A method and system for efficient entry of policy information using a large model

Abstract

The application discloses a method and system for efficiently entering policy information using a large model, and aims at the problems that the existing policy information entry relies on manual work, has low efficiency, has high training cost of traditional NLP annotation, and cannot realize classification, summary extraction and single-file multi-policy splitting. The application adopts a large language model technology, configures field extraction rules through cold start, converts policy files into text through OCR, splits single-file multiple subsidy policies through a large language model, extracts structured data and traceable identifiers, and enters the system after manual confirmation. The application does not require data annotation and model retraining, adapts to variable policy templates, supports complex information extraction, greatly reduces implementation cost, and effectively improves the operation and maintenance efficiency of the policy reporting platform and the information entry accuracy.

Description

Technical Field

[0001] This invention relates to the field of computer information processing technology, specifically to a method and system for efficiently inputting policy information using a large model. Background Technology

[0002] To optimize the business environment, local governments have generally established policy application platforms to provide enterprises with online application services for incentive policies. The efficient and accurate entry of incentive policy information is the main foundation for the stable operation and normal service of the platform.

[0003] Currently, policy information entry mainly relies on manual work or traditional NLP technology, both of which have significant drawbacks. In the manual entry mode, policy documents are diverse in format, complex in content, and lack a unified template. Maintenance personnel must meticulously sift through the documents and manually enter fields, which is not only extremely inefficient but also prone to errors due to misunderstandings or operational mistakes, making it difficult to adapt to the massive policy data entry needs. While traditional NLP technology can assist in information extraction, it suffers from poor adaptability and high implementation costs. Because policy document templates are not standardized, each type of template requires separate data annotation and model training, resulting in a large workload, long cycle, and high implementation difficulty. Furthermore, traditional NLP can only extract fragments of the original text such as policy names and cannot complete classification and summary tasks such as policy category division and implicit deadline summarization. In addition, most policy documents contain multiple incentive policies, and traditional NLP struggles to accurately separate multiple independent policies and extract corresponding information, easily leading to information confusion and incomplete extraction.

[0004] In summary, existing technologies suffer from problems such as low data entry efficiency, insufficient accuracy, high implementation costs, and poor functional adaptability, which seriously restrict the operation and maintenance efficiency of policy application platforms. There is an urgent need for a policy information entry solution that is adaptable to various policy templates, low-cost, and capable of extracting complex information. Summary of the Invention

[0005] To address this, the present invention provides a method and system for efficient policy information input using a large model, solving the problems of existing policy information input relying on manual or traditional NLP, which suffers from low efficiency and poor accuracy; and traditional NLP annotation costs, inability to handle classification, summary extraction, and single-file multi-policy splitting, and difficulty in adapting to changing policy templates.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for efficiently inputting policy information using a large model, comprising:

[0007] S1. System maintenance personnel extract the extraction type and corresponding extraction parameters of each field in the system background. The extraction configuration only needs to be executed once during system cold start.

[0008] S2. Receive policy documents uploaded by users, and convert the policy documents into plain text policy content through the text recognition module;

[0009] S3. Input the plain text policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents;

[0010] S4. Input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the large language model, and the large language model extracts the structured field data and traceability identifier corresponding to each incentive policy;

[0011] S5. The front end displays the structured field data and the corresponding traceability identifier. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

[0012] As a preferred method for efficiently inputting policy information using a large model, in S1, the extraction types include original text extraction, classification extraction, and summary extraction.

[0013] The extraction parameters include the text format, the policy unique identifier, and the field extraction criteria;

[0014] When the extraction type is the classification extraction, the large language model outputs the classification result based on the multi-level policy category semantic matching degree calculation model. The expression of the multi-level policy category semantic matching degree calculation model is:

[0015]

[0016] In the formula, Let P be the content of the incentive policy to be processed and the i-th preset policy category. The overall semantic matching degree, with a value range of [0,1]; Preset policy categories semantic embedding vector; Let P be the semantic embedding vector of the incentive policy content to be processed; This is the function for calculating cosine similarity. Preset policy categories A standard set of keywords; The set of keywords for the subsidy policy content P to be processed; is the Jaccard similarity calculation function; α and β are weight coefficients, satisfying α+β=1; the policy category corresponding to the maximum matching degree is taken as the final classification result;

[0017] When the extraction type is summary extraction, the large language model outputs summary content based on the implicit content confidence verification model, and the expression of the implicit content confidence verification model is:

[0018]

[0019] In the formula, The confidence score of the generated content S is taken in the range of [0,1]; N is the total number of independent semantic sentences obtained by splitting the generated content S. This is the k-th semantic clause summarizing the content; For clauses The maximum semantic similarity between the summary and all clauses of the original text of the incentive policy; the summary is considered valid if the confidence score is not lower than the preset threshold.

[0020] As a preferred method for efficiently inputting policy information using large models, in S1, the extraction configuration also supports dynamic addition, modification and deletion of fields. When it is necessary to add fields to be extracted or adjust existing fields, it is only necessary to configure the corresponding extraction type and extraction parameters in the background, without having to retrain the model and label the data.

[0021] As a preferred method for efficiently inputting policy information using a large model, in S2, the policy documents include edited PDF format, scanned PDF format, Word format, and image format;

[0022] When the policy document is in the scanned PDF format or the image format, the text recognition module uses optical character recognition (OCR) technology to convert the text, and at the same time preprocesses the document based on the scanned document tilt angle detection and correction model.

[0023] The expression for the correction model is:

[0024]

[0025] in, The actual tilt angle of the document, the detection range is to ; Rotate document images The sum of the pixel grayscale values ​​in the y-th row after the angle; H is the pixel height of the document image; The first derivative of the sum of the gray levels of the rows in the rotated image; take the derivative and the value corresponding to the maximum value. To correct the tilt angle, the original image is rotated in reverse.

[0026] After text conversion is completed, the conversion quality is verified based on an OCR recognition accuracy verification model. The expression for the OCR recognition accuracy verification model is as follows:

[0027]

[0028] In the formula, The accuracy score for the OCR recognition result O, with a value range of [0,1]; The total number of characters in the OCR recognition result; To identify the number of incorrect characters; The number of characters missed in recognition; the text conversion is considered valid when the accuracy score is not lower than the preset threshold.

[0029] As a preferred method for efficiently inputting policy information using a large model, in S3, the policy unique identifier distinguishes different policy documents, the large language model assigns a sub-unique identifier to each incentive policy after splitting, and the sub-unique identifier establishes an association with the policy unique identifier of the original policy document to achieve multi-policy traceability association.

[0030] As a preferred method for efficiently inputting policy information using a large model, S3 also includes a split prompt word framework, which stores and manages multiple sets of the preset split prompt word templates. The preset split prompt word templates include policy splitting rules, output JSON format requirements, and policy unique identifier association rules, and select the corresponding template according to the characteristics of the policy documents in different regions.

[0031] As a preferred method for efficiently inputting policy information using a large model, S4 also includes a prompt word extraction framework, which stores and manages multiple sets of the preset prompt word extraction templates. The preset prompt word extraction templates include prompt word structures, field output format requirements, and traceability information generation rules corresponding to different extraction types.

[0032] As a preferred method for efficiently inputting policy information using a large model, in S4, the structured field data includes policy name, application conditions, subsidy standards, application deadline, policy category, and applicable subjects;

[0033] The traceability identifier records the page number, paragraph position, and corresponding original text fragment of the structured field data in the original policy document.

[0034] As a preferred method for efficiently inputting policy information using a large model, in S5, the manual confirmation process includes: verifying the accuracy of the structured field data, verifying the consistency between the field content and the original policy document through the traceability identifier, correcting erroneous extraction results; recording the manually corrected content, and iteratively optimizing the corresponding prompt word template based on the corrected content.

[0035] This invention also provides a system for efficient policy information input using a large model, employing the aforementioned method for efficient policy information input using a large model, comprising:

[0036] Configuration Extraction Unit: This unit is used by system maintenance personnel to extract the extraction type and corresponding extraction parameters of each field in the system backend. The configuration extraction only needs to be executed once during a system cold start.

[0037] Document preprocessing unit: used to receive policy documents uploaded by users and convert the policy documents into plain text policy content through the text recognition module;

[0038] Policy splitting unit: used to input the plain text policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents;

[0039] Information extraction unit: used to input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the large language model, and the large language model extracts the structured field data and traceability identifier corresponding to each incentive policy;

[0040] Result Confirmation and Input Unit: This unit is used to display the structured field data and the corresponding traceability identifier on the front end. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

[0041] The present invention has the following advantages:

[0042] First, this invention does not require data annotation and model training for different policy templates. Through the configurable design of backend fields, it can adapt to multiple policy templates with a single configuration during cold start. No additional annotation and training are required when adding or adjusting extracted fields, which significantly reduces the cost of technology implementation and daily operation and maintenance.

[0043] Secondly, this invention breaks through the limitation of traditional NLP, which can only extract original text. Relying on the semantic understanding and logical reasoning capabilities of large models, it supports three modes: original text extraction, classification extraction, and summary extraction. It can complete complex extraction tasks such as policy category division and implicit information summarization, and adapt to diverse policy information extraction needs.

[0044] Third, by using the unique policy identifier association and the semantic segmentation capability of the large model, this invention can accurately split a single document containing multiple incentive policies into independent policy content, enabling the separate extraction and management of each policy information, and avoiding the problems of information confusion and incomplete extraction.

[0045] Fourth, this invention replaces the manual, word-by-word data entry mode. By using text recognition, large-scale model splitting and extraction, it reduces the amount of manual work. Combined with manual verification of traceability identifiers, it reduces the probability of information errors and omissions, effectively improving the information entry efficiency and operation and maintenance quality of the policy application platform. Attached Figure Description

[0046] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0047] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0048] Figure 1 This is a flowchart illustrating a method for efficiently inputting policy information using a large model, as provided in Embodiment 1 of the present invention.

[0049] Figure 2 This is a technical roadmap for a method of efficiently inputting policy information using a large model, as provided in Embodiment 1 of the present invention.

[0050] Figure 3 This is a schematic diagram of a system architecture for efficiently inputting policy information using a large model, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0051] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] Example 1

[0053] See Figure 1 and Figure 2This invention provides a method for efficiently inputting policy information using a large model, including: S1, system maintenance personnel extract and configure the extraction type and corresponding extraction parameters of each field in the system background, and the extraction configuration only needs to be executed once during system cold start;

[0054] Specifically, traditional policy information extraction relies on manual labor or traditional NLP techniques. Traditional NLP requires separate annotation and training for different templates, resulting in poor adaptability and high implementation costs. This step leverages the characteristic of large-scale model semantic understanding that it eliminates the need for repeated annotation. It completes the field extraction rule configuration once during the initial system deployment and cold start, eliminating the need for subsequent retraining or annotation. This addresses the pain point of high template adaptation costs in traditional NLP from the source, adapting to diverse policy document template scenarios.

[0055] S2. Receive policy documents uploaded by users, and convert the policy documents into plain text policy content through the text recognition module;

[0056] Specifically, government-issued incentive policy documents come in various formats, including editable PDFs, scanned copies, Word documents, and images. These non-plain text formats cannot be directly recognized and processed by large language models. The text recognition module uses optical character recognition (OCR) and document preprocessing technology to convert policy documents of various formats into standard plain text, eliminating format barriers and providing recognizable basic data for subsequent semantic segmentation and information extraction by large models.

[0057] S3. Input the plain text policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents;

[0058] Specifically, policy documents often contain multiple incentive policies within a single document. Traditional NLP cannot accurately distinguish between these independent policies, leading to information confusion. This step leverages the deep semantic understanding capabilities of a large language model, combining unique policy identifiers (used for document tracing and differentiation) with pre-defined splitting prompt templates (defining splitting rules), to guide the large model in accurately identifying the semantic boundaries between policies. This process breaks down a single document into multiple independent incentive policy contents, solving the challenge of splitting multiple policies within a single document.

[0059] S4. Input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the large language model, and the large language model extracts the structured field data and traceability identifier corresponding to each incentive policy;

[0060] Specifically, traditional NLP can only extract fragments of the original text, such as policy names, and cannot complete complex extraction tasks such as classification and summarization. This step, based on the semantic parsing and logical reasoning capabilities of a large language model, combined with preset extraction types, parameters, and extraction prompt word templates, accurately extracts the structured data required for policy declaration from the split, independent policy content; at the same time, it generates traceability identifiers to record the location of the data in the original file, ensuring the traceability and authenticity of the extracted data.

[0061] S5. The front end displays the structured field data and the corresponding traceability identifier. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

[0062] Specifically, while large language models possess strong semantic processing capabilities, extraction errors may still occur due to obscure policy expressions and semantic ambiguities. This step uses a front-end visual display of structured data and traceability identifiers to facilitate operations and maintenance personnel in quickly verifying data accuracy and consistency with the original text. After manual confirmation, the data is entered into the system. This approach leverages the large model to improve efficiency while ensuring data quality through manual verification, thus balancing the efficiency of machine computation with data reliability.

[0063] In one possible embodiment, in S1, the extraction type includes original text extraction, classification extraction, and summary extraction;

[0064] Specifically, to cover the full range of policy information extraction needs, three extraction types are defined based on the capabilities of the large model: original text extraction, which adapts to information directly recorded in the original text such as policy name and policy document number; classification extraction, which adapts to information that requires semantic classification such as policy category and applicable industry; and summary extraction, which adapts to information that is not explicitly stated in the original text and requires full-text reasoning, such as application deadline and main reward and subsidy logic. This comprehensively covers complex extraction scenarios that traditional NLP cannot achieve.

[0065] The extraction parameters include the text format, the policy unique identifier, and the field extraction criteria;

[0066] Specifically, the extracted parameters provide clear execution rules and constraints for the large model: the text format is used to standardize the format consistency of the large model's output data and avoid format chaos; the policy unique identifier is used to link the source relationship between the split sub-policies and the original document; the field extraction standard clarifies the definition, scope and judgment rules of each field to be extracted, ensuring that the large model's extraction results comply with the input specifications of the policy application platform and reduce invalid output.

[0067] When the extraction type is the classification extraction, the large language model outputs the classification result based on the multi-level policy category semantic matching degree calculation model. The expression of the multi-level policy category semantic matching degree calculation model is:

[0068]

[0069] In the formula, Let P be the content of the incentive policy to be processed and the i-th preset policy category. The overall semantic matching degree, with a value range of [0,1]; Preset policy categories semantic embedding vector; Let P be the semantic embedding vector of the incentive policy content to be processed; This is the function for calculating cosine similarity. Preset policy categories A standard set of keywords; The set of keywords for the subsidy policy content P to be processed; is the Jaccard similarity calculation function; α and β are weight coefficients, satisfying α+β=1; the policy category corresponding to the maximum matching degree is taken as the final classification result;

[0070] Specifically, a large language model is used to transform policy texts and predefined categories into high-dimensional semantic embedding vectors. The cosine of the angle between these vectors is calculated to capture deep semantic relationships within the text, such as the semantic relationship between "R&D expense deduction" and "technology support," avoiding semantic bias caused by relying solely on keyword matching. The ratio of the intersection to the union of policy text keywords and predefined category standard keyword sets is calculated to measure the overlap of surface keywords, strengthening the contribution of key policy identifiers, such as "micro and small enterprises" and "high-tech enterprises," to the classification results. The model can be dynamically adjusted according to the characteristics of policy expression in different regions. For example, regions with standardized policy expressions and clear semantics can increase the α weight, while regions with obvious keyword features can increase the β weight, improving the model's regional adaptability. All predefined multi-level policy categories are traversed, and the comprehensive matching degree between the policy to be processed and each category is calculated. The category corresponding to the maximum value is taken as the result, ensuring the uniqueness and accuracy of the classification results.

[0071] When the extraction type is summary extraction, the large language model outputs summary content based on the implicit content confidence verification model, and the expression of the implicit content confidence verification model is:

[0072]

[0073] In the formula, The confidence score of the generated content S is taken in the range of [0,1]; N is the total number of independent semantic sentences obtained by splitting the generated content S. This is the k-th semantic clause summarizing the content; For clauses The maximum semantic similarity between the summary and all clauses of the original text of the incentive policy; the summary is considered valid if the confidence score is not lower than the preset threshold.

[0074] Specifically, the summary content generated by the language model is broken down into independent semantic sentences, and each sentence is verified to ensure semantic consistency with the original text, avoiding the overall verification from masking local errors. For each summary sentence, its semantic similarity to all sentences in the original policy text is calculated, and the maximum value is taken as the confidence score of that sentence. As long as the summary sentence can find a semantically corresponding basis in the original text, the sentence is considered valid. The average of the maximum similarity of all sentences is taken as the confidence score of the entire summary content, reflecting the overall degree of fit between the summary content and the original text. A pre-set confidence threshold is set; summaries with scores higher than the threshold are considered valid and can be directly output; those with scores lower than the threshold are marked for manual review to avoid implicit errors, such as incorrect application deadlines being entered into the system.

[0075] In one possible embodiment, in S1, the extraction configuration also supports dynamic addition, modification and deletion of fields. When it is necessary to add a field to be extracted or adjust an existing field, it is only necessary to configure the corresponding extraction type and extraction parameters in the background, without having to retrain the model and label the data.

[0076] Specifically, the fields entered into the policy application platform are dynamically adjusted according to government needs. Traditional NLP requires re-labeling data and retraining the model for newly added / modified fields, which is time-consuming and costly. This solution relies on the zero-shot / few-shot learning capabilities of large models, eliminating the need for model retraining. Fields can be added, modified, or deleted quickly through a visual configuration interface in the backend, adapting to the dynamic iteration needs of policy fields and significantly reducing operation and maintenance adjustment costs.

[0077] In one possible embodiment, in S2, the policy document may be in the format of an edited PDF, a scanned PDF, a Word document, or an image.

[0078] Specifically, the channels and tools used by local governments to release incentive policies are inconsistent, resulting in a wide variety of policy document formats, including editable electronic documents, scanned archived documents, and image-based public notices. This solution comprehensively covers mainstream policy document formats, avoiding document processing limitations caused by format restrictions, adapting to the policy document input needs of different regions and release scenarios, and improving the system's versatility.

[0079] When the policy document is in the scanned PDF format or the image format, the text recognition module uses optical character recognition (OCR) technology to convert the text, and at the same time preprocesses the document based on the scanned document tilt angle detection and correction model.

[0080] The expression for the correction model is:

[0081]

[0082] in, The actual tilt angle of the document, the detection range is to ; Rotate document images The sum of the pixel grayscale values ​​in the y-th row after the angle; H is the pixel height of the document image; The first derivative of the sum of the gray levels of the rows in the rotated image; take the derivative and the value corresponding to the maximum value. To correct the tilt angle, the original image is rotated in reverse.

[0083] Specifically, when the document image is horizontal, the text pixels are concentrated within the horizontal rows, and the difference in the total grayscale value between adjacent rows is significant. The first derivative of the total grayscale value of each row shows a noticeable peak, and the sum of the absolute values ​​of the derivatives of all rows reaches its maximum. When the image is tilted, the text spans multiple rows, the change in the total grayscale value of each row is gradual, and the sum of the absolute values ​​of the derivatives decreases. The tilt angle detection range is limited to -45° to 45°, covering common tilt scenarios in government document scanning, while avoiding excessive computation caused by traversing all angles. By traversing all possible angles from -45° to 45°, rotating the image, and calculating the sum of the absolute values ​​of the first derivative of the total grayscale value of each row, the angle corresponding to the maximum value is the actual tilt angle of the document. The original image is then rotated in the opposite direction to the tilt angle to correct it to a horizontal state, eliminating the interference of tilt on subsequent OCR text recognition and improving the accuracy of text conversion.

[0084] After text conversion is completed, the conversion quality is verified based on an OCR recognition accuracy verification model. The expression for the OCR recognition accuracy verification model is as follows:

[0085]

[0086] In the formula, The accuracy score for the OCR recognition result O, with a value range of [0,1]; The total number of characters in the OCR recognition result; To identify the number of incorrect characters; The number of characters missed in recognition; the text conversion is considered valid when the accuracy score is not lower than the preset threshold.

[0087] Specifically, the accuracy rate is obtained by dividing the number of correctly recognized characters by the total number of characters, where the number of correctly recognized characters = total number of characters - number of misrecognized characters - number of missed recognized characters, comprehensively covering the two main problems of "misrecognition" and "missed recognition" in OCR recognition; misrecognized characters refer to characters where the OCR output does not match the original text, such as recognizing "award subsidy" as "award supplement"; missed recognized characters refer to characters that exist in the original text but are not output by OCR, such as blurred handwriting or text blocked by a seal; a preset accuracy threshold is set, and the text conversion result with a score higher than the threshold is determined to be valid and can enter the subsequent large model processing link; those lower than the threshold will automatically prompt the user to re-upload a clear file manually or correct the text manually to ensure the quality of the basic data for subsequent policy splitting and information extraction.

[0088] In a possible embodiment, in S3, the policy unique identifier distinguishes different policy documents, and the large language model assigns a sub-unique identifier to each split award subsidy policy, and the sub-unique identifier establishes an association relationship with the policy unique identifier of the original policy document to achieve multi-policy traceability association.

[0089] Specifically, after a single file is split into multiple independent award subsidy policies, problems such as confusion in policy attribution and difficulty in traceability are likely to occur. By assigning a unique identifier to the original file, assigning a sub-unique identifier to the sub-policies after splitting and establishing an association, a hierarchical traceability system of "original file, sub-policy" is constructed, which can quickly trace the original file source and context of any award subsidy policy, facilitating data verification, error tracing and subsequent maintenance, and ensuring the orderliness of multi-policy management.

[0090] In a possible embodiment, in S3, it further includes a splitting prompt word framework for storing and managing multiple sets of the preset splitting prompt word templates. The preset splitting prompt word templates include policy splitting rules, output JSON format requirements, and policy unique identifier association rules, and select the corresponding template according to the characteristics of the policy documents in different regions.

[0091] Specifically, there are differences in the expression habits and structural frameworks of policy documents in different regions, and the adaptability of a single splitting template is limited. The splitting prompt word framework contains multiple sets of preset templates adapted to the policy characteristics of different regions, and the rules such as splitting boundaries, data output formats, and identifier association logics are clearly defined in the templates; users can select templates as needed without repeating the design of prompt words, adapting to the splitting requirements of regionalized policy documents, and improving the splitting accuracy and flexibility.

[0092] In a possible embodiment, in S4, it further includes an extraction prompt word framework for storing and managing multiple sets of the preset extraction prompt word templates. The preset extraction prompt word templates include prompt word structures corresponding to different extraction types, field output format requirements, and traceability information generation rules.

[0093] Specifically, the logic of different extraction types differs significantly, requiring targeted prompts to guide the large model in executing tasks. The extraction prompt framework designs standardized prompt structures for the three extraction types, clarifying the field output format and the logic for generating traceability information; it also supports user-defined templates to adapt to the extraction needs of special policy fields, ensuring a high degree of matching between the large model's extraction logic and task objectives, and improving the consistency and accuracy of structured data extraction.

[0094] In one possible embodiment, in S4, the structured field data includes policy name, application conditions, subsidy standards, application deadline, policy category, and applicable subject;

[0095] Specifically, the aforementioned fields constitute the main information displayed to enterprises by the policy application platform for their application of incentive policies. They form the basis for enterprises to quickly select suitable policies and submit application materials. Accurate extraction of these fields ensures the normal operation of the platform's main service functions, meets enterprises' primary needs for querying and applying for incentive policies, and complies with the standardized requirements for government policy information entry.

[0096] The traceability identifier records the page number, paragraph position, and corresponding original text fragment of the structured field data in the original policy document.

[0097] Specifically, the purpose of traceability identifiers is to establish a direct link between extracted data and the original document: page numbers locate the physical location of the document, paragraph positions locate the context of the data, and original text fragments provide direct comparison evidence. Maintenance personnel can use traceability identifiers to quickly verify the consistency between extracted data and the original text, quickly locate and correct erroneous extracted content, and ensure that the policy data entered into the system is authentic, accurate, and traceable.

[0098] In one possible embodiment, in S5, the manual confirmation process includes: verifying the accuracy of the structured field data, verifying the consistency between the field content and the original policy document through the traceability identifier, correcting erroneous extraction results; recording the manually corrected content, and iteratively optimizing the corresponding prompt word template based on the corrected content.

[0099] Specifically, manual verification involves three layers of logic: first, verifying data accuracy and identifying obvious extraction errors; second, cross-validating through source identification to ensure data consistency with the original text; and third, correcting errors to directly optimize the extraction results. Simultaneously, the system records the corrections, forming optimized samples for iterative optimization of the prompt word template. This allows the large model to avoid similar errors during subsequent extractions, forming a "extraction, verification, correction, optimization" logic that continuously improves the accuracy of information extraction from the large model and reduces the workload of subsequent manual verification.

[0100] The application scenarios of this invention are as follows:

[0101] Government Business Environment Incentive Policy Input Scenarios: Applicable to government information technology operation and maintenance processes at all levels, including government service platforms, business environment management systems, and enterprise service centers. It supports multiple formats of incentive policy materials, such as PDF, Word, images, scanned documents, and exported files from public notice webpages. After the operation and maintenance end uploads policy documents, the system extracts text through OCR and parses semantics using a large model to split a single file into multiple policies. It outputs structured data such as policy names, application conditions, and incentive standards according to cold start configuration rules. Relying on a configurable field framework and multiple prompt word templates, it adapts to various templates. Combined with traceability identification verification and manual confirmation, it completes data entry, meeting the needs of government agencies for rapid input of multiple template incentive documents, dynamic adjustment of adaptable fields, reduced annotation and training costs, improved platform operation and maintenance efficiency, and ensuring accurate and compliant policy information input.

[0102] Intelligent data entry scenario for local industry support policies: Applicable to the operation and maintenance of industry policy management platforms of local development and reform commissions, industry and information technology bureaus, science and technology bureaus, and park enterprise service systems, etc. It is compatible with various formats of industry policy materials such as PDFs, scanned copies, official document images, and Word compilation files related to industry support, science and technology subsidies, and tax incentives. After the operation and maintenance end uploads the policy documents, the system uses OCR text conversion and large-scale model semantic segmentation to extract multiple policy items. It extracts data such as application conditions, subsidy standards, and applicable industries according to preset field configurations. Based on multi-level category matching and summary reasoning, it generates implicit application period fields. After tracing and locating the original text fragments, it manually verifies and enters the data. This meets the needs of efficiently collecting industry policies from multiple departments, accurately segmenting the content of multiple items, extracting implicit application rules, reducing manual input errors, and improving the efficiency of industry policy outreach.

[0103] Batch entry scenario for preferential policies for enterprises in industrial parks: Applicable to the operation and maintenance of government services in industrial parks, high-tech zones, and economic development zones, such as enterprise preferential service platforms and enterprise policy matching systems. It is compatible with various formats of preferential materials for enterprises in industrial parks, including PDFs, scanned documents, announcement images, and batch-compiled documents related to special subsidies, rent reductions, and talent awards. After the operation and maintenance side uploads policy documents in batches, the system uses OCR to extract text in batches, splits multiple policies in a single file in parallel using a large model, dynamically configures the extracted fields to output structured data, and adapts the policy-specific expressions of the park based on customized prompt word templates. After traceability verification, batch manual confirmation and entry are performed to meet the needs of batch processing of preferential documents for enterprises in the park, quickly splitting multiple policy items, adapting to park-specific field adjustments, reducing manual annotation costs, and improving the response speed of park policy services.

[0104] Policy information entry scenario for SME service platforms: Applicable to the operation and maintenance of enterprise services such as provincial, municipal and county SME public service platforms, private economy service systems, and one-stop enterprise service platforms. It is compatible with various formats of SME support materials, including PDFs, scanned copies, public notice images, and Word notification documents, related to SME support, start-up subsidies, and financing interest subsidies. After the operation and maintenance end uploads policy documents, the system uses OCR text recognition and large-scale model semantic decomposition to extract independent policies. It configures three extraction rules: original text, classification, and summary to extract key fields, completes intelligent classification based on a multi-level policy category system, and manually confirms the entry after tracing and verifying the consistency of information. This meets the needs of adapting to various support policies for SMEs, accurately decomposing multiple policy contents, completing intelligent classification of policy categories, extracting implicit application deadlines, and improving the accuracy of SME policy queries.

[0105] Intelligent data entry scenario for government policy database construction: Applicable to the data collection process of government big data platforms, policy resource databases, and government information sharing and exchange platforms at all levels. Adaptable to various policy materials from multiple sources, including PDFs, scanned copies, official documents, and image-based public notices related to awards, subsidies, and support, across departments and fields. After the operation and maintenance end collects and uploads policy documents from multiple sources, the system uses OCR for unified text conversion, large-scale model splitting of single files into multiple independent policies, cold start configuration for general extraction fields to output structured data, and multiple sets of regionally adapted prompt word templates to accommodate different expressions. After source verification, the data is entered into the database, meeting the needs of efficiently collecting cross-departmental multi-format policy documents, uniformly splitting multiple policy items, adapting to regional policy template differences, reducing database construction and annotation costs, and ensuring standardized data entry of government policy data.

[0106] It should be noted that the method of this embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this embodiment, and these multiple devices will interact with each other to complete the method for efficiently inputting policy information using a large model.

[0107] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0108] Example 2

[0109] See Figure 3 Embodiment 2 of the present invention also provides a system for efficiently inputting policy information using a large model, employing the above-described method for efficiently inputting policy information using a large model, including:

[0110] Extraction configuration unit 100: used by system operation and maintenance personnel to initially extract the extraction type and corresponding extraction parameters of each field of the configuration in the system background. The extraction configuration only needs to be executed once during system cold start.

[0111] Document preprocessing unit 200: Used to receive policy documents uploaded by system maintenance personnel, and convert the policy documents into plain text policy content through the text recognition module;

[0112] Policy splitting unit 300: Used to input the plain text format policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents;

[0113] Information extraction unit 400: used to input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the big language model, and the big language model extracts the structured field data and traceability identifier corresponding to each incentive policy;

[0114] Result Confirmation and Input Unit 500: Used to display the structured field data and the corresponding traceability identifier on the front end. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

[0115] It should be noted that the information interaction and execution process between the modules of the above-mentioned device are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0116] Example 3

[0117] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a method of efficiently inputting policy information using a large model. The program code includes instructions for executing the method of efficiently inputting policy information using a large model according to Embodiment 1 or any possible implementation thereof.

[0118] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0119] Example 4

[0120] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0121] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can call the program instructions to execute a method for efficient input of policy information using a large model, as described in Embodiment 1 or any possible implementation thereof.

[0122] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0123] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0124] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0125] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A method for efficiently inputting policy information using a large model, characterized in that, include: S1. System maintenance personnel extract the extraction type and corresponding extraction parameters of each field in the system background. The extraction configuration only needs to be executed once during system cold start. S2. Receive policy documents uploaded by users, and convert the policy documents into plain text policy content through the text recognition module; S3. Input the plain text policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents; S4. Input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the large language model, and the large language model extracts the structured field data and traceability identifier corresponding to each incentive policy; S5. The front end displays the structured field data and the corresponding traceability identifier. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

2. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S1, the extraction types include original text extraction, classification extraction, and summary extraction; The extraction parameters include the text format, the policy unique identifier, and the field extraction criteria; When the extraction type is the classification extraction, the large language model outputs the classification result based on the multi-level policy category semantic matching degree calculation model. The expression of the multi-level policy category semantic matching degree calculation model is: In the formula, Let P be the content of the incentive policy to be processed and the i-th preset policy category. The overall semantic matching degree, with a value range of [0,1]; Preset policy categories semantic embedding vector; Let P be the semantic embedding vector of the incentive policy content to be processed; This is the function for calculating cosine similarity. Preset policy categories A standard set of keywords; The set of keywords for the subsidy policy content P to be processed; is the Jaccard similarity calculation function; α and β are weight coefficients, satisfying α+β=1; the policy category corresponding to the maximum matching degree is taken as the final classification result; When the extraction type is summary extraction, the large language model outputs summary content based on the implicit content confidence verification model, and the expression of the implicit content confidence verification model is: In the formula, The confidence score of the generated content S is taken in the range of [0,1]; N is the total number of independent semantic sentences obtained by splitting the generated content S. This is the k-th semantic clause summarizing the content; For clauses The maximum semantic similarity between the summary and all clauses of the original text of the incentive policy; the summary is considered valid if the confidence score is not lower than the preset threshold.

3. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S1, the extraction configuration also supports dynamic addition, modification and deletion of fields. When it is necessary to add a field to be extracted or adjust an existing field, it is only necessary to configure the corresponding extraction type and extraction parameters in the background, without having to retrain the model and label the data.

4. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S2, the policy documents are available in edited PDF format, scanned PDF format, Word format, and image format. When the policy document is in the scanned PDF format or the image format, the text recognition module uses optical character recognition (OCR) technology to convert the text, and at the same time preprocesses the document based on the scanned document tilt angle detection and correction model. The expression for the correction model is: in, The actual tilt angle of the document, the detection range is to ; Rotate document images The sum of the pixel grayscale values ​​in the y-th row after the angle; H is the pixel height of the document image; The first derivative of the sum of the gray levels of the rows in the rotated image; take the derivative and the value corresponding to the maximum value. To correct the tilt angle, the original image is rotated in reverse. After text conversion is completed, the conversion quality is verified based on an OCR recognition accuracy verification model. The expression for the OCR recognition accuracy verification model is as follows: In the formula, The accuracy score for the OCR recognition result O, with a value range of [0,1]; The total number of characters in the OCR recognition result; To identify the number of incorrect characters; The number of characters missed in recognition; the text conversion is considered valid when the accuracy score is not lower than the preset threshold.

5. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S3, the policy unique identifier distinguishes different policy documents, and the large language model assigns a sub-unique identifier to each incentive policy after it is split. The sub-unique identifier is associated with the policy unique identifier of the original policy document to realize the traceability of multiple policies.

6. The method for efficient policy information input using a large model according to claim 1, characterized in that, S3 also includes a split prompt word framework, which stores and manages multiple sets of the preset split prompt word templates. The preset split prompt word templates include policy splitting rules, output JSON format requirements, and policy unique identifier association rules, and select the corresponding template according to the characteristics of the policy documents in different regions.

7. The method for efficient policy information input using a large model according to claim 1, characterized in that, S4 also includes a prompt word extraction framework, which stores and manages multiple sets of the preset prompt word extraction templates. The preset prompt word extraction templates include prompt word structures, field output format requirements, and traceability information generation rules corresponding to different extraction types.

8. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S4, the structured field data includes policy name, application conditions, subsidy standards, application deadline, policy category, and applicable entities; The traceability identifier records the page number, paragraph position, and corresponding original text fragment of the structured field data in the original policy document.

9. The method for efficient policy information input using a large model according to claim 1, characterized in that, In S5, the manual confirmation process includes: verifying the accuracy of the structured field data, verifying the consistency between the field content and the original policy document through the traceability identifier, correcting erroneous extraction results, recording the manually corrected content, and iteratively optimizing the corresponding prompt word template based on the corrected content.

10. A system for efficient policy information input using a large model, comprising the method for efficient policy information input using a large model as described in any one of claims 1-9, characterized in that, include: Configuration Extraction Unit: This unit is used by system maintenance personnel to extract the extraction type and corresponding extraction parameters of each field in the system backend. The configuration extraction only needs to be executed once during a system cold start. Document preprocessing unit: used to receive policy documents uploaded by users and convert the policy documents into plain text policy content through the text recognition module; Policy splitting unit: used to input the plain text policy content, policy unique identifier and preset splitting prompt word template into the big language model, and the big language model splits the policy document into multiple incentive policy contents; Information extraction unit: used to input the content of the multiple incentive policies, the extraction type, the extraction parameters and the preset extraction prompt word template into the large language model, and the large language model extracts the structured field data and traceability identifier corresponding to each incentive policy; Result Confirmation and Input Unit: This unit is used to display the structured field data and the corresponding traceability identifier on the front end. After manual confirmation that the extraction results are correct, the structured field data is entered into the policy application platform system.

Images