Personnel file auditing method and system based on large model

Abstract

The invention discloses a personnel archive auditing method and system based on a large model. The method comprises the following steps: performing data identification and error correction on an original archive to obtain preliminary identification data; and auditing the archive data by using the auditing model to obtain an auditing result. According to the method, the auditing model is constructed for the personnel archive data, and specific model optimization is performed on the model, so that the accuracy and practicability of an intelligent auditing function are improved, and manpower is liberated.

Description

Technical Field

[0001] The present application relates to a personnel file review method and system based on a large model. Background Art

[0002] Personnel files refer to learning and work experience, professional qualities, work achievements, and family and social concerns. They are important credentials in recruitment, political review, and professional titles. Since personnel files need to be stable enough, current personnel files are still mainly paper files, and there are a lot of handwritten fonts. Therefore, file review is a "physical job" that requires the participation of a large number of professional staff, a large amount of review work, and a cumbersome review process. Even so, the final quality of file review is difficult to control. Summary of the invention

[0003] In order to solve the above problems, the present application proposes a personnel file review method based on a large model, including the following steps:

[0004] Perform data identification and error correction on the original files to obtain preliminary identification data;

[0005] The audit model is used to audit the file data to obtain the audit results. This application constructs an audit model for personnel file data, and by optimizing the model, the accuracy and practicality of the intelligent audit function are improved, freeing up manpower.

[0006] Preferably, the audit model is obtained as follows:

[0007] Adopt the Transformer model, modify the underlying network structure of Transformer, adopt a multi-level deep feedback neural structure, and improve the ability to extract structured data by increasing the number of network layers and breadth;

[0008] Modify the Transformer attention mechanism and adopt a selective multi-head attention mechanism to ignore irrelevant context and improve the accuracy and speed of the model;

[0009] By adding a generation template and modifying the Transformer generation layer to limit the model generation format, the model can generate accurate JSON format data according to human will, which is convenient for subsequent parsing and use;

[0010] In view of the specific business of smart file review, the word library used by some models is modified, and professional knowledge and experience are input into Transformer. This application establishes a proprietary large model for the special requirements of cadre personnel file review. By modifying the underlying network structure of Transformer, modifying the word library, generating restrictions, etc., a proprietary large model for smart file review is established. The smart file review large model conducts the initial review, and the professional team conducts the detailed review. The review results are continuously fed back to the smart file review large model for algorithm optimization.

[0011] Preferably, an audit rule library is established for the audit method, and Transformer is used to extract key information from a large amount of complex archive audit rule data to accurately identify the elements in the rules;

[0012] Use Transformer to extract and quantify complex case review rules;

[0013] By inputting prior knowledge, the archive audit rule prompt project is automatically generated, and the prior knowledge and archive rule prompt project are input into the archive audit rule extraction model. This application establishes an audit rule library and uses Transformer to extract key information from a large amount of complex archive audit rule data, accurately identify the elements in the rules, and improve the audit accuracy.

[0014] Preferably, the data identification is performed in the following manner:

[0015] Perform OCR recognition on the original files to obtain the original data set;

[0016] The original data set is cleaned, key information is extracted, structured data is extracted, and data is vectorized to obtain preliminary identification data.

[0017] Preferably, the data cleaning is performed in the following manner: cleaning the data, including removing stop words, special characters, missing value processing, and marking named entities in the text such as names of people and places;

[0018] The data vectorization processing is performed as follows: the cleaned data is vectorized using the bce-embedding-base vector model. This application uses OCR technology to identify existing archival data to form an original data set, cleans the data, extracts key information, extracts structured data, and vectorizes the data to form an archival data set, which can quickly process paper archives.

[0019] Preferably, the error correction is performed in the following manner:

[0020] Construct a semantic fluency model, analyze and learn different handwritings, and perform text error correction based on the characteristics of the archives industry.

[0021] Preferably, the error correction is based on the qwen2.5 model:

[0022] The semantic fluency model is constructed by expanding and modifying its network structure and adding a multi-head-multi-level attention mechanism; then the semantic fluency model is trained using prior knowledge of the archives industry and manually cleaned data to obtain an optimized semantic fluency model. This application constructs a semantic fluency model, analyzes and learns different handwritings, introduces a text error correction mechanism based on the characteristics of the archives industry, and improves the OCR recognition rate and accuracy of handwriting.

[0023] Preferably, the audit results are reviewed and the audit model is modified according to the review structure feedback:

[0024] For Transformer extraction errors or indicators that do not meet the requirements, they are marked and backfilled during the manual review process, and Transformer performs automatic learning. For indicators with special requirements, a special indicator library is set up. During the review, the Transformer extraction structure is compared with the data in the special indicator library, and the best result is adopted.

[0025] This application also discloses a personnel file review system based on a large model, including the following modules:

[0026] Data identification module, used to identify and correct the original files to obtain preliminary identification data;

[0027] The audit module is used to use the audit model to audit the archive data to obtain the audit results.

[0028] Preferably, it also includes a review module for reviewing the audit results and modifying the audit model according to the review structure feedback.

[0029] This application can bring the following beneficial effects:

[0030] 1. This application constructs an audit model for personnel file data, and optimizes the model to improve the accuracy and practicality of the intelligent audit function, thereby freeing up manpower.

[0031] 2. This application aims to establish a proprietary big model for the special requirements of cadre personnel file review. By modifying the underlying network structure of Transformer, modifying the vocabulary, generating restrictions, etc., a proprietary big model for intelligent file review is established. The intelligent file review big model will conduct the initial review, and the professional team will conduct a detailed review. The review results will be continuously fed back to the intelligent file review big model for algorithm optimization.

[0032] 3. This application establishes an audit rule library and uses Transformer to extract key information from a large amount of complex archive audit rule data, accurately identify the elements in the rules, and improve the audit accuracy.

[0033] 4. This application uses OCR technology to identify existing archival data to form an original data set, cleans the data, extracts key information, extracts structured data, and vectorizes the data to form an archival data set, which can quickly process paper archives. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 This is a schematic diagram of Example 1.

[0036] Figure 2 This is a schematic diagram of Example 2. DETAILED DESCRIPTION

[0037] In order to clearly illustrate the technical features of the present solution, the present application is described in detail below through specific implementation methods and in conjunction with the accompanying drawings.

[0038] In the first embodiment, if Figure 1 As shown, a personnel file review method based on a large model includes the following steps:

[0039] S101. Perform data identification and error correction on the original files to obtain preliminary identification data;

[0040] The data identification is performed as follows:

[0041] Perform OCR recognition on the original files to obtain the original data set;

[0042] The original data set is cleaned, key information is extracted, structured data is extracted, and data is vectorized to obtain preliminary identification data.

[0043] The data cleaning is performed in the following manner: cleaning the data, including removing stop words, special characters, and missing value processing, and marking named entities in the text, such as names of people and places;

[0044] The data vectorization processing is performed in the following manner: the bce-embedding-base vector model is used to vectorize the cleaned data according to the steps of word segmentation, building a word vector library, feature extraction, and feature vectorization.

[0045] The error correction is carried out in the following manner: based on the qwen2.5 model, a semantic fluency model is constructed by expanding and modifying its network structure and adding a multi-head and multi-level attention mechanism; then the semantic fluency model is trained using prior knowledge of the archives industry and manually cleaned data to obtain an optimized semantic fluency model.

[0046] When constructing the semantic fluency model, different handwritings are analyzed and learned, and text error correction is performed according to the characteristics of the archives industry. Data items extracted incorrectly by the model are corrected by manual verification. For example, if the personal identity extraction result is A, the correct result should be B, and the result is corrected to B.

[0047] S102. Use the audit model to audit the file data to obtain the audit results.

[0048] The audit model is obtained as follows:

[0049] Adopt the Transformer model, modify the underlying network structure of Transformer, adopt a multi-level deep feedback neural structure, and improve the ability to extract structured data by increasing the number of network layers and breadth;

[0050] Modify the Transformer attention mechanism and adopt a selective multi-head attention mechanism to ignore irrelevant context and improve the accuracy and speed of the model;

[0051] By adding a generation template and modifying the Transformer generation layer to limit the model generation format, the model can generate accurate JSON format data according to human will, which is convenient for subsequent parsing and use;

[0052] In view of the specific business of intelligent file review, the word library used by some models is modified to inform the big model of more professional prior knowledge and experience. For example, when extracting the decision time, the resolution time of the general meeting should be used as the decision time;

[0053] For the audit method, an audit rule library was established, and Transformer was used to extract key information from a large amount of complex archive audit rule data to accurately identify the elements in the rules;

[0054] Use Transformer to extract and quantify complex case review rules;

[0055] The archive review rule prompting project is automatically generated by inputting prior knowledge, and the prior knowledge and the archive rule prompting project are input into the archive review rule extraction large model.

[0056] S103. Review the audit results and modify the audit model based on the review structure feedback:

[0057] Input the manual verification and correction results into the big model, and let the big model learn autonomously based on the input results to further improve the auditing ability of the big model;

[0058] For Transformer extraction errors or indicators that do not meet the requirements, they are marked and backfilled during the manual review process, and Transformer performs automatic learning. For indicators with special requirements, a special indicator library is set up. During the review, the Transformer extraction structure is compared with the data in the special indicator library, and the best result is adopted.

[0059] In the second embodiment, if Figure 2 As shown, a personnel file review system based on a large model includes the following modules:

[0060] The data identification module 201 is used to perform data identification and error correction on the original file to obtain preliminary identification data;

[0061] The audit module 202 is used to audit the archive data using the audit model to obtain the audit result.

[0062] The review module 203 is used to review the audit results and modify the audit model according to the review structure feedback.

[0063] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.

Claims

1. A personnel file review method based on a large model, characterized by: The steps include: Perform data identification and error correction on the original files to obtain preliminary identification data; The audit model is used to audit the archival data to obtain the audit results.

2. A personnel file review method based on a large model as claimed in claim 1, characterized in that: The audit model is obtained as follows: Adopt the Transformer model, modify the underlying network structure of Transformer, adopt a multi-level deep feedback neural structure, and improve the ability to extract structured data by increasing the number of network layers and breadth; Modify the Transformer attention mechanism and adopt a selective multi-head attention mechanism to ignore irrelevant context and improve the accuracy and speed of the model; By adding a generation template and modifying the Transformer generation layer to limit the model generation format, the model can generate accurate JSON format data according to human will, which is convenient for subsequent parsing and use; In view of the specific business of intelligent file review, the vocabulary used by some models is modified and professional knowledge and experience are input into Transformer.

3. A personnel file review method based on a large model as claimed in claim 1, characterized in that: For the audit method, an audit rule library was established, and Transformer was used to extract key information from a large amount of complex archive audit rule data to accurately identify the elements in the rules; Use Transformer to extract and quantify complex case review rules; The archive review rule prompting project is automatically generated by inputting prior knowledge, and the prior knowledge and the archive rule prompting project are input into the archive review rule extraction large model.

4. A personnel file review method based on a large model as claimed in claim 1, characterized in that: The data identification is performed as follows: Perform OCR recognition on the original files to obtain the original data set; The original data set is cleaned, key information is extracted, structured data is extracted, and data is vectorized to obtain preliminary identification data.

5. A personnel file review method based on a large model as claimed in claim 4, characterized in that: The data cleaning is performed in the following manner: cleaning the data, including removing stop words, special characters, and missing value processing, and marking named entities in the text, such as names of people and places; The data vectorization processing is performed in the following manner: the cleaned data is vectorized using the bce-embedding-base vector model.

6. A personnel file review method based on a large model as claimed in claim 1, characterized in that: The error correction is performed as follows: Construct a semantic fluency model, analyze and learn different handwritings, and perform text error correction based on the characteristics of the archives industry.

7. A personnel file review method based on a large model as claimed in claim 6, characterized in that: The error correction is based on the qwen2.5 model: A semantic fluency model is constructed by expanding and modifying its network structure and adding a multi-head and multi-level attention mechanism. Then, the semantic fluency model is trained using prior knowledge of the archives industry and manually cleaned data to obtain an optimized semantic fluency model.

8. A personnel file review method based on a large model as claimed in claim 7, characterized in that: It also includes reviewing the audit results and modifying the audit model based on the review structure feedback: For Transformer extraction errors or indicators that do not meet the requirements, they are marked and backfilled during the manual review process, and Transformer performs automatic learning. For indicators with special requirements, a special indicator library is set up. During the review, the Transformer extraction structure is compared with the data in the special indicator library, and the best result is adopted.

9. A system for implementing the large model-based personnel file review method according to any one of claims 1 to 8, characterized in that: Includes the following modules: Data identification module, used to identify and correct the original files to obtain preliminary identification data; The audit module is used to use the audit model to audit the archive data to obtain the audit results.

10. A personnel file review system based on a large model as claimed in claim 9, characterized in that: It also includes a review module for reviewing the audit results and modifying the audit model based on the review structure feedback.

Images