KR20210000916A - Retention period recommendation system and method using RNN of records - Google Patents

Abstract

The present invention relates to a record retention period recommendation system using a recurrent neural network technique and to a recommendation method thereof. More specifically, in generating the retention period for public records of public institutions or records of private companies, deep learning training can be applied to quickly recommend the retention period, and the reliability of the recommended retention period can be improved in the record retention period recommendation system using recurrent neural network technology and the recommendation method thereof. The system includes: a learning data input unit; a data preprocessing unit; a deep learning training unit; a new data input unit; and a retention period management unit.

Description

Retention period recommendation system and method using RNN of records}

The present invention relates to a system for recommending a record retention period using a recurrent artificial neural network technology and a method for recommending the same, and more particularly, in creating a retention period for public records of public institutions or records of private companies, deep learning learning The present invention relates to a system for recommending the retention period of records using a recurrent artificial neural network technology that can quickly recommend the retention period by applying the system and improve the reliability of the recommended retention period, and a method for recommending the same.

In general, the public has the right to request the disclosure of information on the records of public institutions in accordance with the law, and public institutions have the scope, frequency, and fraud of disclosure even if there is no request from the public for policy information that affects public life. It is obligated to pre-determine, publish, and disclose regularly.

According to Article 26 (Retention Period) of the Enforcement Decree of the Record Management Act of Public Institutions, it is classified into 7 types: permanent, semi-permanent, 30 years, 10 years, 5 years, 3 years, and 1 year. Accordingly, there is a difference in the retention period.

It is common for a manager who manages such records to create a list of records, create metadata for each record, and store them on the system. In particular, since the retention period is set at the discretion of the manager in the setting of the retention period among the metadata of records, if the manager processes a large amount of records, there is a very high possibility of human error in the process of setting the retention agency. It becomes high.

Of course, since the National Archives of Records discloses the index information of records, the information related to the retention period among the published index information of records is largely inconsistent as a result of the subjective judgment of the person in charge of the records. Even if the preservation period for new records is estimated through this, there is a problem that the reliability of the result value is low because the pre-processing process for the collected data is not performed.

In this regard, in Korean Patent Registration No. 10-1887629 ("Natural Language-Based Information Disclosure Classification System"), a learning model for information disclosure classification is created, and semantic analysis through natural language processing is performed by applying records subject to disclosure classification, Records subject to public classification are automatically classified into one of public, non-public, and partial disclosure.

Domestic Registration Patent No. 10-1887629 (Registration Date 2018.08.06.)

The present invention was conceived to solve the problems of the prior art as described above, and in creating a retention period for public records of public institutions or records of private companies, prior to applying deep learning learning, learning data It is to provide a system for recommending a record retention period using a recurrent artificial neural network technology that can improve the reliability of a calculation result (recommendation of a retention period) by increasing the significance and accuracy of learning data by performing pre-processing for the training data and a method for recommending the same.

The system for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention includes a learning data input unit 100 receiving training data for learning from previously linked records information related databases, and the text mining. Data for analyzing the frequency of appearance of each morpheme by extracting specific information included in the learning data received from the learning data input unit 100, removing duplicate data, and performing morpheme analysis on the extracted specific information The preprocessing unit 200, a deep learning learning unit 300 that performs learning by receiving one or more specific morphemes based on the frequency of appearance analyzed by the data preprocessing unit 200 using a preset deep learning technique, Using the learning result of the deep learning learning unit 300, a data management unit 400 that generates reference data of the type of retention period for the records, and receives new data related to record information for setting the retention period from the outside. A retention period management unit 600 that maps and compares the reference data of the data input unit 500 and the data management unit 400 with the new data of the new data input unit 500, and recommends the most optimal retention period for the new data. It is preferable to be configured to include.

Furthermore, it is preferable that the deep learning learning unit 300 uses a Recurrent Neural Networks (RNN) technique to which a Long Short Term Memory (LSTM) method is applied.

Further, the data preprocessing unit 200 extracts the management department signature information or document name information from among the information included in the learning data, removes duplicate data, and performs morpheme separation analysis on the extracted information, It is desirable to perform ascending sorting by analyzing the frequency of appearance of morphemes.

Furthermore, the deep learning learning unit 300 receives morphemes within a preset frequency ranking order from the data preprocessor 200, generates an embedding layer, converts the received data, and then converts the input data into each of the hidden layers. It is desirable to perform learning by applying the LSTM (Long Short Term Memory) method to the node of.

According to another embodiment of the present invention, a method for recommending a record retention period using a recurrent artificial neural network technology is a learning data input step of receiving training data for learning from previously linked records information related databases in the training data input unit ( S100), in the data preprocessor, extracts specific information included in the learning data input by the learning data input step (S100), removes redundant data, and performs morpheme analysis on the extracted specific information Data preprocessing step (S200) analyzing the frequency of appearance for each morpheme, in the deep learning learning unit, one or more specific characteristics based on the frequency of appearance analyzed by the data preprocessing step (S200) using a preset deep learning technique. Deep learning learning step (S300) of receiving morphemes and performing learning, data management in which the data management unit generates reference data of the type of retention period for records using the learning result of the deep learning learning step (S300) Step (S400), in the new data input unit, a new data input step (S500) of receiving new data related to record information for setting the retention period from the outside, and in the retention period management unit, generated by the data management step (S400) It is preferable to perform a retention period recommendation step (S600) of mapping and comparing the reference data with the new data input by the new data input step (S500), and recommending the most optimal retention period for the new data. .

Furthermore, the data pre-processing step (S200) extracts the management department signature information or document name information from among the information included in the learning data, removes the redundant data, and performs morpheme separation analysis on the extracted information. It is desirable to perform ascending sorting by analyzing the frequency of appearance of morphemes.

Further, in the deep learning learning step (S300), the RNN (Recurrent Neural Networks) technique applied with the LSTM (Long Short Term Memory) method is used, but the morphemes within the frequency rank preset by the data preprocessing step (S200) are It is desirable to perform learning by applying the LSTM method to each node of the hidden layer after receiving the input and transforming the input data by creating an embedding layer.

The system for recommending the retention period of records using the recurrent artificial neural network technology of the present invention and the method for recommending the recordings according to the configuration as described above enables deep learning learning in creating a retention period for public records of public institutions or records of private companies. By performing pre-processing on the learning data prior to application, there is an advantage of improving the reliability of the calculation result (recommendation/inference of the retention period) by increasing the significance and accuracy of the learning data.

1 is a diagram showing a system for recommending a recording retention period using a recurrent artificial neural network technology according to an embodiment of the present invention.
2 is a flowchart illustrating a method of recommending a recording retention period using a recurrent artificial neural network technology according to an embodiment of the present invention. This is an example in which statistical analysis is performed and data is displayed.

Hereinafter, a system for recommending a record retention period using a recurrent artificial neural network technology of the present invention and a method for recommending the same will be described in detail with reference to the accompanying drawings. The drawings introduced below are provided as examples in order to sufficiently convey the spirit of the present invention to those skilled in the art. Accordingly, the present invention is not limited to the drawings presented below and may be embodied in other forms. In addition, the same reference numbers throughout the specification indicate the same elements.

In this case, unless there are other definitions in the technical terms and scientific terms used, they have the meanings commonly understood by those of ordinary skill in the art to which this invention belongs, and the gist of the present invention in the following description and accompanying drawings Descriptions of known functions and configurations that may unnecessarily obscure are omitted.

In addition, the system refers to a set of components including devices, devices, and means that are organized and regularly interact to perform a required function.

A system for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention, and a method for recommending the same, is a technology capable of inferring the retention period of records through the application of a deep learning technique.

In the case of the creation of the retention period during the processing of records by public institutions or private companies that are currently dependent on manual work, problems with reliability and consistency are emerging as the result of the subjective judgment of each record manager.

Therefore, through the record retention period recommendation system and the recommendation method using the recurrent artificial neural network technology according to an embodiment of the present invention, specific information is extracted from the records, and through morpheme separation analysis of the extracted specific information, significance At the same time, it has the advantage of improving the reliability/accuracy of the calculation result (recommended retention period) by applying deep learning technology.

FIG. 1 is a configuration diagram showing a system for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention. Referring to FIG. 1, a record using a recurrent artificial neural network technology according to an embodiment of the present invention The retention period recommendation system will be described in detail.

As shown in FIG. 1, a system for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention includes a learning data input unit 100, a data preprocessor 200, a deep learning learning unit 300, It is preferable to include a data management unit 400, a new data input unit 500, and a retention period management unit 600.

To learn more about each configuration,

It is preferable that the learning data input unit 100 receives learning data for learning from databases related to recorded material information linked in advance.

Specifically, the National Archives of Records is linked to a database of public records, a database of the Information Disclosure Act, a database of the Electronic Signature Act, a database of laws on the management of public records, a database of records management classification standards, a database of records with a preservation period set, It is desirable to receive learning data for learning.

It is preferable that the data preprocessor 200 extracts specific information included in the training data received from the training data input unit 100 through text mining and removes redundant data.

That is, the data preprocessing unit 200 processes the learning data received from the learning data input unit 100, that is, records-related information as metadata, an institution code, a unit work code, a year of production, and a serial number for registration. , Number of books, title of record, type of record, summary of contents, number of pages of record, number of records registered, number of electronic files, record classification, whether or not to modify, name of old record type production institution, old record type number, start year, end year, end of installation The production year that is set as meaningful information by the data preprocessing unit 200 among them, including the date, reason for non-preservation, preservation period, preservation method, preservation location, business person in charge, and handover category. It is desirable to extract at least one of information, management department name information, document name information, or retention period. In addition, it is desirable to remove duplicate input data since it is stored in multiple databases.

Thereafter, the data preprocessor 200 analyzes the frequency of appearance of each morpheme by performing morpheme analysis on the extracted specific information, that is, production year information, management department name information, document name information, or retention periods. It is desirable.

In this case, to perform morpheme analysis, it is most preferable to use a morpheme having the highest classification accuracy among words, words, and morphemes constituting a sentence.

It is preferable that the deep learning learning unit 300 receives one or more specific morphemes and performs learning based on the frequency of appearance analyzed by the data preprocessor 200 using a preset deep learning technique.

In this case, in the system for recommending a recording period using a recurrent artificial neural network technology according to an embodiment of the present invention, the data preprocessor 200 performs morpheme separation analysis on specific information, and simply In order not to analyze only the frequency of appearance, but to improve the accuracy/reliability of the result by improving the significance of morphemes, it is preferable to perform the sorting in ascending order.

Through this, the deep learning learning unit 300 selects a morpheme from 1st to 1000th of the appearance frequency as input data of the deep learning technique set in advance, and then generates and receives an embedding layer defining the vector length. It is desirable to transform the input data.

In particular, it is preferable that the deep learning learning unit 300 uses a Recurrent Neural Networks (RNN) technique to which a Long Short Term Memory (LSTM) method is applied.

Prior to the description of the RNN method to which the LSTM method is applied, the RNN method is when learning input data (time-series data) received in chronological order, the hidden layer has a memory function to store and learn each state. It is a neural network model used for

However, since the RNN technique only reflects the results of the near past in judgment, the longer the sentence or the distance between the sentence and the sentence becomes, the faint memory is, and thus there is a problem that it cannot be reflected in the previous context.

In order to solve this problem, the LSTM block replaces the neurons in the hidden layer of the RNN with the LSTM scheme.

For LSTM block, adjust with forget gate(ft) how much to reflect the previous result (cell state), and adjust with ignore gate(it) how much to reflect current input data and previous output data. It takes the form of a weighted average as a weight.

Using the RNN technique to which the LSTM method is applied, the deep learning learning unit 300 selects morphemes from the 1st to 1000th of the appearance frequency as input data, and the weight of the hidden layer is to limit the vector length to 200. desirable.

The data management unit 400 uses the learning result of the deep learning learning unit 300, that is, the output data obtained by the deep learning technique of the deep learning learning unit 300, to determine the type of retention period for the record. It is desirable to generate the data.

As for the type of retention period, it is preferable to limit it to a total of 9 labels that are no retention period, 1 year, 3 years, 5 years, 10 years, 25 years, 30 years, 40 years, and permanent, but this is in accordance with an embodiment of the present invention. It's just that.

In addition, in order to reduce the loss of the output data of the deep learning learning unit 300, the data management unit 400 is preferably defined using a softmax function, which is a technique already known in the relevant field and further described. Is omitted.

It is preferable that the new data input unit 500 receives new data related to record information for setting a retention period from an outside (record manager, etc.).

In other words, it is desirable to input data of a record that needs to be newly set, or that requires confirmation of an existing set retention period.

The retention period management unit 600 may map and compare the reference data of the data management unit 400 with the new data of the new data input unit 500 to recommend the most optimal retention period for the new data. .

In detail, the retention period management unit 600 includes retention period information recommended through deep learning learning of training data by a user using the recording retention period recommendation system using the recurrent artificial neural network technology according to an embodiment of the present invention. By mapping and comparing the last selected retention period information or modified and selected retention period information (new data) among the reference data of the data management unit 400, it can be stored and managed in a separate database (not shown). In this way, the retention period information selected by the user or the retention period information selected by modification can be used as information with a high level of reliability.

Through this, it is possible to reflect subtle differences in the retention period of records and the characteristics of each institution that manages records (using recommended retention period information and actually selected retention period information), tailored information for each user It is possible to increase the reliability of the recommendation of preservation agencies by actively providing the archives and reflecting the differentiation of the records management of each archive in the retention period recommendation (recommendation of the most optimal retention period).

FIG. 2 is a flow chart showing a method for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention. Referring to FIG. 2, a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention The recommended method will be described in detail.

As shown in FIG. 2, a method for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention includes a learning data input step (S100), a data preprocessing step (S200), and a deep learning learning step (S300). , It is preferable to consist of a data management step (S400), a new data input step (S500), and a retention period recommendation step (S600).

To learn more about each step,

In the learning data input step (S100), it is preferable that the learning data input unit 100 receives learning data for learning from previously linked databases related to recording information.

Specifically, it is desirable to receive learning data for learning in connection with the Information Disclosure Act database, the electronic signature method database, the law database on the management of public records, the records management classification standard database, and the record database for which the retention period is set.

In the data pre-processing step S200, it is preferable that the data pre-processing unit 200 extracts specific information included in the training data received by the training data input step S100 and removes duplicate data.

In this case, the data pre-processing step (S200) is most preferably performed through text mining.

Specifically, the learning data input by the learning data input step (S100), that is, the information related to the record is processed as metadata and the institution code, the unit work code, the year of production, the registration serial number, the number of volumes, and the record. Steel title, type of record, summary of contents, number of pages of record, number of records registered, number of electronic files, record classification, whether or not to be modified, name of old record material production institution, old record material classification number, start year, end year, date of installation, and reason , Preservation period, preservation method, storage place, business person in charge, transfer category, etc., among them, the management department's signature information or document name information that is set as meaningful information in the data preprocessor 200 through text mining. It is preferable to extract. In addition, it is desirable to remove duplicate input data since it is stored in multiple databases.

Thereafter, in the data pre-processing step S200, it is preferable to analyze the frequency of appearance of each morpheme by performing morpheme analysis on the extracted specific information, that is, management department signature information or document name information.

In this case, to perform morpheme analysis, it is most preferable to use a morpheme having the highest classification accuracy among words, words, and morphemes constituting a sentence.

In the deep learning learning step (S300), the deep learning learning unit 300 receives and learns one or more specific morphemes based on the frequency of appearance analyzed in the data preprocessing step (S200) by using a preset deep learning technique. It is preferable to perform.

To this end, in the method for recommending a record retention period using a recurrent artificial neural network technology according to an embodiment of the present invention, through the data pre-processing step (S200), morpheme separation analysis for specific information is performed, and simply each morpheme is It is preferable to perform sorting in ascending order in order to improve the accuracy/reliability of the result by improving the significance of morphemes, not just analyzing the frequency of occurrence of the morpheme.

Through this, in the deep learning learning step (S300), after selecting a morpheme from the 1st to 1000th of the appearance frequency as input data of a preset deep learning technique, an embedding layer defining the vector length is generated and input It is desirable to convert the received input data.

In particular, in the deep learning learning step (S300), it is preferable to use a Recurrent Neural Networks (RNN) technique to which the LSTM (Long Short Term Memory) method is applied. Prior to the description of the RNN technique to which the LSTM method is applied, the RNN technique is time When learning input data (time-series data) received in sequence, the hidden layer has a memory function, so it is a neural network model that stores each state and uses it for learning.

However, since the RNN technique only reflects the results of the near past in judgment, the longer the sentence or the distance between the sentence and the sentence becomes, the faint memory is, and thus there is a problem that it cannot be reflected in the previous context.

In order to solve this problem, the LSTM block replaces the neurons in the hidden layer of the RNN with the LSTM scheme.

For LSTM block, adjust with forget gate(ft) how much to reflect the previous result (cell state), and adjust with ignore gate(it) how much to reflect current input data and previous output data. It takes the form of a weighted average as a weight.

Using the RNN technique to which the LSTM method is applied, the deep learning learning unit 300 selects morphemes from the 1st to 1000th of the appearance frequency as input data, and the weight of the hidden layer is preferably to limit the vector length. .

In the data management step (S400), the data management unit 400 uses the learning result of the deep learning learning step (S300), that is, the output data according to the deep learning technique, to determine the type of retention period for the record. It is desirable to generate the data.

As for the type of retention period, it is preferable to limit it to a total of 9 labels that are no retention period, 1 year, 3 years, 5 years, 10 years, 25 years, 30 years, 40 years, and permanent, but this is in accordance with an embodiment of the present invention. It's just that.

In addition, in order to reduce the loss of the output data by the deep learning technique, it is preferable to define it using a softmax function, which is a technique already known in the relevant field, and further description is omitted.

In the new data input step (S500), it is preferable that the new data input unit 500 receives new data related to record information for setting a retention period from an outside (record manager, etc.).

In other words, it is desirable to input data of a record that needs to be newly set, or that requires confirmation of an existing set retention period.

In the retention period recommendation step (S600), the retention period management unit 600 includes the reference data generated by the data management step (S400) and the new data input by the new data input step (S500). It is desirable to map and compare, and recommend the most optimal retention period for new data.

In detail, the retention period recommendation step (S600) is the retention period information recommended through deep learning learning of the training data by a user who uses the recording retention period recommendation method using the recurrent artificial neural network technology according to an embodiment of the present invention. By mapping and comparing the last selected retention period information or modified and selected retention period information (new data) among (reference data of the data management unit 400), it can be stored and managed in a separate database (not shown). In this way, the retention period information selected by the user or the retention period information selected by modification can be used as information with a high level of reliability.

Through this, it is possible to reflect subtle differences in the retention period of records and the characteristics of each institution that manages records (using recommended retention period information and actually selected retention period information), tailored information for each user It can increase the credibility of the recommendation of preservation agencies by actively providing the archives and reflecting the differentiation of the records management of each archive in the retention period recommendation (recommendation of the most optimal retention period).

That is, in other words, the system for recommending the retention period of the records using the recurrent artificial neural network technology according to an embodiment of the present invention and the recommendation method thereof, in inferring the retention period of records by applying a deep learning technique, include learning data (input data ) Has the advantage of improving the accuracy and reliability of the result data, effectively improving the accuracy of the recommended retention period.

As described above, in the present invention, specific matters such as specific components, etc. and limited embodiments have been described, but this is provided only to aid in a more general understanding of the present invention, and the present invention is limited to the above-described embodiment. It is not, and those of ordinary skill in the field to which the present invention belongs can make various modifications and variations from this description.

Therefore, the spirit of the present invention is limited to the described embodiments and should not be determined, and all things equivalent or equivalent to the claims as well as the claims to be described later belong to the scope of the spirit of the present invention. .

100: learning data input unit
200: data preprocessor
300: Deep Learning Department
400: data management unit
500: new data input unit
600: Retention period management department

Claims (7)

A learning data input unit 100 receiving learning data for learning from previously linked records information-related databases;
After extracting specific information included in the learning data received from the learning data input unit 100 through text mining, removing duplicate data, and performing morpheme analysis on the extracted specific information, the appearance of each morpheme A data preprocessing unit 200 for analyzing frequency;
A deep learning learning unit 300 that receives one or more specific morphemes based on the frequency of appearance analyzed by the data preprocessor 200 using a preset deep learning technique and performs learning;
A data management unit 400 for generating reference data of a retention period type for a record using the learning result of the deep learning learning unit 300;
A new data input unit 500 for receiving new data related to record information for setting a retention period from the outside; And
A retention period management unit 600 that maps and compares reference data of the data management unit 400 with new data of the new data input unit 500, and recommends the most optimal retention period for new data;
Record retention period recommendation system using a recurrent artificial neural network technology, characterized in that configured to include.
The method of claim 1,
The deep learning learning unit 300
A system for recommending a record retention period using a recurrent artificial neural network technology, characterized by using a Recurrent Neural Networks (RNN) technique applying a Long Short Term Memory (LSTM) method.
The method of claim 2,
The data preprocessor 200
After extracting the management department signature information or document name information among the information included in the learning data, removing duplicate data, and performing morpheme separation analysis on the extracted information, the frequency of appearance of each morpheme is analyzed and sorted in ascending order. A system for recommending record retention periods using recurrent artificial neural network technology, characterized in that performing.
The method of claim 3,
The deep learning learning unit 300
After receiving the morphemes within a preset frequency ranking from the data preprocessor 200, converting the input data by creating an embedding layer, a Long Short Term Memory (LSTM) method is applied to each node of the hidden layer. Record retention period recommendation system using recurrent artificial neural network technology, characterized in that the learning is performed by applying.
In the learning data input unit, a learning data input step (S100) of receiving learning data for learning from previously linked records information-related databases;
In the data preprocessor, after extracting specific information included in the learning data input by the learning data input step (S100), removing duplicate data, and performing morpheme analysis on the extracted specific information, each morpheme Data pre-processing step of analyzing the frequency of appearance of (S200);
A deep learning learning step (S300) of receiving one or more specific morphemes based on the frequency of appearance analyzed by the data preprocessing step (S200) using a preset deep learning technique and performing learning;
A data management step (S400) of generating, in the data management unit, reference data of a retention period type for a record using the learning result of the deep learning learning step (S300);
In the new data input unit, a new data input step (S500) of receiving new data related to record information for setting a retention period from outside; And
The retention period management unit maps and compares the reference data generated by the data management step (S400) with the new data input by the new data input step (S500), and the most optimal storage for new data Retention period recommendation step of recommending a period (S600);
A method for recommending a record retention period using a recurrent artificial neural network technology, characterized in that it consists of.
The method of claim 5,
The data pre-processing step (S200)
After extracting the management department signature information or document name information among the information included in the learning data, removing duplicate data, and performing morpheme separation analysis on the extracted information, the frequency of appearance of each morpheme is analyzed and sorted in ascending order. A method of recommending a record retention period using a recurrent artificial neural network technology, characterized in that performing.
The method of claim 6,
The deep learning learning step (S300)
Using the Recurrent Neural Networks (RNN) technique applying the Long Short Term Memory (LSTM) method,
By receiving the morphemes in the pre-processing step (S200), the morphemes within the pre-set frequency ranking are received, the embedding layer is created and the received data is converted, and then the LSTM method is applied to each node of the hidden layer to perform learning. A method for recommending a record retention period using a recurrent artificial neural network technology, characterized in that:

Images