterization sequence belong, determine the window assignment size corresponding to each policy characterization sequence; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy characterization sequence, and then update the policy characterization sequence list through the attention mechanism; Obtain the policy interaction data in the office software involved in the current policy matching organization; among them, the policy interaction data at least includes: operation inquiry data and policy rule notification data involved in the policy rules; Input a number of unlabeled data into the self-supervised learning ability model to pre-train the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy characterization sequence list and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model to obtain a trained training model; obtain new tax preference policies and interpretation documents, and then extract the corresponding policy characterization sequence list; then input the policy characterization sequence list into the trained training model to obtain the policy interaction data corresponding to the current policy matching organization.

[0007] In an implementation manner of the present application, obtaining tax preference policies and interpretation documents specifically includes: Obtain tax preference policies and interpretation documents by means of web crawlers; Read whether there is a preset file interpretation program at the preset address. When there is a preset interpretation program, interpret the tax preference policies and interpretation documents, and use the interpreted documents as the tax preference policies and interpretation documents; When there is no preset interpretation program, obtain the externally added preset file interpretation program through the input interface corresponding to the preset address.

[0008] In an implementation manner of the present application, extracting the policy characterization sequence list specifically includes: Use a recurrent neural network to perform sequence modeling on the tax preference policies and interpretation documents to capture the temporal dependence relationship between each piece of data in the tax preference policies and interpretation documents; Combine a convolutional neural network to perform paragraph feature extraction on the text to capture the keywords and key phrases corresponding to each paragraph; Based on the sequence order between paragraphs, store the temporal dependence relationship, keywords and key phrases in a paragraph as a policy characterization sequence in the policy characterization sequence list.

[0009] In an implementation manner of the present application, based on whether a preset high-weight keyword exists in the current policy representation sequence and the weight interval to which the policy keyword existing in the current policy representation sequence belongs, the window assignment size corresponding to each policy representation sequence is determined, specifically including: When a preset high-weight keyword exists in the current policy representation sequence, it is determined that the policy representation sequence belongs to a preset high-weight interval; Furthermore, based on the number of existing preset high-weight keywords, the sorting situation of each policy representation sequence in the preset high-weight interval is determined. Then, based on the first assignment range of the preset window assignment corresponding to the preset high-weight interval and the number of policy representation sequences in the preset high-weight interval, the window assignment is performed in a decreasing manner; among them, the window assignments of the policy representation sequences in the preset high-weight interval are all within the first assignment range; When the preset high-weight keyword does not exist in the current policy representation sequence, it is determined that the policy representation sequence belongs to a preset low-weight interval; Based on the total number of time-dependent relationships, keywords, and key phrases included in the policy representation sequences in the preset low-weight interval, the sorting situation of each policy representation sequence in the preset low-weight interval is determined according to the order of the total number. Then, based on the second assignment range of the preset window assignment corresponding to the preset low-weight interval and the number of policy representation sequences in the preset low-weight interval, the window assignment is performed in a decreasing manner; among them, the window assignments of the policy representation sequences in the preset low-weight interval are all within the second assignment range.

[0010] In an implementation manner of the present application, the new tax preference policies and interpretation documents are obtained, and then the corresponding policy representation sequence list is extracted, specifically including: The new tax preference policies and interpretation documents are obtained, and the policy representation sequence list is extracted; Based on whether a preset high-weight keyword exists in the policy representation sequence and the weight interval to which the policy keyword existing in the current policy representation sequence belongs, the window assignment size corresponding to each policy representation sequence is determined; according to the window assignment size, the window size of the attention mechanism corresponding to the policy representation sequence is determined, and then the policy representation sequence list is updated through the attention mechanism.

[0011] In an implementation manner of the present application, after obtaining the policy interaction data corresponding to the current policy matching organization, the method further includes: obtaining the tax preference policies queried by the business personnel through a preset service operation interface, and then feedbacking the corresponding policy interaction data.

[0012] In a second aspect, the present application provides a tax preference policy automatic matching system based on a training model, and the system includes: An extraction module, configured to obtain tax preference policies and interpretation documents, and extract a policy representation sequence list; wherein, the policy representation sequence list is composed of a plurality of policy representation sequences; An update module, configured to determine the window assignment size corresponding to each policy representation sequence based on whether a preset high-weight keyword exists in the current policy representation sequence and the weight interval to which the policy keyword exists in the current policy representation sequence; determine the window size of the attention mechanism corresponding to the policy representation sequence according to the window assignment size, and then update the policy representation sequence list through the attention mechanism; An acquisition module, configured to acquire policy interaction data in the office software involved in the current policy matching organization; wherein, the policy interaction data at least includes: operation inquiry data involved in policy rules, policy rule notification data; An obtaining module, configured to input a number of unlabeled data into a self-supervised learning ability model, pre-train the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy representation sequence list and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model and obtain a trained training model; acquire a new tax preference policy and its interpretation document, and then extract the corresponding policy representation sequence list; then input the policy representation sequence list into the trained training model to obtain the policy interaction data corresponding to the current policy matching organization.

[0013] In an implementation manner of the present application, the extraction module includes a first extraction unit, configured to obtain a tax preference policy and its interpretation document by means of a crawler; read whether a preset file interpretation program exists at a preset address, and when the preset interpretation program exists, interpret the tax preference policy and its interpretation document, and use the interpreted document as the tax preference policy and its interpretation document; when the preset interpretation program does not exist, obtain a preset file interpretation program added externally through an input interface corresponding to the preset address.

[0014] In an implementation manner of the present application, the extraction module includes a second extraction unit, configured to perform sequence modeling on the tax preference policy and its interpretation document by using a recurrent neural network to capture the temporal dependence relationship between each piece of data in the tax preference policy and its interpretation document; combine a convolutional neural network to extract paragraph features of the text and capture the keywords and key phrases corresponding to each paragraph; Based on the sequence order between paragraphs, store the temporal dependence relationship, keywords and key phrases located in a paragraph as a policy representation sequence in the policy representation sequence list.

[0015] In an implementation manner of the present application, the update module includes an update unit, configured to determine that the policy representation sequence belongs to a preset high-weight interval when a preset high-weight keyword exists in the current policy representation sequence; Furthermore, based on the number of preset high-weight keywords, determine the sorting of each policy representation sequence in the preset high-weight interval, and then perform window assignment in a decreasing manner based on the first assignment range of the preset window assignment corresponding to the preset high-weight interval and the number of policy representation sequences in the preset high-weight interval; among them, the window assignments of the policy representation sequences in the preset high-weight interval are all within the first assignment range. If there is no preset high-weight keyword in the current policy representation sequence, it is determined that the policy representation sequence belongs to the preset low-weight interval. Based on the total number of temporal dependence relationships, keywords, and key phrases included in the policy representation sequences within the preset low-weight interval, determine the sorting of each policy representation sequence in the preset low-weight interval in the order of the total quantity sum, and then perform window assignment in a decreasing manner based on the second assignment range of the preset window assignment corresponding to the preset low-weight interval and the number of policy representation sequences in the preset low-weight interval; among them, the window assignments of the policy representation sequences in the preset low-weight interval are all within the second assignment range.

[0016] In a third aspect, the present application provides a non-volatile computer storage medium, on which computer instructions are stored, and when the computer instructions are executed, a method for automatically matching tax preference policies based on a training model as described in any one of the above is implemented.

[0017] Those skilled in the art can understand that the present application has at least the following beneficial effects: The present application proposes a method, system, and medium for automatically matching tax preference policies based on a training model. By constructing a dynamically updated policy representation sequence table and introducing an attention mechanism, the technical defects of traditional fixed-rule matching methods are solved. Specifically, (1) adopting a dynamic window assignment mechanism based on weighted keywords enables the model to adaptively capture key features in policy texts, overcoming the problem of a single artificial rule matching relationship; (2) through the self-supervised pre-training and fine-tuning mechanism that integrates policy interaction data, the model has the ability to continuously learn, significantly reducing the manual maintenance cost during policy updates; (3) using the training model to automatically process newly added policy files ensures the timeliness and accuracy of preferential policy matching, avoiding the problem of omission of policy application caused by delayed rule updates. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings required for description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1It is a flowchart of an automatic matching method for tax preferential policies based on a training model provided by an embodiment of the present application.

[0020] Figure 2 It is a schematic diagram of the internal structure of an automatic matching system for tax preferential policies based on a training model provided by an embodiment of the present application. Detailed implementation manners

[0021] Those skilled in the art should understand that the embodiments described below are only the preferred embodiments of the present disclosure, and do not mean that the present disclosure can only be implemented through these preferred embodiments. These preferred embodiments are only used to explain the technical principles of the present disclosure and are not used to limit the protection scope of the present disclosure. Based on the preferred embodiments provided by the present disclosure, all other embodiments obtained by those of ordinary skill in the art without creative efforts should still fall within the protection scope of the present disclosure.

[0022] It should also be noted that the term "including", "comprising" or any other variant thereof is intended to cover a non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, commodity or device. Without further limitation, the element defined by the statement "including one..." does not exclude the existence of additional identical elements in the process, method, commodity or device including the element.

[0023] The technical solutions proposed in the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0024] The embodiment provides an automatic matching method for tax preferential policies based on a training model. As Figure 1 shown, the method provided by the embodiment of the present application mainly includes the following steps: Step 110: Obtain tax preferential policies and interpretation documents, and extract a policy characterization sequence list.

[0025] It should be noted that the policy characterization sequence list is composed of several policy characterization sequences.

[0026] In some embodiments, obtaining tax preferential policies and interpretation documents specifically includes: Obtaining tax preferential policies and interpretation documents by means of web crawling; reading whether there is a preset file interpretation program at a preset address. When there is a preset interpretation program, interpreting the tax preferential policies and interpretation documents, and using the interpreted documents as the tax preferential policies and interpretation documents; when there is no preset interpretation program, obtaining a preset file interpretation program added externally through the input interface corresponding to the preset address.

[0027] Exemplarily, extract the policy representation sequence table, specifically including: Use a recurrent neural network to perform sequence modeling on tax preference policies and interpretation documents, capturing the temporal dependence relationship between each piece of data in the tax preference policies and interpretation documents; combine a convolutional neural network to extract paragraph features from the text, capturing the keywords and key phrases corresponding to each paragraph; based on the order of the paragraphs, store the temporal dependence relationship, keywords, and key phrases in a paragraph as a policy representation sequence in the policy representation sequence table.

[0028] Those skilled in the art can understand that this step ensures that the model can obtain the latest and most accurate tax preference policy information. Automatically obtaining files through web crawling improves the efficiency and timeliness of data acquisition. At the same time, according to the existence of a preset interpretation program, the interpretation method is flexibly selected to ensure the readability and usability of the data. Using a recurrent neural network and a convolutional neural network to perform sequence modeling and feature extraction on the files can capture the temporal dependence relationship and keywords in the policy text, forming a policy representation sequence, providing a basis for subsequent matching.

[0029] Step 120: Determine the window assignment size corresponding to each policy representation sequence based on whether a preset high-weight keyword exists in the current policy representation sequence and the weight interval to which the policy keyword in the current policy representation sequence belongs; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy representation sequence, and then update the policy representation sequence table through the attention mechanism.

[0030] Among them, determining the window assignment size corresponding to each policy representation sequence based on whether a preset high-weight keyword exists in the current policy representation sequence and the weight interval to which the policy keyword in the current policy representation sequence belongs specifically includes: When there are preset high-weight keywords in the current policy representation sequence, it is determined that the policy representation sequence belongs to a preset high-weight interval; furthermore, based on the number of preset high-weight keywords present, the sorting of each policy representation sequence in the preset high-weight interval is determined, and then, based on the first assignment range of the preset window assignment corresponding to the preset high-weight interval and the number of policy representation sequences in the preset high-weight interval, window assignment is performed in a decreasing manner; wherein, the window assignments of the policy representation sequences in the preset high-weight interval are all within the first assignment range; when there are no preset high-weight keywords in the current policy representation sequence, it is determined that the policy representation sequence belongs to a preset low-weight interval; based on the total number of temporal dependence relationships, keywords, and key phrases included in the policy representation sequences in the preset low-weight interval, the sorting of each policy representation sequence in the preset low-weight interval is determined in the order of the total number, and then, based on the second assignment range of the preset window assignment corresponding to the preset low-weight interval and the number of policy representation sequences in the preset low-weight interval, window assignment is performed in a decreasing manner; wherein, the window assignments of the policy representation sequences in the preset low-weight interval are all within the second assignment range.

[0031] Those skilled in the art can understand that this step improves the model's attention to key information by introducing an attention mechanism. According to the presence or absence and quantity of preset high-weight keywords, the window assignment size of the policy representation sequence is determined, and then the window size of the attention mechanism is adjusted, enabling the model to focus more on important policy information. This way of dynamically adjusting the window size enhances the flexibility and adaptability of the model and improves the matching accuracy.

[0032] Step 130: Obtain policy interaction data in the office software involved in the current policy matching organization; wherein, the policy interaction data at least includes: operation inquiry data related to policy rules and policy rule notification data.

[0033] Those skilled in the art can understand that this step provides rich actual application scenario data for the model by obtaining policy interaction data in the office software. This data includes operation inquiry data and notification data related to policy rules, which can reflect the situations and problems of policies in actual applications. Using this data as one of the inputs for model training helps the model better understand and adapt to actual business requirements and improves the practicality and accuracy of matching.

[0034] Step 140: Input a number of unlabeled data into the self-supervised learning ability model for pre-training the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy representation sequence list and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model and obtain a trained training model; obtain the newly added tax incentive policies and interpretation documents, and then extract the corresponding policy representation sequence list; further input the policy representation sequence list into the trained training model to obtain the policy interaction data corresponding to the current policy-matching organization.

[0035] Those skilled in the art can understand that this step reduces the dependence on a large amount of labeled data by introducing a self-supervised learning ability model. First, use unlabeled data for initial training to obtain a preliminary training model; then input the updated policy representation sequence list and policy interaction data into the preliminary training model for fine-tuning to obtain a trained training model. This method not only improves the training efficiency of the model but also enhances the generalization ability of the model, enabling it to better adapt to new tax incentive policies.

[0036] In some embodiments, obtaining the newly added tax incentive policies and interpretation documents and then extracting the corresponding policy representation sequence list specifically includes: Obtain the newly added tax incentive policies and interpretation documents and extract the policy representation sequence list; Based on whether the preset high-weight keywords exist in the policy representation sequence and the weight interval to which the policy keywords in the current policy representation sequence belong, determine the window assignment size corresponding to each policy representation sequence; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy representation sequence, and then update the policy representation sequence list through the attention mechanism.

[0037] Those skilled in the art can understand that by automatically obtaining and extracting the policy representation sequence list, the rapid structured processing of the newly added tax incentive policy documents is realized; based on the window assignment mechanism of the weight keywords, the attention window size can be intelligently adjusted to improve the extraction accuracy of policy key information; by adopting the weight interval determination mechanism, the importance difference of policy keywords can be effectively identified to improve the accuracy of policy matching; by dynamically updating the policy representation through the attention mechanism, the feature expression requirements of different policy documents can be adapted.

[0038] In some embodiments, after obtaining the policy interaction data corresponding to the current policy-matching organization, the method further includes: Obtain the tax incentive policies queried by business personnel through a preset business operation interface, and then feedback the corresponding policy interaction data.

[0039] In addition, this application Figure 2 This is a tax incentive policy automatic matching system provided by an embodiment of this application. AsFigure 2 As shown in Figure 2 , the system provided by the embodiment of the present application mainly includes: An extraction module 210, configured to obtain tax preference policies and interpretation documents, and extract a policy characterization sequence table; wherein, the policy characterization sequence table is composed of a number of policy characterization sequences.

[0040] The extraction module 210 includes a first extraction unit, configured to obtain tax preference policies and interpretation documents by means of web crawlers; read whether there is a preset document interpretation program at a preset address, and when there is a preset interpretation program, perform the interpretation of tax preference policies and interpretation documents, and use the interpreted documents as tax preference policies and interpretation documents; when there is no preset interpretation program, obtain a preset document interpretation program added externally through an input interface corresponding to the preset address.

[0041] The extraction module 210 includes a second extraction unit, configured to perform sequence modeling on tax preference policies and interpretation documents by using a recurrent neural network, capture the temporal dependence relationship between each piece of data in the tax preference policies and interpretation documents; combine a convolutional neural network to extract paragraph features of the text, and capture keywords and key phrases corresponding to each paragraph; based on the order of paragraphs, store the temporal dependence relationship, keywords, and key phrases in a paragraph as a policy characterization sequence in the policy characterization sequence table.

[0042] An update module 220, configured to determine the window assignment size corresponding to each policy characterization sequence based on whether a preset high-weight keyword exists in the current policy characterization sequence and the weight interval to which the policy keyword belongs in the current policy characterization sequence; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy characterization sequence, and then update the policy characterization sequence table through the attention mechanism.

[0043] An acquisition module 230, configured to acquire policy interaction data in the office software involved in the current policy matching organization; wherein, the policy interaction data at least includes: operation inquiry data and policy rule notification data involved in the policy rules.

[0044] An obtaining module 240, configured to input a number of unlabeled data into a self-supervised learning ability model to pre-train the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy characterization sequence table and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model to obtain a trained training model; acquire new tax preference policies and interpretation documents, and then extract the corresponding policy characterization sequence table; then input the policy characterization sequence table into the trained training model to obtain the policy interaction data corresponding to the current policy matching organization.

[0045] In addition, an embodiment of the present application further provides a non-volatile computer storage medium, on which executable instructions are stored, and when the executable instructions are executed, the automatic matching method for tax preferential policies based on a training model as described above is implemented.

[0046] So far, the technical solutions of the present disclosure have been described in combination with multiple foregoing embodiments. However, it is easy for those skilled in the art to understand that the protection scope of the present disclosure is not limited to these specific embodiments. Without departing from the technical principle of the present disclosure, those skilled in the art can split and combine the technical solutions in the foregoing embodiments, and can also make equivalent changes or replacements to relevant technical features. Any changes, equivalent replacements, improvements, etc. made within the technical concept and / or technical principle of the present disclosure will fall within the protection scope of the present disclosure.

Claims

1. An automatic matching method for tax preference policies based on a training model, characterized in that, The method includes: Obtain tax preferential policies and interpretation documents, and extract a policy representation sequence list; wherein, the policy representation sequence list is composed of a number of policy representation sequences; Based on whether a preset high-weight keyword exists in the current policy representation sequence, and the weight interval to which the policy keyword belongs in the current policy representation sequence, determine the window assignment size corresponding to each policy representation sequence; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy representation sequence, and then update the policy representation sequence list through the attention mechanism; Obtain the policy interaction data in the office software involved in the current policy matching organization; wherein, the policy interaction data at least includes: operation inquiry data related to policy rules, policy rule notification data; Input a number of unlabeled data into the self-supervised learning ability model, pre-train the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy representation sequence list and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model and obtain a trained training model; Obtain the newly added tax preferential policies and interpretation documents, and then extract the corresponding policy representation sequence list; then input the policy representation sequence list into the trained training model to obtain the policy interaction data corresponding to the current policy matching organization.

2. The automatic matching method for tax preference policies based on a training model according to claim 1, wherein Obtain tax preferential policies and interpretation documents, specifically including: Obtain tax preferential policies and interpretation documents by means of web crawlers; Read whether there is a preset file interpretation program at a preset address. When there is a preset interpretation program, interpret the tax preferential policies and interpretation documents, and use the interpreted documents as the tax preferential policies and interpretation documents; When there is no preset interpretation program, obtain the externally added preset file interpretation program through the input interface corresponding to the preset address.

3. The automatic matching method for tax preference policies based on a training model according to claim 1, characterized in that Extract the policy representation sequence list, specifically including: Use a recurrent neural network to perform sequence modeling on tax preferential policies and interpretation documents, and capture the temporal dependence relationship between each piece of data in the tax preferential policies and interpretation documents; Combine a convolutional neural network to extract paragraph features of the text, and capture the keywords and key phrases corresponding to each paragraph; Based on the order between paragraphs, store the temporal dependence relationship, keywords and key phrases in a paragraph as a policy representation sequence into the policy representation sequence list.

4. The method for automatically matching tax preference policies based on a training model according to claim 3, characterized in that Based on whether a preset high-weight keyword exists in the current policy representation sequence, and the weight interval to which the policy keyword belongs in the current policy representation sequence, determine the window assignment size corresponding to each policy representation sequence, specifically including: When a preset high-weight keyword exists in the current policy representation sequence, determine that the policy representation sequence belongs to a preset high-weight interval; Then, based on the number of preset high-weight keywords that exist, determine the sorting situation of each policy representation sequence in the preset high-weight interval, and then perform window assignment in a decreasing manner based on the first assignment range of the preset window assignment corresponding to the preset high-weight interval and the number of policy representation sequences in the preset high-weight interval; wherein, the window assignments of the policy representation sequences in the preset high-weight interval are all within the first assignment range; In the current policy representation sequence, there is no preset high-weight keyword, and it is determined that the policy representation sequence belongs to the preset low-weight interval; Based on the total number of temporal dependencies, keywords, and key phrases included in the policy representation sequences within the preset low-weight interval, and in the order of the total number, determine the sorting of each policy representation sequence in the preset low-weight interval. Then, based on the second assignment range of the preset window assignment corresponding to the preset low-weight interval and the number of policy representation sequences in the preset low-weight interval, perform window assignment in a decreasing manner; among them, the window assignments of the policy representation sequences in the preset low-weight interval are all within the second assignment range.

5. The method for automatically matching tax incentive policies based on a training model according to claim 1, characterized in that, Obtain the newly added tax preference policies and interpretation documents, and then extract the corresponding policy representation sequence list, specifically including: Obtain the newly added tax preference policies and interpretation documents, and extract the policy representation sequence list; Based on whether the preset high-weight keywords exist in the policy representation sequence and the weight interval to which the policy keywords in the current policy representation sequence belong, determine the window assignment size corresponding to each policy representation sequence; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy representation sequence, and then update the policy representation sequence list through the attention mechanism.

6. The automatic matching method for tax preference policies based on a training model according to claim 1, characterized in that After obtaining the policy interaction data corresponding to the current policy matching organization, the method further includes: Obtain the tax preference policies queried by the business personnel through the preset business operation interface, and then feedback the corresponding policy interaction data.

7. An automatic matching system for tax incentive policies based on a training model, characterized in that, The system includes: An extraction module, configured to obtain tax preference policies and interpretation documents, and extract a policy representation sequence list; among them, the policy representation sequence list is composed of several policy representation sequences; An update module, configured to determine the window assignment size corresponding to each policy representation sequence based on whether the preset high-weight keywords exist in the current policy representation sequence and the weight interval to which the policy keywords in the current policy representation sequence belong; according to the window assignment size, determine the window size of the attention mechanism corresponding to the policy representation sequence, and then update the policy representation sequence list through the attention mechanism; An acquisition module, configured to obtain the policy interaction data in the office software involved in the current policy matching organization; among them, the policy interaction data at least includes: operation inquiry data related to policy rules, policy rule notification data; An obtaining module, configured to input a number of unlabeled data into the self-supervised learning ability model to pre-train the self-supervised learning ability model to obtain a pre-trained model; then input the updated policy representation sequence list and policy interaction data into the pre-trained model to complete the fine-tuning of the pre-trained model to obtain a trained training model; obtain the newly added tax preference policies and interpretation documents, and then extract the corresponding policy representation sequence list; then input the policy representation sequence list into the trained training model to obtain the policy interaction data corresponding to the current policy matching organization.

8. The tax preference policy automatic matching system based on a training model according to claim 7, characterized in that, The extraction module includes a first extraction unit, configured to obtain tax preference policies and interpretation documents by means of web crawlers; Read whether there is a preset file interpretation program at the preset address. When there is a preset interpretation program, interpret the tax preference policies and interpretation documents, and use the interpreted documents as the tax preference policies and interpretation documents; When there is no preset interpretation program, obtain a preset file interpretation program added externally through the input interface corresponding to the preset address.

9. The tax preference policy automatic matching system based on a training model according to claim 7, characterized in that, The extraction module includes a second extraction unit. It is used to perform sequence modeling on tax preference policies and interpretation documents using a recurrent neural network to capture the temporal dependence relationships between each piece of data in the tax preference policies and interpretation documents. Combine a convolutional neural network to extract paragraph features from the text and capture the keywords and key phrases corresponding to each paragraph. Based on the sequence order between paragraphs, store the temporal dependence relationships, keywords, and key phrases in a paragraph as a policy representation sequence in the policy representation sequence table.

10. A non-volatile computer storage medium, characterized in that, It stores computer instructions, and when the computer instructions are executed, they implement a method for automatically matching tax preference policies based on a training model as described in any one of claims 1-6.

Images