Personnel information correction method and system based on large model and storage medium

Abstract

The application discloses a personnel information error correction method and system based on a large model and a storage medium, and the method comprises the following processes: S1, data cleaning is performed on personnel information to be corrected, and the post of the personnel information to be corrected is disassembled; S2, the personnel information to be corrected is matched with a human resource library, if the matching is successful, it is determined that the information is correct, and the process is ended; otherwise, S3 is entered; S3, relevant personnel information data in the human resource library is searched to construct an initial error correction data set; S4, the comprehensive similarity of each information in the initial error correction data set is calculated, and N personnel information with the highest similarity is selected to form a final error correction data set; S5, the personnel information to be corrected and the final error correction data set are input into a large language model to obtain correct name and post information; and S6, the correct information is updated to the human resource library and a character confusion matrix, and the application can efficiently and accurately correct personnel information and improve the error correction accuracy according to historical error correction behaviors.

Description

Technical Field

[0001] This invention relates to personnel information management, and in particular to a method, system, and storage medium for personnel information error correction based on a large model. Background Technology

[0002] In important organizations or locations, such as hospitals, traffic control centers, fire stations, power stations, and logistics centers, 24 / 7 on-call personnel are required to ensure the continuous, safe, and efficient operation of critical services. Errors in personnel names and job titles, or discrepancies between names and titles, can lead to various problems: incorrect job titles can confuse responsibilities and cause blame-shifting; incorrect names can reduce communication efficiency, delay emergency response, and negatively impact the organization's image. The accuracy of names and job titles is fundamental to the efficient operation of an organization. Errors in names and titles not only affect internal management but can also result in multiple losses, including legal, economic, and social reputational damage. Therefore, establishing an effective error correction mechanism for personnel names and titles is crucial.

[0003] Chinese invention patent application CN119003752A discloses a method and system for extracting and correcting names and titles in text based on a large model, belonging to the field of artificial intelligence technology. The technical problem it addresses is: how to accurately extract and correct names and titles from user-input text based on a large model to improve text accuracy. The method includes the following steps: writing prompt words for the large model; inputting the original text containing names and titles to be corrected into the large model using these prompt words; extracting names and titles from the original text using the large model and returning the original sentences containing names, titles, and related information as search results; comparing and correcting the names and titles in the search results with those recorded in the database to obtain the correct names and titles; and replacing the original sentences in the search results based on the correct names and titles.

[0004] While the aforementioned methods attempt to use large language models for name and job title error correction, they suffer from several drawbacks: they directly feed all raw text, including a significant amount of correct information, into a large model, resulting in high computational costs and low response efficiency; furthermore, they lack refined preprocessing such as cleaning and decomposing non-standard job title expressions, allowing noise and ambiguity in the raw data to directly interfere with model judgment; they fail to quantify and integrate common character-level error patterns in Chinese names and job titles (such as confusion between similar-looking and similar-sounding characters), making it difficult to accurately match candidate records when character-level errors exist; additionally, this solution is a static system and cannot learn error patterns from historical error correction to achieve self-optimization. These shortcomings collectively limit its application in real-world business scenarios requiring high efficiency, high accuracy, and continuous evolution capabilities. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method, system, and storage medium for personnel information error correction based on a large model that can perform error correction efficiently and accurately and improve the accuracy of error correction based on historical error correction behavior.

[0006] Technical solution: The personnel information error correction method based on a large model described in this invention includes the following processes:

[0007] S1. Clean the data of the names and positions of the personnel to be corrected, and break down the positions of the personnel to be corrected into their departments and standard positions according to the standardized position database.

[0008] S2. Perform an exact match between the name and job title of the person to be corrected and the human resources database. If the match is successful, the information is considered correct and the process ends; otherwise, proceed to step S3.

[0009] S3. Search for personnel information data containing any character from the name and position of the person to be corrected in the name and position data of the human resources database to construct the initial error correction dataset. When the personnel information to be corrected has a job title that is split, only search for personnel information data containing any character from the job title that is split in the job title data of the human resources database to construct the initial error correction dataset.

[0010] S4. Construct a personnel information character confusion matrix based on all characters in the human resources database to reflect the similarity between the character shape and the pinyin; based on the personnel information character confusion matrix, use an improved edit distance algorithm to calculate the comprehensive similarity between the personnel information to be corrected and each personnel information in the initial correction dataset, and select the N personnel information with the highest similarity to form the final correction dataset;

[0011] S5. Input the information of the personnel to be corrected and the final error correction dataset into the big language model. The big language model identifies and outputs the correct name and job information from the final error correction dataset based on the preset error correction task prompt words.

[0012] S6. Update the correct information output by S5 to the human resources database, and increase the relevant weights in the personnel information character confusion matrix according to the character replacement situation that occurs during the error correction process.

[0013] By constructing a character confusion matrix for personnel information, we not only provide a basis for calculating the edit distance of character replacements for subsequent improvements to the edit distance algorithm, but also, combined with the post-correction update mechanism, increase the weight of characters that frequently replace each other, thereby reducing the edit distance between two character replacements. The smaller the edit distance, the higher the overall similarity between the two, and the easier it is to filter this information from the initial error correction dataset for large models to correct errors. In other words, the more frequently replaced characters are, the easier they are to be filtered out. This achieves the goal of improving the accuracy of subsequent error correction based on historical error correction behavior, making this method a dynamic system that can learn from historical behavior to improve its own accuracy.

[0014] In addition, data cleaning and job decomposition can filter out a large amount of invalid noise information in the original personnel information to be corrected, greatly reducing the subsequent computation and model call overhead. Performing equality matching on the cleaned and decomposed information further filters out personnel information that can be directly identified and does not require correction, further reducing the subsequent model call overhead. An initial correction dataset is constructed based on the remaining personnel information to be corrected. Then, an improved edit distance algorithm is used to calculate the comprehensive similarity. The N personnel information with the highest similarity are selected to form the final correction dataset for the large model to identify. This reduces the large amount of information that the large model originally had to process to a very few highly relevant personnel information, greatly improving the processing efficiency of the large model. In summary, this method, based on a large model, can achieve efficient and accurate correction of personnel information, and can further improve the accuracy of correction based on historical correction behavior.

[0015] Preferably, the elements in the personnel information character confusion matrix are confusion weights for a pair of characters, and the formula for calculating the confusion weights is as follows:

[0016]

[0017] in, For the confusion weights of characters A and B, and These are the glyph similarity and pinyin similarity between characters A and B, respectively. and Let be the weighting coefficient, satisfying .

[0018] The confusion weight introduces a character-level similarity quantification model that integrates glyphs and pinyin. This quantification model is pre-calculated as a confusion matrix, which enables subsequent comprehensive similarity calculations to obtain accurate data support at the character level, significantly improving the algorithm's sensitivity to and recognition accuracy of Chinese misspellings.

[0019] Preferably, the job decomposition process for the personnel to be corrected is as follows: standard job titles in the standardized job title library are sorted from most to least number of characters, and substring matching is performed with the job title to be corrected in turn. When the first matching is successful, the substring before the matched standard job title is taken as the department to which the personnel belong, and the subsequent part is discarded.

[0020] Sort by character count from most to least and match first, solving the structured parsing challenge of non-standard job titles. For example, it ensures that "Deputy Director of the Finance Bureau" is correctly broken down into "Finance Bureau" and "Deputy Director," rather than incorrectly into "Deputy Director of the Finance Bureau" and "Director." This job title breakdown helps identify correct job titles and departmental information while filtering out invalid noise, reducing subsequent computational load and model call overhead.

[0021] Preferably, the improved edit distance algorithm is as follows: when calculating the edit distance of converting string S1 to string S2, the edit distance of inserting or deleting a character is 1, and the edit distance of replacing the i-th character S1[i] in string S1 with the j-th character S2[j] in string S2 is 1 - mixWeight(S1[i], S2[j]), where mixWeight(S1[i], S2[j]) is the confusion weight of the two characters S1[i] and S2[j].

[0022] The cost of the "replacement" operation in the classic edit distance algorithm has been improved by replacing the fixed cost 1 of the "replacement" operation with a dynamic cost 1 - mixWeight(S1[i], S2[j]) related to the character confusion weight. When two characters are highly similar (the confusion weight is close to 1), the replacement cost approaches 0, and the algorithm considers the replacement to be "very costly"; otherwise, the cost is high. The confusion weight quantifies the similarity between the character shape and the pinyin. This method allows the edit distance to reflect not only the number of editing steps of the string, but also the shape and sound characteristics of Chinese characters. Moreover, combined with the confusion weight update strategy after error correction, the edit distance can reflect not only the shape and sound characteristics of Chinese characters, but also the frequency of confusion between two characters in the past. This makes the edit distance calculation more comprehensive and accurate, and also improves the accuracy of subsequent comprehensive similarity calculation.

[0023] Preferably, when a character is replaced in the personnel information to be corrected, the confusion weight of the corresponding character in the personnel information character confusion matrix is ​​updated according to the following formula.

[0024]

[0025] in, The obfuscation weights of characters A and B that were replaced before this update. The number of times characters A and B are replaced.

[0026] By using the above formula for updating the obfuscation weights, the obfuscation weights of frequently replaced character pairs can be gradually increased. The design allows the update magnitude to decay with increasing frequency, eventually converging to a stable value. This gradually increases the obfuscation weight for frequently replaced character pairs without saturating the obfuscation weight filtering, while ensuring that the updated obfuscation weight never exceeds 1. This enables the method to continuously adapt to the specific writing habits and common errors of the user group, achieving a personalized adaptive effect that becomes more accurate with use—an advantage that static rule systems cannot match.

[0027] The personnel information error correction system based on a large model described in this invention includes:

[0028] Information filtering module: used to clean the data of the names and positions of the personnel to be corrected, and to break down the positions of the personnel to be corrected into their departments and standard positions according to the standardized position library;

[0029] The equal matching module is used to perform an equal match between the name and job title of the person to be corrected and the human resources database. If the match is successful, the information is determined to be correct and the process ends; otherwise, the process proceeds to the initial error correction dataset construction module.

[0030] Initial Error Correction Dataset Construction Module: This module is used to search for personnel information data containing any character from the name and position data of the personnel to be corrected in the human resources database to construct the initial error correction dataset. When the personnel information to be corrected has a job title that is split, the human resources database will only search for personnel information data containing any character from the job title that is split to construct the initial error correction dataset.

[0031] The final error correction dataset construction module is used to construct a personnel information character confusion matrix that reflects the similarity between the character shape and the pinyin based on all characters in the human resources database. Based on the personnel information character confusion matrix, an improved edit distance algorithm is used to calculate the comprehensive similarity between the personnel information to be corrected and each personnel information in the initial error correction dataset. The N personnel information with the highest similarity are selected to form the final error correction dataset.

[0032] Error correction module: Used to input the information of the personnel to be corrected and the final error correction dataset into the big language model. The big language model identifies and outputs the correct name and job information from the final error correction dataset based on the preset error correction task prompt words.

[0033] Update module: Used to update the correct information output by the error correction module to the human resources database, and to increase the relevant weights in the personnel information character confusion matrix based on the character replacements that occur during the error correction process. The computer-readable storage medium storing one or more programs according to this invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0034] Beneficial effects: Through pre-processing data cleaning, job title structure decomposition, and identical matching, a large amount of correct information can be efficiently intercepted and verified, significantly reducing invalid calls to large models and improving overall processing efficiency. The improved similarity algorithm based on character confusion weights that integrate glyph and pinyin features quantifies common Chinese errors. The comprehensive similarity calculated based on this quantifies the initial error correction dataset, achieving progressive filtering from broad initial screening to highly precise filtering, providing high-quality input data for large models. Specialized prompt word engineering constrains large models to make stable and reliable final judgments. Furthermore, the closed-loop mechanism of dynamically updating confusion weights based on error correction results endows the system with the long-term ability to continuously learn and adaptively optimize from actual use. Therefore, this invention can achieve efficient, accurate, and self-learning personnel information error correction. Attached Figure Description

[0035] Figure 1 This is a simplified flowchart of the present invention;

[0036] Figure 2 This is a detailed schematic diagram of the process of the present invention;

[0037] Figure 3 This is a diagram showing the structure and relationship of the standardized job database, personnel to be corrected, error correction dataset, and human resources database in this invention.

[0038] Figure 4 This is the system architecture diagram of the present invention. Detailed Implementation

[0039] As shown in the figure, the personnel information error correction method based on a large model according to the present invention includes the following process:

[0040] S1. Clean the data of the names and positions of the personnel to be corrected, and break down the positions of the personnel to be corrected into their departments and standard positions according to the standardized position database.

[0041] Data cleaning includes removing spaces, non-Chinese characters, and punctuation marks from names and job titles.

[0042] The process of breaking down the job titles of personnel to be corrected is as follows: Standard job titles in the standardized job title library are sorted by the number of characters from longest to shortest. Each standard job title is then matched with the job title to be corrected. Upon the first successful match, the substring preceding the matched standard job title is taken as the department, and the remaining substring is discarded. It's possible that there are no characters preceding the standard job title, resulting in an empty department information, but this is still considered a successful breakdown. Sort the standardized job titles by the number of characters from longest to shortest and match them with the job title to be corrected sequentially to avoid misidentifying a deputy position as a full position. For example, "Deputy Director of the Finance Bureau." If "Director" is matched first, the job title would be broken down into "Deputy Director of the Finance Bureau + Director," which is clearly incorrect. Matching the longer, longer title first, "Deputy Director," results in the correct breakdown of "Finance Bureau + Deputy Director."

[0043] It should be noted that whether the job decomposition is successful or not, step S2 must be performed for full equality matching.

[0044] S2. Perform an exact match between the name and job title of the person to be corrected and the human resources database. If the match is successful, the information is considered correct and the process ends; otherwise, proceed to step S3.

[0045] Specifically, the name of the person to be corrected is used to perform an exact match on the name of the person in the human resources database. At the same time, the job title of the person to be corrected (including standard job title + department information) is used to perform an exact match on the "full department name + standard job title" or "abbreviated department name + standard job title" or standard job title of the person in the human resources database. The human resources database is searched for personnel information that matches the name and job title of the person to be corrected. If unique personnel information is found, the personnel information to be corrected is determined to be correct and the process ends; otherwise, step S3 is entered.

[0046] S3. Search for personnel information data containing any character from the name and position of the person to be corrected in the name and position data of the human resources database to construct the initial error correction dataset. When the personnel information to be corrected has a job title that can be broken down, only search for personnel information data containing any character from the job title that can be broken down in the job title data of the human resources database to construct the initial error correction dataset.

[0047] Specifically, the system searches the human resources database for personnel whose names contain any character from the name of the person whose error is to be corrected. It also searches the database for personnel whose department name, abbreviation, or standard job title contains any character from the job title of the person whose error is to be corrected (if the job title has been successfully broken down, then the job title here refers to the broken job title, i.e., the standard job title + department information). The union of the personnel information found in the above two steps forms the initial error correction dataset. Special handling is applied when job titles are broken down to prevent noise information (characters) contained in the original job title information from being matched to personnel information, thus increasing subsequent unnecessary workload.

[0048] S4. Construct a personnel information character confusion matrix based on all characters in the human resources database to reflect the similarity between the character shape and the pinyin. Based on the personnel information character confusion matrix, use an improved edit distance algorithm to calculate the comprehensive similarity between the personnel information to be corrected and each personnel information in the initial correction dataset. Select the N personnel information with the highest similarity to form the final correction dataset.

[0049] Based on the human resources database, a character confusion matrix for personnel information is constructed. All characters involved in the personnel information in the human resources database are obtained to form a personnel information character set. For any two characters in the character set (such as character A and character B), their comprehensive similarity is calculated as the character confusion weight.

[0050] The formula for calculating obfuscation weights is as follows:

[0051]

[0052] in, This represents the confusion weight between two characters A and B. and Weighting coefficients (e.g.) =0.5, =0.5), The similarity between characters A and B is the similarity of their Chinese character shapes, which represents the degree of similarity between the two Chinese characters in terms of visual features such as structure, strokes and radicals. The value range is [0,1]. This patent uses the nlp-hanzi-similar tool to calculate the similarity between the two Chinese characters. The similarity between the pinyin of characters A and B represents the degree of closeness between the pinyin strings of two Chinese characters. This patent uses the pinyin4j tool to calculate the similarity between the pinyin of two Chinese characters.

[0053] A personnel information confusion matrix is ​​constructed based on character similarity, recording the confusion weights between different characters. An example matrix is ​​shown below.

[0054] Table 1. Schematic diagram of personnel information character obfuscation matrix

[0055] Character / obfuscation weight open chapter Wei Wei plum ... open 1.0 0.9 0.1 0.1 0.1 ... chapter 0.9 1.0 0.1 0.1 0.1 ... Wei 0.1 0.1 1.0 0.8 0.1 ... Wei 0.1 0.1 0.8 1.0 0.1 ... plum 0.1 0.1 0.1 0.1 1.0 ... ... ... ... ... ... ... ...

[0056] Edit distance is used to quantify the difference between two strings, that is, the minimum editing operation cost required to transform one string into another, including the insertion, deletion, and replacement of characters. The improved edit distance algorithm of this invention is as follows: when calculating the edit distance of transforming string S1 into string S2, the edit distance of inserting or deleting a character is 1, and the edit distance of replacing the i-th character S1[i] in string S1 with the j-th character S2[j] in string S2 is 1 - mixWeight(S1[i], S2[j]), where mixWeight(S1[i], S2[j]) is the confusion weight of the two characters S1[i] and S2[j].

[0057] The formula for calculating the overall similarity between the personnel information to be corrected and the information of each personnel in the initial correction dataset is as follows:

[0058]

[0059] in, The combined similarity between string S1 and string S2, where S1 and S2 can represent the string containing the personnel information to be corrected and the personnel information in the initial correction dataset, respectively. Let S1 be the edit distance between string S1 and string S2. and These are the lengths of string S1 and string S2, respectively;

[0060] Calculated based on the string transition state equation.

[0061] The expression for the string transition state equation is:

[0062]

[0063] in, This is the edit distance for converting the first i characters of string S1 to the first j characters of string S2. For example, if S1 and S2 have a total of M and N characters respectively, then... .

[0064] when This means that the i-th character S1[i] of string S1 and the j-th character S2[j] of string S2 are the same. Therefore, finding the edit distance for converting string S1 to string S2 is equivalent to finding the minimum edit operation cost for converting substring S1(i-1) to S2(j-1), i.e. , where S1(i-1) is the substring consisting of the first i-1 characters of string S1, and S2(j-1) is the substring consisting of the first j-1 characters of string S2.

[0065] when At that time, there are three ways to... Convert to Choose the method with the lowest editing cost as the final edit distance:

[0066] Method 1: Delete the i-th character S1[i] in substring S1(i), the operation cost is 1. Then refer to the operation cost of the edit operation to convert S1(i-1) to S2(j). The final editing operation cost is .

[0067] Method 2: Insert the j-th character S2[j] from S2(j) into substring S1(i). The operation cost is 1. At this time, the last character S1[i+1] of the inserted substring S1(i+1) is the same as the last character S2[j] of the substring S2(j). Only the editing operation cost of substring S1(i) and substring S2(j-1) needs to be considered. The final editing operation cost is .

[0068] Method 3: Replace the i-th character S1[i] of substring S1(i) with the j-th character S2[j] of substring S2(j), converting substring S1(i) to S2(j). Considering the confusion weights of characters S1[i] and S2[j] in the personnel confusion moment, the cost of the edit operation is... .

[0069] The N personnel information records with the highest overall similarity are selected as the final error correction dataset. In this patent, N is chosen to be 5.

[0070] S5. Input the personnel information to be corrected and the final correction dataset into the big language model. The big language model identifies and outputs the correct name and job information from the final correction dataset based on the preset correction task prompts.

[0071] Large models applicable to this invention include, but are not limited to: Tongyi Qianwen, DeepSeek, Doubao Large Model, Baichuan Large Model, Wenxin Yiyan, Xunfei Xinghuo, Kimi Intelligent Assistant, Zhipu Qingyan, Hunyuan Large Model, and Jimeng AI.

[0072] This invention provides a set of prompt words related to personnel names and job titles for large-scale models. The main content of the prompt words is as follows:

[0073] "You need to verify and correct the information of the personnel to be corrected based on the input final error correction dataset, and present this information in a table. This table should have three columns: Correct Name, Correct Job Title, and Verification Status, with the following specific requirements:"

[0074] Correct Name: Check for typos, extra words, or omissions in the name entered, and fill in the correct name.

[0075] Correct Job Title: Check for typos, extra words, or omissions in the job title entered, and fill in the correct job title.

[0076] Verification Status: Indicates the accuracy of the information, including values ​​such as "Information Correct", "Information Incorrect", and "No Such Person Found". If the input name and job title information can be found to be completely identical to the correct name and job title, the result is "Information Correct". If no corresponding person can be found, and the name and job title are significantly different, the result is "No Such Person Found", and the information in the output name and job title columns remains unchanged from the input. In the middle of the two states of "Information Correct" and "No Such Person Found", if complete consistency is not possible, such as the name being the same but the job title having a few extra or missing characters, or the job title being the same but the name being different, the result is "Information Incorrect", and the output is the name and job title of the person who was found.

[0077] During the verification process, please refer to the similarity score and prioritize verifying candidates with high scores.

[0078] Note: Verify that the personnel information in the output table matches the personnel information in the final error correction dataset. Do not add or fabricate personnel arbitrarily, and do not output thought chain content containing "think".

[0079] Next, I will provide you with 5 complete examples:

[0080] --------------------------------------------

[0081] Example 1: Able to find completely identical correct names and job titles.

[0082] Final error correction dataset comparison results table:

[0083] {Name|Position|Similarity Score}

[0084] ---|---|---

[0085] Zhang San | Position 1 | 0.91

[0086] Li Si | Position 2 | 0.85

[0087] Wang Wu | Position 3 | 0.77

[0088] Personnel Information Confirmation Form:

[0089] {Name|Position}

[0090] ---|---

[0091] Zhang San | Position 1

[0092] Output the table after verification:

[0093] {Correct Name|Correct Job Title|Verification Status}

[0094] ---|---|---

[0095] Zhang San | Position 1 | Information Correct

[0096] Example 2: The correct name can be found to be exactly the same, but the job titles are different.

[0097] Final error correction dataset comparison results table:

[0098] {Name|Position|Similarity Score}

[0099] ---|---|---

[0100] Zhang San | Position 1 | 0.91

[0101] Li Si | Position 2 | 0.85

[0102] Wang Wu | Position 3 | 0.77

[0103] Personnel Information Confirmation Form:

[0104] {Name|Position}

[0105] ---|---

[0106] Li Si | Position 4

[0107] Output the table after verification:

[0108] {Correct Name|Correct Job Title|Verification Status}

[0109] ---|---|---

[0110] Li Si | Position 2 | Information Error}

[0111] Example 3: No matching person found

[0112] Final error correction dataset comparison results table:

[0113] {Name|Position|Similarity Score}

[0114] ---|---|---

[0115] Zhang San | Position 1 | 0.91

[0116] Li Si | Position 2 | 0.85

[0117] Wang Wu | Position 3 | 0.77

[0118] Personnel Information Confirmation Form:

[0119] {Name|Position}

[0120] ---|---

[0121] Zhao Er | Position 5

[0122] Output the table after verification:

[0123] {Correct Name|Correct Job Title|Verification Status}

[0124] ---|---|---

[0125] Zhao Er | Position 5 | No such person found}

[0126] Example 4: The correct job title can be found, but the name is different.

[0127] Final error correction dataset comparison results table:

[0128] {Name|Position|Similarity Score}

[0129] ---|---|---

[0130] Zhang San | Position 1 | 0.91

[0131] Li Si | Position 2 | 0.85

[0132] Wang Wu | Position 3 | 0.77

[0133] Personnel Information Confirmation Form:

[0134] {Name|Position}

[0135] ---|---

[0136] Wu Liu | Position 1

[0137] Output the table after verification:

[0138] {Correct Name|Correct Job Title|Verification Status}

[0139] ---|---|---

[0140] Zhang San | Position 1 | Information Error}

[0141] Example 5: Able to find similar names and job titles

[0142] Final error correction dataset comparison results table:

[0143] {Name|Position|Similarity Score}

[0144] ---|---|---

[0145] Zhang San | Position 1 | 0.91

[0146] Li Si | Position 2 | 0.85

[0147] Wang Wu | Position 3 | 0.77

[0148] Personnel Information Confirmation Form:

[0149] {Name|Position}

[0150] ---|---

[0151] Li Si | Position 6

[0152] Output the table after verification:

[0153] {Correct Name|Correct Job Title|Verification Status}

[0154] ---|---|---

[0155] Li Si | Position 2 | Information Error}

[0156] --------------------------------------------

[0157] End of example

[0158] Note: Do not add or fabricate personnel arbitrarily. The table headers must be constructed according to the content within the brackets; do not create new formatting yourself.

[0159] (Correct Name | Correct Job Title | Verification Status)

[0160] ---|---|---)

[0161] You need to insert a left curly brace and a newline "{" before starting the task, and a right curly brace "}" after ending the task. Do not fabricate answer formats.

[0162] When entering each line in the table, follow the example and always use a newline character "\r\n" to end the line. Do not use spaces to fill in the blanks, and do not fabricate answer formats.

[0163] If the verification status in the output of the large model service is "information error", it means that there is an error in the name or position of the person to be corrected, and the corresponding correct data has been found in the correction dataset.

[0164] If the large model service outputs a "No such person found" status, it means the service believes the correct data for the person to be corrected cannot be found in the correction dataset. This generally indicates recent personnel changes within the organization, and the human resources database has not been updated in time. Manual verification of the person to be corrected is required. If the person is incorrect, discard the record; if correct, return the record as the correct result and update the human resources database.

[0165] S6. Update the correct information output by S5 to the human resources database, and increase the relevant weights in the personnel information character confusion matrix according to the character replacement situation that occurs during the error correction process.

[0166] When a character is replaced in the personnel information to be corrected, the confusion weight of the corresponding character in the personnel information character confusion matrix is ​​updated according to the following formula.

[0167]

[0168] in, The obfuscation weights of characters A and B that were replaced before this update. The number of times characters A and B are replaced.

[0169] In addition to the formula above, the following formula can also be used.

[0170]

[0171] Will The value is replaced with a fixed value of 2, which ensures that the confusion weight will not exceed 1 no matter how many iterations there are, and there is no need to record the number of times characters A and B are replaced, further reducing the computational load of the system.

[0172] The personnel information error correction system based on a large model described in this invention includes:

[0173] Information filtering module: This module is used to clean the names and positions of the personnel to be corrected, and to break down the positions of the personnel to be corrected into their departments and standard positions based on a standardized position library. If the breakdown is successful, the module will proceed to the equal matching module; otherwise, the module will proceed to the initial error correction dataset construction module.

[0174] The equal matching module is used to perform an equal match between the name and job title of the person to be corrected and the human resources database. If the match is successful, the information is determined to be correct and the process ends; otherwise, the process proceeds to the initial error correction dataset construction module.

[0175] Initial Error Correction Dataset Construction Module: This module is used to search for personnel information data containing any character from the name and position data of the personnel to be corrected in the human resources database to construct the initial error correction dataset. When the personnel information to be corrected has a job title that is split, the human resources database will only search for personnel information data containing any character from the job title that is split to construct the initial error correction dataset.

[0176] The final error correction dataset construction module is used to construct a personnel information character confusion matrix that reflects the similarity between the character shape and the pinyin based on all characters in the human resources database. Based on the personnel information character confusion matrix, an improved edit distance algorithm is used to calculate the comprehensive similarity between the personnel information to be corrected and each personnel information in the initial error correction dataset. The N personnel information with the highest similarity are selected to form the final error correction dataset.

[0177] Error correction module: Used to input the information of the personnel to be corrected and the final error correction dataset into the big language model. The big language model identifies and outputs the correct name and job information from the final error correction dataset based on the preset error correction task prompt words.

[0178] Update module: Used to update the correct information output by the error correction module to the human resources database, and to increase the relevant weights in the personnel information character confusion matrix according to the character replacements that occur during the error correction process.

[0179] The computer-readable storage medium for storing one or more programs according to the present invention includes one or more programs comprising instructions that, when executed by a computing device, cause the computing device to perform the method described above.

Claims

1. A large model-based personnel information correction method, characterized by, The method comprises the following steps: S1, data cleaning is performed on the name and position of the person to be corrected, and the position of the person to be corrected is disassembled into a department and a standard position according to a standardized position library; S2, the name and disassembled position of the person to be corrected are matched with the human resource library, if the matching is successful, it is determined that the information is correct, and the process ends; otherwise, step S3 is entered; S3, in the name and position data in the human resource library, personnel information data containing any character in the name and position of the person to be corrected is found to construct an initial correction data set, when the person to be corrected has a disassembled position, only the personnel information data containing any character in the disassembled position in the position data in the human resource library is found to be used for constructing the initial correction data set; S4, a personnel information character confusion matrix reflecting the similarity of character shape and pinyin is constructed based on all characters in the human resource library; Based on the personnel information character confusion matrix, an improved edit distance algorithm is used to calculate the comprehensive similarity of the person to be corrected and each personnel information in the initial correction data set, and the N personnel information with the highest similarity is selected to form a final correction data set; The element in the personnel information character confusion matrix is a confusion weight of a pair of characters, and the calculation formula of the confusion weight is ， wherein, is the confusion weight of characters A and B, and are the glyph similarity and the phonetic similarity of characters A and B, respectively; and are weight coefficients, satisfying ; The improved edit distance algorithm is as follows: when calculating the edit distance of converting a string S1 into a string S2, the edit distance of inserting or deleting one character is 1, and the edit distance of replacing the i th character S1 [i] in the string S1 with the j th character S2 [j] in the string S2 is 1 - mixWeight (S1 [i], S2 [j]), wherein mixWeight (S1 [i], S2 [j]) is the confusion weight of the two characters S1 [i] and S2 [j]; When the characters in the person to be corrected information are replaced, the confusion weight of the corresponding characters in the personnel information character confusion matrix is updated according to the following formula ， wherein, the obfuscation weight of characters A and B before this update, the number of times of replacement of characters A and B; S5, the person to be corrected information and the final correction data set are input into a large language model, and the large language model identifies and outputs the correct name and position information from the final correction data set according to a preset correction task prompt word; S6, the correct information output by S5 is updated to the human resource library, and related weights in the personnel information character confusion matrix are increased according to the character replacement in the correction process.

2. The large model-based personnel information correction method according to claim 1, characterized in that: The data cleaning includes removing spaces, non-Chinese characters and punctuation marks in the name and position.

3. The large model-based personnel information correction method of claim 1, wherein, The disassembled position of the person to be corrected is as follows: the standard positions in the standardized position library are sorted in descending order of character number, and are matched with the position to be corrected in turn, when the first matching is successful, the substring before the matched standard position is taken as the department, and the part after is discarded.

4. The large model-based personnel information correction method of claim 1, wherein: The calculation formula of the comprehensive similarity is ， wherein, is the combined similarity of string S1 and string S2, is the edit distance of string S1 and string S2, and are the lengths of string S1 and string S2, respectively. According to the string conversion state transition equation calculation, the expression of the string conversion state transition equation is ， wherein, is the edit distance of the first i characters in string S1 to the first j characters in string S2.

5. The large model-based personnel information correction method according to claim 1, characterized in that: The preset correction task prompt word contains task definition, output format specification and multiple examples, which is used to guide the large language model to understand the task of comparing and selecting from the given final correction data set with the person to be corrected information, and to output according to the set format.

6. A large model-based personnel information correction system, characterized by, It comprises: The information filtering module is configured to perform data cleaning on the name and position of the person to be corrected, and to decompose the position of the person to be corrected into a department and a standard position according to a standardized position library. The congruent matching module is configured to perform congruent matching of the name and decomposed position of the person to be corrected with the human resource library, and if the matching is successful, it is determined that the information is correct, and the process ends; otherwise, the initial correction data set construction module is entered. The initial correction data set construction module is configured to search for personnel information data containing any character in the name and position of the person to be corrected in the name and position data in the human resource library to construct an initial correction data set, and when the person to be corrected has a decomposed position, only search for personnel information data containing any character in the decomposed position in the position data in the human resource library to construct an initial correction data set. The final correction data set construction module is configured to construct a personnel information character confusion matrix reflecting the similarity of character shapes and pinyin based on all characters in the human resource library. Based on the personnel information character confusion matrix, an improved edit distance algorithm is used to calculate the comprehensive similarity of the person to be corrected and each personnel information in the initial correction data set, and the N personnel information with the highest similarity is selected to form a final correction data set. The elements in the personnel information character confusion matrix are the confusion weights of a pair of characters, and the calculation formula of the confusion weight is ， wherein, is the confusion weight of characters A and B, and are the glyph similarity and the phonetic similarity of characters A and B, respectively; and are weight coefficients, satisfying ; The improved edit distance algorithm is as follows: when calculating the edit distance of converting a string S1 into a string S2, the edit distance of inserting or deleting a character is 1, and the edit distance of replacing the i-th character S1[i] in the string S1 with the j-th character S2[j] in the string S2 is 1 - mixWeight(S1[i], S2[j]), wherein mixWeight(S1[i], S2[j]) is the confusion weight of the two characters S1[i] and S2[j]; When the characters in the person to be corrected information are replaced, the confusion weight of the corresponding character in the personnel information character confusion matrix is updated according to the following formula ， wherein, the obfuscation weight of characters A and B before this update, the number of times of replacement of characters A and B; The correction module is configured to input the person to be corrected information and the final correction data set into a large language model, and the large language model identifies and outputs the correct name and position information from the final correction data set according to a preset correction task prompt word. The update module is configured to update the correct information output by the correction module to the human resource library, and to increase the related weights in the personnel information character confusion matrix according to the character replacement in the correction process.

7. A computer-readable storage medium storing one or more programs, the one or more programs comprising instructions that when executed by a computer cause the computer to perform a method comprising: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods of claims 1-5.

Images