KR102511139B1 - Reclassification of Electronic Records Disclosure system and method thereof - Google Patents

Abstract

The present invention relates to a system and method for disclosing and reclassifying electronic records, which enable public institutions to classify, hide, and disclose information subject to non-disclosure, such as personal information, in electronic records produced by public institutions.
The disclosed system for open reclassification of electronic records includes a document database for storing at least one electronic record; a document converting unit for fetching electronic records to be disclosed among the one or more electronic records and converting them into text data; An analysis model unit that loads the converted text data as learning data, performs machine learning through a KOBERT model, and simultaneously performs pattern analysis and address DB to classify and reclassify private parts. ; and a control unit controlling reclassification and disclosure of the one or more electronic records.

Description

Reclassification of Electronic Records Disclosure system and method thereof}

The present invention relates to a method for reclassifying electronic records for disclosure, and more specifically, for electronic records managed by public institutions, to classify non-disclosed information such as personal information to determine whether or not to disclose electronic records. It relates to a system and method for reclassification.

Public institutions produce, receive, and preserve and manage documents related to their work. In addition, it is registered in the electronic records production system, and the registration information determined by the head of the central records management institution is electronically produced and managed. In addition, for the safe and systematic management and utilization of electronic documents, an electronic records management system is established and operated.

The heads of public institutions and records management institutions comply with the 『Public Records Management Act』 (hereinafter referred to as the "Public Records Act") and the 『Information Disclosure Act of Public Institutions』 (hereinafter referred to as the "Information Disclosure Act") so that records can be disclosed and used by the public. ") to manage the disclosure of records.

Disclosure of records refers to allowing public institutions to peruse information in accordance with the Information Disclosure Act, or issuing copies or copies thereof, or information or records through information and communications networks (hereinafter referred to as “information and communications networks”) pursuant to Article 2, subparagraph 10 of the 『Electronic Government Act』 means to provide

Whether or not records are disclosed is registered by classifying them into open, partial, and non-disclosure, and public institutions classify and manage whether or not they are open on a record-by- record basis when producing records. When a public institution transfers records to a competent records management institution, it reclassifies whether or not to disclose and transfers them, and for records reclassified as undisclosed, it must reclassify whether to disclose every five years from the year following the year of reclassification. In principle, all undisclosed records are disclosed 30 years after the end of the production year.

In addition, public institutions periodically reclassify in accordance with Article 35 of the Public Records Act to satisfy the people's right to know and prepare a basis for various records information services, and when the reason for non-disclosure disappears, the relevant records are converted to public. In order to determine whether or not to disclose records, whether or not to disclose them is determined through a comprehensive analysis, including whether or not non-disclosed information corresponding to each subparagraph of Article 9 (1) of the Information Disclosure Act is included. Among them, it can be said that it is very important to detect personal information No. 6, which accounts for the largest part and the importance of protection is increasing.

As the amount of electronic records produced, managed, and transferred by public institutions increases, the subject of periodic disclosure and reclassification is also increasing. To this end, it is necessary to efficiently manage open reclassification of large amounts of electronic records. There is a problem that a large amount of time and budget is required for manual processing to determine whether confidential information such as personal information in electronic records is included. However, commercially available personal information detection products can only detect limited number patterns, and have limitations in detecting various types of personal information including names and addresses included in records.

Republic of Korea Patent Registration No. 10-1504577 (registration date: 2015.03.16) describes a long-term verification method and system for electronic records.

An object of the present invention to solve the above problems is to provide a system and method for analyzing and extracting non-disclosure target information such as personal information with respect to electronic records in public institutions.

In order to achieve the above object, a system for open reclassification of electronic records according to an embodiment of the present invention includes a document database for storing at least one electronic record; a document converting unit for fetching target electronic records from among the one or more electronic records and converting them into text data; For the converted text data, pattern analysis corresponding to each personal information is performed in a hybrid method in which one or more of a machine learning dictionary learning pattern, a regular expression expression pattern, a regular expression expression pattern + DB, a word dictionary + character count pattern, and a word dictionary are combined. an analysis model unit that performs and classifies and reclassifies the non-public part; and a control unit controlling reclassification and disclosure of the one or more electronic records.

In addition, when the reclassified electronic record is disclosed, a mask processing unit may be further included to cover and mask the non-public portion.

The regular expression expression pattern includes an address, and the analysis model unit may process a character string including a word to be disclosed of the database among addresses corresponding to the regular expression expression pattern as an exception.

In addition, the regular expression expression pattern includes e-mail, passport number, resident registration number, alien registration number, business registration number, serial number, driver's license number, card number, address, and lawsuit number, and the machine learning pre-learning pattern is number 6 (Personal information) includes undisclosed words and personal names, and the word dictionary + character count pattern includes No. 6 undisclosed words and No. 2 (National Security/Defense/Unification/Diplomatic Relations) undisclosed words, and the word dictionary The pattern may contain other private words.

The kobert model is a PyTorch model using the Korean Name Entity Recognition (NER) Tagger based on BERT + CRF (Conditional Random Field), and the PER (person) and LOC (place names) of the NER Tagger , ORG (Institution Name), POH (Other), DAT (Date), TIM (Time), DUR (Period), MNY (Currency), PNT (Rate), NOH (Other Quantity Expression) can be used.

In addition, the kobert model processes the converted text data into natural language by applying learning data suitable for the purpose based on a pre-learning model using a sentence-by-sentence embedding technique to which computer-understandable digitization is applied, , Through fine-tuning of the learning data, the name of the person corresponding to the non-public information and, if necessary, the name of the organization, etc. can be extracted and used to determine the non-public information.

The hybrid model of the present invention classifies the private word No. 6 (personal information) in the converted text data into two criteria by personal information type and business classification system, and adds spaces between letters in the basic private word, , it can learn to detect even words with space errors during text conversion.

In addition, the non-disclosed part is information that can be viewed as prescribed by law, information that does not unreasonably infringe on the privacy or freedom of privacy, information that is deemed necessary for the public interest or relief of individual rights, and duties that are not disclosed. Exceptions may be made if they fall under one of the names and positions of the public officials who performed them, or the names and occupations of individuals to whom the state or local governments have entrusted or commissioned part of their duties.

On the other hand, a method for publicly reclassifying electronic records according to an embodiment of the present invention for achieving the above object is a method of storing at least one or more electronic records in a DB and processing target electronic records among the electronic records, ( a) converting a document converting unit into text data by bringing a target electronic record among the one or more electronic records; (b) In addition to performing machine learning by loading the converted text data as learning data, the analysis model unit loads one or more of a regular expression expression pattern, a machine learning dictionary learning pattern, a word dictionary + letter count pattern, and a word dictionary pattern. Performing pattern analysis in a hybrid manner that includes; and (c) a step of classifying and reclassifying, by the control unit, non-public parts from the target electronic records based on the results of the machine learning and pattern analysis.

In addition, in the step (b), the analysis model unit builds a library for undisclosed words to be distinguished from the one or more electronic records in relation to the machine learning; (b-2) With regard to the above-mentioned closed words, analysis frames for each content file including the 6th (personal information) hidden words, 2nd (National Security/Defense/Unification/Diplomatic Relations) hidden words, and other hidden words building a work; (b-3) B (begin) -I (inside) -O (outside) tagging ( tagging); and (b-4) Recognize object names including not only PER (human name) but also LOC (place name) and ORG (organization name) through post-processing (fine-tuning) of the learning data based on the B-I-O tagging and extracting, if necessary, POH (other), DAT (date), TIM (time), DUR (period), MNY (currency), PNT (ratio), NOH (other quantity expression), etc. can be additionally extracted.

In addition, in the step (c), when disclosing the reclassified electronic record, the control unit may cover and mask the non-public portion through the mask processing unit, and then disclose the reclassified electronic record.

In addition, in the step (b), the analysis model unit classifies the private word No. 6 (personal information) in the converted text data into two criteria by personal information type and by business classification system, and the converted text It can learn to classify the second (National Security, Defense, Unification, Diplomatic Relations) closed words in the data, add spaces between letters in the basic closed words, and detect even words with spacing errors when converting text.

According to the present invention, by using a 'hybrid method' that implements an efficient personal information identification method through a "combination of various methods", it is possible to easily identify non-disclosed information for public reclassification of electronic records.

In addition, it is possible to perform reclassification for disclosure of electronic records by using a hybrid method using machine learning + regular expression + DB + word dictionary and letter patterns appropriate for each personal information.

In addition, for electronic records, when reclassifying through machine learning of the Covert model, a higher recognition rate can be obtained in human life.

Further, according to the present invention, when using the hybrid method, a person can perform highly reliable reclassification by using a Covert model, another a regular expression, and another a word dictionary.

In addition, it is possible to derive and pattern the non-disclosed information criteria for open reclassification of electronic records by referring to the National Archives and Records Administration's open reclassification standards.

In addition, according to the present invention, automation and intelligence technology for automatic filtering of undisclosed information can be applied to various types of electronic records such as Hangul, MS-OFFICE, and PDF.

In addition, according to the present invention, it is possible to acquire information for improving the accuracy and suitability of non-disclosed information, and build learning data for automatic filtering with the derived pattern.

In addition, according to the present invention, it is possible to build a prototype system for linking future records management systems (CAMS, RMS).

In addition, according to the present invention, it is possible to build an automatic classification test bed using an artificial intelligence platform and automatically classify the text of electronic records.

In addition, according to the present invention, machine learning can be performed by tagging syllable-based parts of speech based on sequence labeling, and a classification algorithm can be completed using structural SVMs.

In addition, according to the present invention, the highest performance through pre-training and post-processing (fine-tuning) of various natural language processing (text analysis) tasks (entity recognition model, document classification, conversation model, etc.) results can be produced.

In addition, according to the present invention, all natural language processing (text analysis) tasks can be solved with minimal effort through a large-capacity pre-learning model.

In addition, according to the present invention, the type of records file for public reclassification machine learning targets original text files with fewer errors when converted to text for use of learning data, and machine learning and regular expression (pattern) for non-disclosure filtering of records And a hybrid method that uses DB, word dictionary, and letter pattern in combination can be applied. When this is applied, it can show high accuracy in extracting confidential information from records.

In addition, according to the present invention, data obtained by directly converting original records into text data was used as basic input data without going through a process of converting the original records into PDF, and in this case, the process of converting the original records into text was used. Text loss can be minimized.

1 is a configuration diagram schematically showing the overall configuration of a system for open reclassification of electronic records according to an embodiment of the present invention.
2 is a diagram illustrating a process of analyzing non-public information objects of electronic records according to an embodiment of the present invention.
3 is a diagram showing an example of a process in which the system for publicly reclassifying electronic records according to an embodiment of the present invention retrieves records from the National Archives and reclassifies them publicly.
4 is a diagram illustrating a test bed construction and learning process for a system for open reclassification of electronic records according to an embodiment of the present invention.
5 is a flowchart illustrating a process of a masking server when a mask processing unit according to an embodiment of the present invention is implemented as a masking server, which is a separate device.
6 is a diagram showing an example of a masking editing screen of a masking server according to an embodiment of the present invention.
7 is a diagram showing examples of learning target records for public reclassification according to an embodiment of the present invention.
8 is a flowchart illustrating an operation method for reclassifying electronic records according to an embodiment of the present invention.
9 is a flowchart illustrating a process of machine learning and pattern analysis of text data in an analysis model unit according to an embodiment of the present invention.
10 is a diagram showing an example of converting a PDF file written in lower Hangul into a text file.
11 is a diagram showing an example of converting an Excel (MicroSoft Excel) file into a text file.
12 is a diagram illustrating an example of converting a Microsoft Powerpoint file into a text file.
13 is a diagram illustrating a machine learning procedure of an analysis model unit according to an embodiment of the present invention.
14 is a diagram illustrating an analysis model upgrade iterative process according to an embodiment of the present invention.
15 is a diagram schematically showing the configuration of the bert model applied to the Kobert model of the analysis model unit according to an embodiment of the present invention.
16 is a diagram illustrating a hybrid machine learning process of an analysis model unit according to an embodiment of the present invention.
17 is a diagram schematically showing a configuration diagram linking a system for open reclassification of electronic records according to an embodiment of the present invention with a CAMS (RMS) and a masking server.
18 is a diagram showing an exception handling process of a system for open reclassification of electronic records according to an embodiment of the present invention.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily carry out the present invention. The present invention may be embodied in many different forms and is not limited to the embodiments described herein.

In order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals are assigned to the same or similar components throughout the specification.

Throughout the specification, when a part is said to be "connected" to another part, this includes not only the case where it is "directly connected" but also the case where it is "electrically connected" with another element interposed therebetween. . In addition, when a certain component is said to "include", this means that it may further include other components without excluding other components unless otherwise stated.

When a part is referred to as being “on” another part, it may be directly on top of the other part or may have other parts in between. In contrast, when a part is said to be “directly on” another part, there are no other parts in between.

Terms such as first, second and third are used to describe, but are not limited to, various parts, components, regions, layers and/or sections. These terms are only used to distinguish one part, component, region, layer or section from another part, component, region, layer or section. Accordingly, a first part, component, region, layer or section described below may be referred to as a second part, component, region, layer or section without departing from the scope of the present invention.

The terminology used herein is only for referring to specific embodiments and is not intended to limit the present invention. As used herein, the singular forms also include the plural forms unless the phrases clearly indicate the opposite. The meaning of "comprising" as used herein specifies particular characteristics, regions, integers, steps, operations, elements and/or components, and the presence or absence of other characteristics, regions, integers, steps, operations, elements and/or components. Additions are not excluded.

Terms indicating relative space, such as “below” and “above,” may be used to more easily describe the relationship of one part to another shown in the drawings. These terms are intended to include other meanings or operations of the device in use with the meaning intended in the drawings. For example, if the device in the figures is turned over, certain parts described as being “below” other parts will be described as being “above” the other parts. Thus, the exemplary term "below" includes both directions above and below. The device may rotate 90 degrees or other angles, and terms denoting relative space are interpreted accordingly.

Although not defined differently, all terms including technical terms and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present invention belongs. Terms defined in commonly used dictionaries are additionally interpreted as having meanings consistent with related technical literature and currently disclosed content, and are not interpreted in ideal or very formal meanings unless defined.

Hereinafter, with reference to the accompanying drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily carry out the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein.

1 is a configuration diagram schematically showing the overall configuration of a system for open reclassification of electronic records according to an embodiment of the present invention.

Referring to FIG. 1, a system 100 for open reclassification of electronic records according to an embodiment of the present invention includes a communication unit 110, a document conversion unit 120, a control unit 130, an analysis model unit 140, It may include a mask processing unit 150 and a database (DB) 160 .

The communication unit 110 communicates with an external CAMS (Central Archive Management System), an electronic record server (device), or an electronic document system of each level of institution through a communication network, and collects at least one electronic record from them. .

The document conversion unit 120 may bring a target electronic record among one or more electronic records and convert it into text data. Here, the document conversion unit 120 is illustrated as an internal configuration, but is not limited thereto and may be implemented as a separate device. For example, it may be implemented as a document conversion server that is installed outside and communicates remotely through a communication network. .

The control unit 130 may control reclassification and disclosure of one or more electronic records. In addition, the control unit 130 may classify the undisclosed part, inspect the reclassified electrical records, and store and store them in the database 160 .

The analysis model unit 140 loads the converted text data as learning data, performs machine learning through a KOBERT (Korea Bidirectional Encoder Representations from Transformers; KOBERT) model, and simultaneously performs pattern analysis, thereby performing non-public parts. can be classified and reclassified.

The analysis model unit 140 uses a hybrid method including at least one of a machine learning pre-learning pattern, a regular expression expression pattern, a regular expression expression pattern + DB, a word dictionary + character count pattern, and a word dictionary pattern with respect to the converted text data. Pattern analysis corresponding to each personal information can be performed, and non-public parts can be classified and reclassified.

The regular expression expression pattern may include an address, and a process of verifying whether the detected pattern matches the address may be performed using an updated address DB. The analysis model unit 140 may exceptionally process a character string including a word to be disclosed in the database 160 among addresses corresponding to a regular expression expression pattern.

The machine learning pre-learning pattern includes human names, and the analysis model unit 140 may extract desired information such as human names using a KOBERT model.

The regular expression expression pattern may include e-mail, passport number, resident registration number, alien registration number, business registration number, serial number, driver's license number, card number, address, and lawsuit number. The machine learning dictionary learning pattern includes No. 6 (personal information) private words (persons), and the word dictionary + letter pattern includes No. 6 private words (others - gender, certificate, date of birth, etc.) and No. 2 (national security Security·Defense·Unification·Diplomatic Relations), and the word dictionary pattern may include other undisclosed words.

The word dictionary includes private words of No. 6 (personal information) in the converted text data, classifies the private words into two criteria by personal information type and business classification system, and removes spaces between letters in basic private words. May contain additional words. Through this, the word dictionary pattern or the word dictionary + character count pattern can detect even words in which space errors occur during text conversion.

In addition, the word dictionary may include non-public words of No. 2 (national security/defense/unification/diplomatic relations) and may include words in which spaces are added between letters from basic non-public words. Through this, the word dictionary pattern or the word dictionary + character count pattern can detect even words in which space errors occur during text conversion.

The mask processing unit 150 may mask the undisclosed portion of the reclassified electronic record when it is disclosed.

The DB 160 stores at least one electronic record, stores text data converted by the document conversion unit 120, or stores learning data or analysis results processed by the analysis model unit 140.

On the other hand, the kobert model according to the present invention is a PyTorch model using Korean Name Entity Recognition (NER) Tagger based on BERT + CRF (Conditional Random Field), and PER (person) of NER Tagger, 10 including LOC (place name), ORG (organization name), POH (other), DAT (date), TIM (time), DUR (period), MNY (currency), PNT (rate), NOH (other quantity expression) You can use dog basic delimiter tags.

In addition, the Kobert model processes converted text data into natural language through application of training data suitable for the purpose based on a pre-learning model using a sentence-by-sentence embedding technique that applies digitization that a computer can understand, and learns Through data post-processing (fine-tuning), it is possible to recognize and extract the name of an entity including a person's name, place name, organization name, and date.

The present invention relates to information whose non-public parts can be viewed as prescribed by laws and regulations, information that does not unreasonably infringe on the privacy or freedom of privacy, information that is deemed necessary for the public interest or relief of individual rights, and duties to be disclosed. Exceptions may be made in the case of one of the names and positions of public officials who performed the duties, or the names and occupations of individuals to whom the state or local governments entrusted or commissioned part of the work.

Meanwhile, the mask processing unit 150 is illustrated as an internal configuration of the system 100 for open reclassification of electronic records according to the present invention, but is not limited thereto and can be implemented as a separate external device. For example, the mask processing unit 150 may be implemented as a masking server capable of remotely communicating with the system 100 for open reclassification of electronic records according to the present invention through a communication network. At this time, since the masking server performs the same function as the mask processing unit 150, the same reference number as the masking server 150 can be given.

2 is a diagram illustrating a process of analyzing non-public information objects of electronic records according to an embodiment of the present invention.

Referring to FIG. 2 , the system 100 for open reclassification of electronic records according to an embodiment of the present invention, for non-public information in electronic records, refers to the National Archives of Records Disclosure Reclassification Standards, etc. data can be analyzed.

For example, the system 100 for open reclassification of electronic records according to an embodiment of the present invention analyzes data by production organization, record title, record type, target record, non-disclosed information, public value, number, etc. can do.

The non-disclosed subject information filtering and masking system 100 for public reclassification of electronic records according to the present invention may perform public reclassification of electronic records in the following process.

1) As a standard preparation step, the type of information extraction target, related regulations, target information, etc. can be listed.

2) As a basic analysis step, basic analysis such as type, related rule frequency analysis, or morpheme analysis of target information for each related rule can be performed on the extracted information.

3) As an automation/intelligence review step, through basic analysis, it is possible to search for a pattern whether it is possible as a general rule or a case where artificial intelligence application is necessary.

4) As a labeling step, a labeling task for generating a training data set may be performed.

On the other hand, the system 100 for open reclassification of electronic records according to an embodiment of the present invention derives standards for non-public information (centered on Article 9 (1) 6 of the Information Disclosure Act) and patternizes for reclassifying electronic records open. can do.

That is, the system 100 for open reclassification of electronic records according to an embodiment of the present invention presents and patterns applicable standards by analyzing archivist standards, related data, and case studies, and performs standard forms and patterns for non-disclosed information. It is possible to perform non-disclosed information scope, rules, and pattern learning on electronic documents and attached materials.

In addition, the system 100 for open reclassification of electronic records according to an embodiment of the present invention, bi-directional pre-learning through unsupervised learning from a large amount of text, and post-processing (fine-tuning) of analysis of various tasks This can be done through the process. In other words, the highest performance results are produced through pre-training and fine-tuning processes for various natural language processing (text analysis) tasks (entity recognition model, document classification, conversation model, etc.) And, through this large-capacity pre-learning model, all natural language processing (text analysis) tasks can be solved with minimal effort.

In addition, the system 100 for open reclassification of electronic records according to an embodiment of the present invention uses a masked language model (MLM), a language model that masks and predicts 15% of words of characters (the most matching word) predictive model), 80% of the 15% can be masked, 10% can be another random word, and 10% can be left as is.

In addition, the system 100 for open reclassification of electronic records according to an embodiment of the present invention implements a model that predicts the next sentence through NSP (Next Sentence Prediction), so that 50% proceeds to the actual sentence. and the remaining 50% can proceed with non-continuous sentences.

3 is a diagram showing an example of a process in which a system for open reclassification of electronic records according to an embodiment of the present invention brings records from the National Archives of Korea and publicly reclassifies them.

Referring to FIG. 3 , the system 100 for open reclassification of electronic records according to an embodiment of the present invention selects a subject of personal information, designates a processing department, and reflects the examination result when the opinions of the processing department are registered. In the process of doing so, the public reclassification function can be provided.

First, ① After selecting a record, if you press the 'Search' button, a list of records in the corresponding range is provided on the screen, ② A screen is provided to select the range of records to detect personal information, and through this, the detection conditions can be set. If selected, ③ Personal information detection result screen can be provided.

In addition, if you click ④ the record, you can check the record on the screen, and ⑤ if you click the title of the record, the list of attached electronic files is displayed on the screen, and you can also check the number of personal information by type of personal information included. can

In addition, ⑥ By clicking on the electronic file name, you can check the contents of the document through the screen, ⑦ You can select the subject of disclosure and reclassification of the personal information detection record and register the reason, ⑧ Non-disclosure records except for non-disclosure reason No. 6 For detection, in order to detect whether private keywords are included through the multi-private keyword search box provided by the solution, you can organize the private keywords you want to detect based on the 'Public Reclassification Criteria' and enter the private keywords on the keyword management screen. there is.

In addition, you can designate a processing department for ⑨ inspection and opinion registration, and through this, you can proceed with inspection of private keyword detection details, and you can register expert agent opinions in the same way as ⑩ personal information detection records, and finally It can be confirmed after review by the review committee.

In addition, you can select the number of ⑪ reasons for non-disclosure, and when the deliberation is confirmed by the ⑫ Deliberation Committee, you can update the DB by clicking the Reflect Result button.

4 is a diagram illustrating a test bed construction and learning process for a system for open reclassification of electronic records according to an embodiment of the present invention.

Referring to FIG. 4, the system 100 for open reclassification of electronic records according to an embodiment of the present invention first analyzes target records and reference documents (S410), Analyze. At this time, the classification of target records can be defined through the standard document.

Subsequently, as a step of converting target records into text (S420), records subject to disclosure and reclassification are converted into text and then stored. That is, the electronic records delivered from the National Archives are converted into text files to create the type of text data that can be input to the learner/analyzer. At this time, various types of records may be converted into text and stored.

Subsequently, as a step of learning the target record (S430), learning and correction work are performed through the learner.

In addition, as a step of generating an analysis model (pattern) (S440), the analysis model (pattern) is derived through learning. That is, in order to perform machine learning of the converted record, it is loaded into a learner/analyzer to derive an analysis model (pattern) through learning. At this time, derivation and correction of classification criteria through the learner can be repeated, and an analysis model (pattern) can be derived and verified.

Next, as a step of classifying the target records using the analysis model (S450), the target records are classified through the analysis model. At this time, open reclassification of the target records is performed.

Next, as a step of storing the classification result in the DB (S460), the classification result is stored in the DB together with the non-disclosure reason.

That is, a given record is classified using the finally created analysis model, and the result is stored in the DB.

Subsequently, as a step of inspecting the classification result (S470), the classified result of open reclassification is inspected. In other words, as an inspection of reclassified records, it is to review whether the nondisclosure/partial disclosure/disclosure classification and reasons are correct.

Subsequently, when the inspection is completed (S480-Yes), the inspection result is stored in the DB (S482).

However, if the inspection is not completed (S480-No), the result is corrected (S490), and the process of inspecting the classification result (S470) is repeated.

In addition, the system 100 for public reclassification of electronic records according to the present invention may disclose the records by masking the non-public keyword when receiving the DB result and releasing the records.

5 is a diagram showing a processing flow of the masking server when the mask processing unit according to an embodiment of the present invention is implemented as a masking server, which is a separate device, and FIG. 6 is a masking editing screen of the masking server according to an embodiment of the present invention. It is a drawing showing an example.

5 and 6, the masking server 150 according to an embodiment of the present invention, ① registers the classified record names and private keyword results of the learner / analyzer in the DB, ② DB Demon of the PDF conversion server DB The file name and private keywords registered in are read and transmitted to the PDF conversion module. ③ The masking module receives the DB value transmitted to the PDF conversion module, searches for the private keywords in the PDF file, and highlights them.

Subsequently, ④ it is inspected whether the highlighted masking target keyword is accurately detected, and ⑤ the erroneously detected keyword is corrected with PDF Editor as shown in FIG. 6 and transmitted to the masking module.

Then, ⑥ stores the private keyword information (private keyword page and history list) of the final corrected record in the DB, ⑦ the masking module masks the private keywords and stores them as a file in the data storage, Masked records are provided.

7 is a diagram showing examples of learning target records for public reclassification according to an embodiment of the present invention.

Referring to FIG. 7 , the system 100 for open reclassification of electronic records according to an embodiment of the present invention may perform data analysis according to types of undisclosed information by referring to the National Archives and Records Administration open reclassification standards.

In addition, the system 100 for open reclassification of electronic records according to an embodiment of the present invention may perform comparative analysis of data and records according to personal information types. At this time, the types of personal information are general information (name, resident registration number, address, etc.), family information, education and training information, military service information, real estate information, income information, other income information, credit information, employment information, legal information, and medical information. Information, organization information, communication information, location information, body information, habit and hobby information, etc. may be included.

In addition, the system 100 for open reclassification of electronic records according to the present invention utilizes the National Archives of Records Open Reclassification Standards, open reclassification electronic records (No. 6), and open reclassification subject electronic records for machine learning. Machine learning can be performed and a pattern model capable of classifying records can be derived.

Here, the types of non-public information objects include the type of records for non-public information subject No. 6 by business classification system (records management public standard) as shown in Table 1 to Table 4 below. It may be modified, deleted or updated according to the needs of the National Archives of Korea.

In addition, the 6 records subject to non-disclosed information have common personal information types as shown in Tables 5 to 7 below, and can be modified, deleted, or updated according to the national records management standards or the needs of the National Archives of Korea.

In addition, as shown in Table 8 below, non-disclosed information includes Type 2 National Security, National Defense, Unification, and Diplomatic Relations, which can be modified, deleted, or updated according to the national records management standards or the needs of the National Archives of Korea.

8 is a flowchart illustrating an operation method for reclassifying electronic records according to an embodiment of the present invention.

Referring to FIG. 8 , in the system 100 for open reclassification of electronic records according to an embodiment of the present invention, the document converting unit 120 brings a target electronic record among one or more electronic records and shows them in FIGS. 10 to 12 Converted to text data as shown in (S810).

Here, FIG. 10 is a diagram showing an example of converting a PDF file written in Korean below into a text file, FIG. 11 is a diagram showing an example of converting an Excel (MicroSoft Excel) file into a text file, and FIG. 12 is a PowerPoint ( This is a diagram showing an example of converting a Microsoft Powerpoint) file into a text file.

Subsequently, the analysis model unit 140 loads the converted text data as learning data, performs machine learning through a KOBERT model, and simultaneously performs pattern analysis in a hybrid method (S820).

That is, the analysis model unit 140 performs machine learning by loading the converted text data as learning data, and at least one of a regular expression expression pattern, a machine learning dictionary learning pattern, a word dictionary + character count pattern, and a word dictionary pattern. Pattern analysis can be performed simultaneously in a hybrid method involving two or more.

At this time, the Covert model uses two criteria: No. 6 (Personal Information) by type of personal information and No. 2 (National Security/Defense/Unification/Diplomatic Relations) by business classification system. Classify the type, classify the 2nd or 6th hidden words for the converted text data, add spaces between letters in the basic hidden words, and learn to detect even words with spacing errors during text conversion.

As a result of analyzing the 6th and 2nd hidden words according to an embodiment of the present invention, it is preferable to combine pattern analysis and machine learning in consideration of numerous types, characteristics, or patterns of undisclosed information.

Next, the control unit 130 classifies and reclassifies the non-public parts of the target electronic records based on the results of machine learning and pattern analysis (S830).

Then, when disclosing the reclassified electronic records, the control unit 130 masks the undisclosed portion through the mask processing unit 150, and then discloses the reclassified electronic records (S840).

9 is a flowchart illustrating a process of machine learning and pattern analysis of text data in an analysis model unit according to an embodiment of the present invention.

Referring to FIG. 9 , the analysis model unit 140 according to an embodiment of the present invention may build (set) a library for undisclosed words to be distinguished from one or more electronic records (S910).

Next, the analysis model unit 140 is a content file including the 6th (personal information) hidden words, the 2nd (national security/defense/unification/diplomatic relations) hidden words, and other hidden words with respect to the hidden words. A star analysis framework may be constructed (created) (S920).

Subsequently, in the analysis model unit 140, the covert model receives the closed word and utilizes the Korean Name Entity Recognition (NER) tagger based on CRF (Conditional Random Field), B (begin) -I (inside) )-O (outside) tagging is performed (S930).

At this time, the Covert model is PER (person name), LOC (place name), ORG (organization name), POH (other), DAT (date), TIM (time), DUR (period), MNY of the CRF-based Korean NER tagger B-I-O tagging of learning data can be performed using 10 basic classification tags including (currency), PNT (ratio), and NOH (other quantity expressions).

Subsequently, the Covert model modified in the analysis model unit 140 recognizes and extracts the entity name of a person based on the B-I-O tagging (S940).

The analysis model unit 140 according to the present invention has a variety of object names that can be recognized directly from text through machine learning, such as name, gender, place, time, etc., analyzed the characteristics of No. 6 and No. 2 confidential information, and analyzed the original information. Human-centered data was extracted by comprehensively considering sample test results of text files converted from files or PDF files, and if necessary, entity names can be added to determine confidential information.

In addition, the analysis model unit 140 can achieve a relatively high recognition rate compared to machine learning when undisclosed information in the form of a regular expression is extracted or retrieved through regular expression pattern recognition.

Also, the analysis model unit 140 may set a regular expression related pattern. That is, the analysis model unit 140 can set a total of 10 regular expression patterns such as e-mail, passport number, resident registration number, alien registration number, business registration number, serial number, driver's license number, card number, address, and lawsuit number, and different types patterns can be added.

In addition, the analysis model unit 140 can analyze or recognize a relatively detailed address compared to machine learning when using an address regular expression pattern and an address DB in the case of a place where entity name recognition is possible.

In addition, the analysis model unit 140 may separately manage words to be excluded from the address regular expression pattern to exclude patterns that are mistakenly recognized as addresses.

In addition, the analysis model unit 140 may apply a word dictionary and a letter count pattern to the 6th and 2nd confidential information.

That is, the open reclassification system 100 of electronic records according to an embodiment of the present invention may construct an undisclosed word dictionary according to No. 6 (personal information). At this time, No. 6 closed words are personal identification information and personal authentication records, and the types of closed words are classified into two criteria: by personal information type and by business classification system, and after adding spaces between letters in basic non-public words, It is built, and it can be built to detect even words in which space errors occur during text conversion.

In addition, the electronic records open reclassification system 100 according to the present invention can apply the following pattern of the number of characters to the 6th closed word when the letter is an even word and partially blank, and if necessary, the letter pattern can be added

ㅇ 2 letters + space + 2 letters

ㅇ 2 letters + space + 2 letters + space + 2 letters

ㅇ 2 letters + space + 4 letters

ㅇ 3 letters + space + 3 letters

ㅇ 4 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 2 letters + space + 2 letters

ㅇ 3 letters + space + 2 letters + space + 3 letters

ㅇ 4 letters + space + 4 letters

ㅇ 3 characters + space + 5 characters

ㅇ 3 letters + space + 1 letter + space + 4 letters

ㅇ 5 characters + space + 3 characters

ㅇ 5 characters + space + 5 characters

ㅇ 7 letters + space + 3 letters

In addition, the system 100 for open reclassification of electronic records according to the present invention can apply the following pattern of the number of characters to the second hidden word when the character is an odd word and partially blank, and if necessary, the number of characters patterns can be added

ㅇ 1 letter + space + 2 letters

ㅇ 2 letters + space + 3 letters

ㅇ 3 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 3 letters

ㅇ 3 letters + space + 2 letters + space + 2 letters

ㅇ 3 letters + space + 4 letters

ㅇ 4 letters + space + 3 letters

ㅇ 5 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 2 letters + space + 3 letters

ㅇ 4 letters + space + 2 letters + space + 3 letters

ㅇ 5 characters + space + 4 characters

ㅇ 6 characters + space + 3 characters

Meanwhile, the system 100 for open reclassification of electronic records according to an embodiment of the present invention may build an undisclosed word dictionary according to No. 2 (national security, defense, unification, and diplomatic relations). At this time, there are a total of 86 undisclosed words, and it is possible to build a dictionary after adding spaces between letters in basic undisclosed words, and to detect even words with spacing errors during text conversion.

In addition, the system 100 for open reclassification of electronic records according to the present invention may apply the following pattern of the number of characters to the second hidden word when the letters are even words and some are blank.

ㅇ 2 letters + space + 2 letters

ㅇ 2 letters + space + 2 letters + space + 2 letters

ㅇ 2 letters + space + 4 letters

ㅇ 3 letters + space + 3 letters

ㅇ 4 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 2 letters + space + 2 letters

ㅇ 3 letters + space + 2 letters + space + 3 letters

ㅇ 4 letters + space + 4 letters

ㅇ 3 characters + space + 5 characters

ㅇ 3 letters + space + 1 letter + space + 4 letters

ㅇ 5 characters + space + 3 characters

ㅇ 5 characters + space + 5 characters

ㅇ 7 letters + space + 3 letters

In addition, the system 100 for open reclassification of electronic records according to the present invention can apply the following pattern of the number of characters to the second hidden word when the character is an odd word and partially blank, and if necessary, the number of characters patterns can be added

ㅇ 1 letter + space + 2 letters

ㅇ 2 letters + space + 3 letters

ㅇ 3 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 3 letters

ㅇ 3 letters + space + 2 letters + space + 2 letters

ㅇ 3 letters + space + 4 letters

ㅇ 4 letters + space + 3 letters

ㅇ 5 characters + space + 2 characters

ㅇ 2 letters + space + 2 letters + space + 2 letters + space + 3 letters

ㅇ 4 letters + space + 2 letters + space + 3 letters

ㅇ 5 characters + space + 4 characters

ㅇ 6 characters + space + 3 characters

13 is a diagram illustrating a machine learning procedure of an analysis model unit according to an embodiment of the present invention.

Referring to FIG. 13, the analysis model unit 140 according to an embodiment of the present invention brings pre-learning and electronic record learning data from the learning target DB and performs model learning through the PyTorch learner to perform morphological analysis and fine-tuning and entity name recognition.

Next, the analysis model unit 140 stores the final model and documents to be reclassified in the reclassification target DB, imports them to the PyTorch learner, performs reclassification analysis such as data preprocessing, morphological analysis, and model application result output, and extracts necessary information. do.

In addition, the analysis model unit 140 may perform exception processing in the following cases according to Article 9 (1) 6 of the Information Disclosure Act. That is, information that can be viewed as prescribed by law, information that does not unreasonably infringe on privacy or freedom, information that is deemed necessary for the public interest or relief of individual rights to be disclosed, Exception handling is carried out in the case of names, positions, and names and occupations of individuals to whom the state or local government has entrusted or commissioned part of the work.

Meanwhile, the analysis model unit 140 according to the present invention may repeat upgrading of the analysis model as shown in FIG. 14 . 14 is a diagram illustrating an analysis model upgrade iterative process according to an embodiment of the present invention.

The analysis model unit 140 according to the present invention may analyze a text document in a data analysis process and set a non-disclosure criterion.

In addition, the analysis model unit 140 according to the present invention can analyze patterns for each private word type through an analysis model, and upgrade an address DB, word dictionary and analysis model, and pattern analysis.

In addition, the analysis model unit 140 according to the present invention can analyze the machine learning model, determine the performance of the machine learning model, and perform repeated learning.

In addition, the analysis model unit 140 according to the present invention can analyze the needs of the National Archives through a user interface and implement an application service reflecting user requirements.

Meanwhile, the analysis model unit 140 according to the present invention may select a Kobert model by applying the Bert model as shown in FIG. 15 as a machine learning pre-learning model. 15 is a diagram schematically showing the configuration of the bert model applied to the Kobert model of the analysis model unit according to an embodiment of the present invention.

The analysis model unit 140 according to the present invention selects the Kobert model as a machine learning pre-learning model, and can extract human names in national records based on the pre-learned model.

In addition, the analysis model unit 140 may separately store the Kobert pre-learning data and the selected machine learning best model in the pytorch-bert-crf-ner folder.

Kobert is a PyTorch model using Korean NER Tagger based on BERT + CRF. The basic classification tags of NER Tagger are PER (person name), LOC (place name), ORG (organization name), POH (other), DAT (date), TIM (time), DUR (period), MNY (currency), PNT (rate) ) and NOH (other quantity expressions) are used.

The Kobert model uses a sentence-by-sentence embedding technique (applying text to computer-understandable digitization) and can solve natural language processing problems by applying training data suitable for the purpose based on a pre-learning model.

The Kobert model can build a final model (learner) through fine-tuning of training data based on a pre-trained model.

The Kobert model is capable of recognizing entity names such as names of people, places of origin, names of organizations, and dates, but it is possible to extract names of people by implementing the modeling centered on people required for the undisclosed No. 6 word, and, if necessary, placenames and institution names necessary for extracting undisclosed information. It can be used for extracting confidential information by additionally recognizing entity names such as

16 is a diagram illustrating a hybrid machine learning process of an analysis model unit according to an embodiment of the present invention.

Referring to FIG. 16, the analysis model unit 140 according to the present invention, No. 6 (personal information related) hidden word, No. 6 (person name, etc.) hidden word, No. 2 (national defense related) hidden word, etc. For private words (occupation, rank, etc.), you can build a Python library or build an analysis framework (contents by folder/file).

In addition, the analysis model unit 140 may perform pattern recognition and machine learning as follows.

No. 6 Private Word (Regular Expression Pattern) - E-mail address, passport number, social security number, etc.

No. 6 Undisclosed words (others) - Gender, certificate, date of birth, etc.

No. 2/Other Undisclosed Words - North Korea, defense industry, military planning, military personnel, minister, etc.

In addition, the analysis model unit 140 may perform CRF (BIO tagging) - NER (Entity Recognition) - entity name extraction process with respect to the sixth undisclosed word (person's name) through the Covert model.

In addition, the analysis model unit 140 may perform fine-tuning of an object name recognition model for documents and texts of the National Archives of Korea by utilizing the Kobert pre-learning model.

17 is a diagram schematically showing a configuration diagram linking a system for open reclassification of electronic records according to an embodiment of the present invention with a CAMS (RMS) and a masking server.

Referring to FIG. 17 , the system 100 for open reclassification of electronic records according to an embodiment of the present invention can be linked with CAMS/RMS and a masking server through a communication network. Here, the system 100 for open reclassification of electronic records is simply abbreviated as 'open reclassification server 100' and described.

CAMS/RMS includes a storage, a WEB/WAS server, and a DB server, and transmits records subject to public reclassification to the learning/analyzer of the analysis model unit 140 (①).

Subsequently, the learning/analyzer of the analysis model unit 140 transmits the open reclassification result to the CAMS/RMS (②).

Also, when the CAMS/RMS transmits the partially open records to the masking server 150 (②), the masking server 150 transmits the masking records to the CAMS/RMS (③).

In addition, CAMS/RMS provides the records viewing service to the business PC when a request to view records is made on a business PC (①).

That is, the CAMS/RMS and masking server 150 provides the following functions.

1) Through machine learning-based machine learning, records with hidden words are detected and automatically classified.

2) Create a learning model (pattern) through machine learning for non-public classification based on the public reclassification standard.

3) Using the created learning model (pattern), records subject to public reclassification are automatically classified, words subject to non-disclosure are stored in the DB, and these are transmitted to CAMS.

4) When there is a request to read a record subject to partial disclosure, the masking server provides a browsing service by masking the private words of the record.

5) The masking server only accepts PDF files, and masking records can also be provided as PDF files.

On the other hand, the open reclassification server 100-linked agent is linked with the open reclassification function of CAMS and RMS through API, transmits records subject to public reclassification to the public reclassification server 100, and transmits the public reclassification result DB. is transmitted and stored in the open reclassification DB table of CAMS and RMS using API.

The masking server 150-linked agent transmits CAMS and RMS partial disclosure records to the masking server 150 through API, and transmits the open reclassification result DB of the learning/analyzer to the masking server 150 through API, The masked record is received from the masking server 150 using an API, and the record is stored in the DB.

The public reclassification server 100 transfers the records subject to public reclassification to the public reclassification server 100 through the API in connection with the public reclassification function of CAMS and RMS through the CAMS/RMS public reclassification agent. It is received and stored, and the open reclassification server DB is transmitted to the open reclassification DB table of CAMS and RMS through API.

In addition, the open reclassification server 100 converts and stores records transmitted from the CAMS/RMS API into text through a text conversion module, and transmits the text-converted file to the learner/analyzer.

In addition, the open reclassification server 100 generates a learning model through repeated learning of the text records transmitted from the text conversion module through the learning/analyzer, classifies the open reclassification target records through the generated learning model, The classification result and reason for non-disclosure are stored in the learning/analyzer DB, and the open reclassification result is transmitted to the DB through the API linkage module of CAMS and RMS.

The masking server 150 receives a CAMS/RMS partial disclosure request through the CAMS/RMS linkage masking agent, transmits the record to the masking server 150 through the API, and transmits the masked record through the API. It is transmitted to CAM/RMS through and registered in DB.

In addition, the masking server 150 receives the PDF file and the public reclassification DB result value through the masking module, searches for and detects a keyword corresponding to the result value, stores the keyword search detection result value in the DB, 1 The primary search result is inspected, the items to be corrected are edited with a PDF editor, the keywords of the edited PDF file are masked, and the masked file is transmitted to CAMS/RMS through the masking API.

Meanwhile, the system 100 for open reclassification of electronic records according to an embodiment of the present invention can prevent missing letters and files when converting electronic records into text.

That is, the open reclassification server 100 does not convert the PDF conversion file into text and uses it, but converts the original document into text and applies it to the work to minimize missing letters and files. As a result of verifying this, it was confirmed that even special characters in PowerPoint were converted into text close to 100%.

In addition, the system 100 for open reclassification of electronic records according to an embodiment of the present invention may perform exception processing as shown in FIG. 18 for documents that do not need to be disclosed. 18 is a diagram showing an exception handling process of a system for open reclassification of electronic records according to an embodiment of the present invention. In FIG. 18, the public reclassification server 100 retrieves the non-public keyword exception processing and reclassification target documents from the document database 160, determines whether they are private, and executes the non-public classification as described above in case of non-public (Yes). If not (No), exception handling can be performed through the analyzer.

The open reclassification server 100, when a corresponding keyword appears in a record subject to open reclassification, proceeds with post-processing for exception processing of information disclosure after preemptive work of classifying it as undisclosed. Then, exception handling patterning management and corresponding properties are converted from private information to public information.

In addition, in an embodiment of the present invention, the masking server 150 may provide the following functions.

- You can read the results of the learner/analyzer (title of records classified as non-public, list of non-public words), find the non-public words of the PDF file, and process masking.

- To implement this, PDF file Editor and Making function are provided, and DB Demon is provided to receive the result of learner/analyzer.

- It is configured to primarily detect undisclosed words to be masked, correct them and perform final inspection, and more stable operation is possible.

- Connection with CAMS is established by using API (Application Programmable Interface), and the result value of the DB is received from the learner/analyzer and performs masking.

- The masking server runs, but the load is not high, so it can be logically separated rather than physically separated and built.

- Partial disclosure of records that have been requested for disclosure As a method for disclosing more records, it can provide the effect of enhancing transparent administration by increasing the disclosure rate of institutions.

As described above, according to the present invention, it is possible to realize a system and method for disclosing and reclassifying electronic records, which classifies, hides, and discloses non-public parts, such as personal information, of electronic records disclosed by public institutions.

Those skilled in the art to which the present invention pertains should understand that the embodiments described above are illustrative in all respects and not limiting, since the present invention can be embodied in other specific forms without changing the technical spirit or essential characteristics thereof. only do The scope of the present invention is indicated by the following claims rather than the detailed description, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. .

100: system for open reclassification of electronic records
110: Communication Department
120: document conversion unit
130: control unit
140: analysis model unit
150: mask processing unit
160: database (DB)

Claims (15)

a database (DB) for storing at least one or more electronic records;
a document converting unit for fetching target electronic records from among the one or more electronic records and converting them into text data;
For the converted text data, pattern analysis corresponding to each personal information is performed in a hybrid method including at least two or more of a machine learning dictionary learning pattern, a regular expression expression pattern, a word dictionary + letter count pattern, and a word dictionary pattern, An analysis model unit that separates and reclassifies the non-public part; and
a control unit for controlling reclassification and disclosure of the one or more electronic records;
A system for public reclassification of electronic records that includes.
According to claim 1,
a mask processing unit for masking the non-disclosed portion when the reclassified electronic records are disclosed;
System for public reclassification of electronic records further comprising a.
According to claim 1,
The regular expression expression pattern includes an address,
and the analysis model unit exception-processing a string including a word to be disclosed of the database among addresses corresponding to the regular expression expression pattern.
According to claim 1,
The machine learning pre-learning pattern includes human life,
The system for reclassifying electronic records, comprising machine learning and extracting necessary entity names using a Korea Bidirectional Encoder Representations from Transformers (KOBERT) model for the human life.
According to claim 4,
The kobert model is a PyTorch model using the Korean Name Entity Recognition (NER) Tagger based on BERT + CRF (Conditional Random Field), and the PER (person name) and LOC (place name) of the NER tagger , ORG (Institution Name), POH (Other), DAT (Date), TIM (Time), DUR (Period), MNY (Currency), PNT (Rate), NOH (Other Quantity Expression) A system for open reclassification of electronic records, using
According to claim 5,
The Kobert model processes the converted text data into natural language by applying training data suitable for the purpose based on a prior learning model using a sentence-by-sentence embedding technique to which computer-understandable digitization is applied, and the A system for open reclassification of electronic records that recognizes and extracts necessary entity names such as names of people, place names, and institutions through post-processing (fine-tuning) of learning data.
A system for open reclassification of electronic records that recognizes and extracts necessary entity names such as names of people, place names, and institutions.

According to claim 1,
The word dictionary + number of characters pattern or the word dictionary pattern classifies the private word No. 6 (personal information) in the converted text data into two criteria by personal information type and business classification system, and A system for open reclassification of electronic records, characterized by adding spaces between letters in and detecting even words in which spacing errors occur during text conversion.
According to claim 1,
The word dictionary + number of characters pattern or the word dictionary pattern classifies the second (National Security, Defense, Unification, Diplomatic Relations) private word in the converted text data, adds spaces between letters in the basic private word, , A system for open reclassification of electronic records, characterized in that it detects even words with space errors during text conversion
According to claim 5 or 6,
In the kobert model, the non-public part is information that can be viewed as prescribed by law, information that does not unreasonably infringe on the privacy or freedom of privacy, and disclosure is necessary for the public interest or to rescue individual rights. Disclosure of electronic records, which is subject to exception processing, in cases where it falls under one of the following: information deemed to be relevant, the name and position of public officials who performed their duties, and the names and occupations of individuals to whom the state or local governments have entrusted or commissioned part of their duties. A system for reclassification.
According to claim 1,
The regular expression expression pattern includes e-mail, passport number, resident registration number, alien registration number, business registration number, serial number, driver's license number, card number, address, and lawsuit number,
The machine learning pre-learning pattern includes No. 6 (personal information) private words and human names,
The word dictionary + character count pattern includes No. 6 hidden words and No. 2 (National Security/Defense/Unification/Diplomatic Relations) undisclosed words,
wherein the word dictionary pattern includes other non-public words.
A method for reclassifying electronic records in a system for storing at least one electronic record in a DB and reclassifying electronic records to be disclosed among the one or more electronic records,
(a) converting the electronic records to be disclosed among the one or more electronic records into text data by a document converting unit;
(b) The analysis model unit performs machine learning by loading the converted text data as learning data, and at least two of a regular expression expression pattern, a machine learning dictionary learning pattern, a word dictionary + letter count pattern, and a word dictionary pattern. Simultaneously performing pattern analysis in a hybrid manner comprising; and
(c) classifying and reclassifying non-public parts of the electronic records to be disclosed based on the results of the machine learning and pattern analysis by the control unit;
A method for public reclassification of electronic records, including
According to claim 11,
In the step (b), the analysis model unit for the machine learning,
(b-1) setting a library for undisclosed words to be distinguished from the one or more electronic records;
(b-2) With regard to the above-mentioned closed words, analysis frames for each content file including the 6th (personal information) hidden words, 2nd (National Security/Defense/Unification/Diplomatic Relations) hidden words, and other hidden words creating a work;
(b-3) B (begin) -I (inside) -O (outside) tagging ( tagging); and
(b-4) recognizing and extracting entity names including human names, place names, and institution names through post-processing (fine-tuning) of the learning data based on the BIO tagging;
A method for public reclassification of electronic records, including
According to claim 12,
In the step (b-3), the analysis model unit PER (person name), LOC (place name), ORG (organization name), POH (others), DAT (date), TIM (time) of the CRF-based Korean NER tagger , DUR (period), MNY (currency), PNT (ratio), using 10 basic classification tags including NOH (other quantity expressions) to perform the BIO tagging on the learning data, electronic records disclosure material method for classification.
According to claim 11,
In the step (c), the control unit, when disclosing the reclassified electronic records, masks the undisclosed portion through a mask processing unit, and then discloses the reclassified electronic records.
According to claim 11,
In the step (b), the word dictionary + character count pattern or word dictionary pattern classifies the private word No. 6 (personal information) in the converted text data into two types based on personal information type and business classification system. and classifying the second (National Security, Defense, Unification, Diplomatic Relations) private words in the converted text data, adding spaces between letters in the basic private words, and detecting even words with spacing errors during text conversion , Methods for Disclosure Reclassification of Electronic Records.

Images