KR102507925B1 - Method for Korean correction service and apparatus for performing the method - Google Patents

Abstract

The present invention relates to a method for providing a Korean language correction service and a device for performing the method. The method for providing a Korean language correction service comprises the steps of: a data center collecting original texts and generating erroneous sentences for the original texts to generate a learning data set; a grammar error detection/grammar error correction (GED/GEC) service engine performing grammar correction learning for extracting grammatical errors based on the learning data set; and a Korean language correction service platform providing part providing a grammar error correction service, a semantic word recommendation service, an erratum correction service or a punctuation correction service based on the GED/GEC service engine. Therefore, accuracy in the vocabulary command of a user can be improved.

Description

Method for providing Korean correction service and apparatus for performing the method {Method for Korean correction service and apparatus for performing the method}

The present invention relates to a method for providing a Korean language correction service and an apparatus for performing the method. More specifically, it relates to a method for providing a Korean language correction service that provides supportable learning tools so that correct sentences can be written and read in any environment within an online environment, and a device for performing the method.

Due to the emergence of various online video media and platforms, students who are easily exposed to the media environment are accustomed to acquiring information through watching media videos rather than reading, which leads to problems in literacy.

Currently, education offices and institutions across the country are establishing policies and introducing learning tools to promote literacy at the national level, and the core of literacy promotion has been specified as strengthening reading/writing skills.

In order to strengthen the reading and writing skills of Generation Z adapted to online learning, learning tools that can support writing and reading correct sentences in any environment in the online environment are needed. This can help.

The global population of Korean speakers is about 81 million, and Korean ranks 14th according to the number of native speakers among all languages in the world. .

According to the '2021 Overseas Korean Language Education Support Project Basic Plan' announced by the Ministry of Education, the government plans to invest about 23.6 billion won to promote Korean language education textbooks and teachers abroad, and to develop 'Overseas Korean language curriculum for elementary and secondary schools'. Through this, we plan to promote systematic expansion of Korean language education. However, due to the linguistic characteristics of Korean (word order/sentence structure/grammar/pronunciation/homonyms), it is considered one of the most difficult languages for foreigners, and the lack of Korean self-assessment/learning tools causes problems with accessibility to Korean education. .

To this end, research is being conducted on question-answering systems, translation systems, summary systems, etc. of various documents using artificial intelligence language models. In data construction, there are still many documents stored in paper form, especially in libraries, which carry out a DB project for digitalization of more than 10 billion documents every year. As for the performance of domestic and foreign OCR programs, the OCR performance of documents mixed with Korean/English/Chinese characters/special characters is still around 70-80%.

Therefore, the OCR post-processing (POST-OCR) process through GEC technology is essential, and the development of a grammar and semantic-based Korean sentence error detection and sentence error correction (Grammar Error Detection + Grammar Error Correction) system is essential for Korean language education. Research is needed to increase accessibility.

The object of the present invention is to solve all of the above problems.

In addition, an object of the present invention is to help users use accurate vocabulary based on a system for detecting and correcting Korean sentence errors based on grammar and semantics.

In addition, an object of the present invention is to develop grammatical ability that can express accurately and effectively according to the ability to use sentences and words appropriate to the situation and the elements of language.

Representative configurations of the present invention for achieving the above object are as follows.

According to an embodiment of the present invention, a method for providing Korean language correction service includes the steps of a data center collecting original text and generating an error sentence for the original text to generate a training data set, GED/GEC (grammar error detection/grammar error correction) ) The service engine performs grammar correction learning to extract grammatical errors based on the learning data set, and the Korean correction service platform provider provides grammatical error correction service and semantic word recommendation service based on the GED / GEC service engine , providing a typo correction service or a punctuation mark correction service.

Meanwhile, the GED/GEC service engine includes a data noising unit, the data noising unit includes a plurality of data set generating modules, and each of the plurality of data set generating modules generates the above information for the original text based on the original text. It can be implemented to generate an error statement.

In addition, the GED/GEC service engine includes a transform encoder, and the transform encoder defines and maps a relationship between the original text and the erroneous sentence based on a tag matched to the analysis unit of the original text and the analysis unit of the erroneous sentence. learning can be performed.

According to another embodiment of the present invention, a Korean language correction system providing a Korean language correction service includes a data center implemented to collect original texts and generate error sentences for the original texts to generate a learning data set, based on the learning data set GED/GEC (grammar error detection/grammar error correction) service engine implemented to perform grammar correction learning to extract grammatical errors and grammar error correction service based on the GED/GEC service engine, semantic word recommendation service, misspelling A Korean correction service platform providing unit implemented to provide a correction service or a punctuation mark correction service may be included.

Meanwhile, the GED/GEC service engine includes a data noising unit, the data noising unit includes a plurality of data set generating modules, and each of the plurality of data set generating modules generates the above information about the original text based on the original text. It can be implemented to generate an error statement.

In addition, the GED/GEC service engine includes a transform encoder, and the transform encoder defines and maps a relationship between the original text and the erroneous sentence based on a tag matched to the analysis unit of the original text and the analysis unit of the erroneous sentence. learning can be performed.

According to the present invention, the user's ability to use accurate vocabulary can be improved based on the system for detecting and correcting Korean sentence errors based on grammar and semantics.

In addition, according to the present invention, the ability to use sentences/words appropriate to the situation and the grammatical ability to express accurately and effectively according to language elements can be cultivated.

1 is a conceptual diagram showing a Korean language correction system according to an embodiment of the present invention.
2 is a conceptual diagram illustrating the operation of a data center according to an embodiment of the present invention.
3 is a conceptual diagram illustrating the operation of a GED/GEC service engine according to an embodiment of the present invention.
4 is a conceptual diagram illustrating the operation of a GED/GEC service engine according to an embodiment of the present invention.
5 is a conceptual diagram illustrating the operation of a data set generation module according to an embodiment of the present invention.
6 is a conceptual diagram illustrating the operation of a data set generation module according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The detailed description of the present invention which follows refers to the accompanying drawings which illustrate, by way of illustration, specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable any person skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different from each other but are not necessarily mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented from one embodiment to another without departing from the spirit and scope of the present invention. It should also be understood that the location or arrangement of individual components within each embodiment may be changed without departing from the spirit and scope of the present invention. Therefore, the detailed description to be described later is not performed in a limiting sense, and the scope of the present invention should be taken as encompassing the scope claimed by the claims and all scopes equivalent thereto. Like reference numbers in the drawings indicate the same or similar elements throughout the various aspects.

Hereinafter, various preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings in order to enable those skilled in the art to easily practice the present invention.

1 is a conceptual diagram showing a Korean language correction system according to an embodiment of the present invention.

In FIG. 1, a Korean language correction system capable of grammatical and semantic correction of errors in a sentence beyond the existing Korean spelling correction is disclosed.

Referring to FIG. 1 , the Korean language correction system may include a data center 100, a grammar error detection/grammar error correction (GED/GEC) service engine 120, and a Korean language correction service platform provider 140.

The data center 100 may be implemented to collect original text and erroneous sentences by being connected to an artificial intelligence (AI) hub. In addition, the data center 100 may be implemented to construct a synthetic dataset using a data-noising technique.

Also, the data center 100 may generate an error sentence based on a data set generation module. Specifically, the data set generation module may generate an error sentence corresponding to the original text by applying a self-defined error generation rule to a given sentence, or generate an error sentence corresponding to the original text by randomly applying space errors to the given sentence.

Data collected by the data center 100 may be used for learning of the GED/GEC service engine 120 .

The GED/GEC service engine 120 may provide a correction service for an erroneous sentence input by a user.

The GED/GEC service engine 120 may be a GED/GEC model based on a transformer encoder using a sequence tagging technique.

In addition, the GED/GEC service engine 120 may provide a semantic word recommendation service based on a word embedding model.

Error occurrence rules such as spelling of words separated into nouns, verbs, adverbs, and adjectives, G2P (Grapheme To Phoneme), sound rules, and abbreviations are classified as self-defined code values (tags), and the classified error occurrence rules are It can be applied to the collected raw data for learning Korean.

A self-defined code value (tag) is defined as a tag dictionary, and an annotator that creates a training dataset by defining the relationship between words in the original text and words in the error sentence based on the tag dictionary and mapping the tags can be created

Learning of the transform encoder model may be performed using training data generated through the annotator. An error tag value for each input token (sentence) can be predicted by stacking a linear layer and a softmax layer on top of the transform encoder.

The Korean language correction service platform providing unit 140 may be implemented to provide a Korean language correction service. The Korean language correction service may include a grammatical error correction service, a semantic word recommendation service, a typo correction service, and a punctuation mark correction service. .

Hereinafter, in the embodiment of the present invention, specific operations of each component are disclosed.

2 is a conceptual diagram illustrating the operation of a data center according to an embodiment of the present invention.

2 discloses a method of collecting and generating data for learning of a GED/GEC service engine in a data center.

Referring to FIG. 2 , the data center may include a training data collection unit 230, an original text storage unit 210, a crawling engine unit 200, and a data noise unit 220.

The training data set unit 230 may be a data set including an original text and erroneous sentences for the original text. Error sentences for the original text may be generated based on the data noise unit 220 and stored in the training data set unit 230 . A data set including the original text and erroneous sentences stored in the training data set unit 230 may be transferred to the GED/GEC service engine.

The original text storage unit may be implemented to store the original text and transmit the original text to the data noising unit 220 and the GED/GEC service engine. The original text storage unit may receive the original text from the crawling engine unit.

Crawling engine unit 200 may be implemented to collect the original text from an external server through crawling. Original information for learning may be collected through various documents (news, Wikipedia, etc.) existing on the web.

The data noising unit 220 may be implemented to generate erroneous sentences, and a synthetic dataset using a data noising technique may be constructed.

The data noise unit 220 includes a first dataset generation module 250 using the G2PK Python module, a second dataset generation module 260 that applies a self-defined error generation rule to a given sentence, and spacing errors. It may include a third dataset generation module 270 that randomly applies to a given sentence.

Original text data collected by the crawling engine unit 200 may be generated as a learning data set in which original text/error sentences are paired by passing through a noise module.

A data center can create a training data set in two ways. The first learning data set generation method is a method of defining raw data as error data (or error sentences) and then generating error correction data (error correction sentences). The second learning data set generation method is a method of generating error data (or error sentences) after defining raw data as error-free data.

In the present invention, the first learning data set generation method and the second training data set generation method may be set differently according to data sources. Based on the judgment result of the existing GED / GEC service engine for the data source crawling by the crawling engine unit, if the error rate is greater than the critical percentage as a result of the judgment of the GED / GEC service engine, the data source is 1 A training data set can be created using the learning data set creation method. Conversely, as a result of the determination of the GED/GEC service engine, when the error rate is less than the critical percentage, the corresponding data source may generate a training data set using the second training data set generation method.

The error rate can be adaptively changed. For example, an error rate generated as a result of data accumulation may be determined by assigning a little more weight to the latest data. A data division period may be set in consideration of a new generation rate of data. As the new generation rate increases, the data division period may be determined to be shorter. An error rate may be determined by assigning a higher weight to data collected in a recent data partitioning period among data collected in a plurality of data partitioning periods. That is, when data is crawled relatively much, the data division period can be set relatively short, and conversely, when data is crawled relatively little, the data division period can be set relatively long.

Also, according to an embodiment of the present invention, selection of a first training data set generation method or a second training data set generation method for a specific data source may be performed when data is accumulated to a threshold amount of data or more. If the amount of data accumulated in a specific data source is less than the threshold data amount, the first training data set generation method and the second training data set generation method may be simultaneously applied to the same data to generate a data set.

When the first training data set generation method and the second training data set generation method are used simultaneously, the first error rate and the second training data set of the judgment result of the GED/GEC service engine based on the first training data set generation method A second error rate of the determination result of the GED/GEC service engine may be determined based on the generation method.

When the difference between the first error rate and the second error rate exceeds the threshold value, the first training data set generation method or the second training data set generation method having a smaller error rate is selected, and thereafter, the selected method is the main generation method can be determined and used.

In addition, according to an embodiment of the present invention, when error data is generated in the second learning data set generation method, the learning data set may be created by defining the location of error occurrence and the number of occurrences of errors differently in the original text.

When generating an erroneous sentence, the data noising unit 220 may determine the number of errors based on the length of the sentence and the type of grammatical parts of speech located in the sentence, and group the sentences to determine the location of the error occurrence. More specifically, as the length of the sentence is relatively long, the number of errors is relatively increased, and among parts of speech, parts of speech with a lot of errors (for example, verbs with many variations, postpositions, etc.) are relatively prioritized to generate errors. can make it In addition, in order to determine the position of occurrence of an error, a plurality of parts of speech constituting a sentence may be grouped into sub-sentence units based on nouns and verbs. Among the grouped sub-sentence groups, a sub-sentence group including a relatively large number of parts of speech with frequent errors may be determined as an error occurrence position. When setting a subsentence group, the first group is created by grouping other parts of speech located before and after the noun and verb first, and then the rest of the parts of speech are included in the nearest first group, so that the subsentence group is can be determined

3 is a conceptual diagram illustrating the operation of a GED/GEC service engine according to an embodiment of the present invention.

3 discloses a method for performing learning based on data transferred from a data center in a GED/GEC service engine.

Referring to FIG. 3 , the GED/GEC service engine may receive a training data set paired with original/erroneous sentences.

The GED/GEC service engine may perform pre-processing to generate a sequence tagging-based learning data set through pre-processing.

The pre-processing unit 300 may generate original text data, erroneous sentence data, and converted data. Conversion data can be divided into nouns, verbs, adverbs, and adjectives for analysis units of original sentences and error sentences, and can be assigned self-defined tag values such as spelling, G2P (Grapheme To Phoneme), sound rules, and abbreviations.

For example, for analysis units such as $KEEP, $NOUN_xxxx, $VB_xxx, $ADV_xxx, $ADJ_xxxx, etc., the part of analysis of the unit of analysis and information about the error type generated in the corresponding part of speech may be tagged.

The preprocessed training data set is based on the transformer-based encoder 350, based on the tag, the relationship between the analysis unit (e.g., word) of the original text and the analysis unit (e.g., word) of the error sentence Learning can be performed by defining and mapping. Through this learning process, the transform encoder 350 may predict an error tag value for an input sentence (input token).

That is, the conversion encoder 350 may perform learning by defining and mapping a relationship between the original text and the erroneous sentence based on a tag matched to the analysis unit of the original text and the analysis unit of the erroneous sentence.

As disclosed in the drawing, Guji->God ($ADV_G2P) is an adverb, which is a part of speech that is an analysis unit, such as grammatical errors caused by using actual pronunciation as it is, and so far->so ($KEEP) error-free parts. and information on the reason for occurrence of grammatical errors may be tagged.

Through the conversion encoder 350, GED detects grammatical errors, GEC is implemented to correct grammatical errors, and corrected sentences based on GED and GEC may be output.

4 is a conceptual diagram illustrating the operation of a GED/GEC service engine according to an embodiment of the present invention.

In FIG. 4, a GED/GEC service engine performs learning based on data transmitted from a data center, and a calibration method using a conversion-based GED/GEC engine and a knowledge-based GED/GEC engine is disclosed.

Referring to FIG. 4, the GED/GEC service engine performs fine-tuning using training data including additionally collected actual error sentences, and the GED/GEC prediction model based on the fine-tuning Upgrading can be done.

(1) Transformation-based predictive model (400)

The transform-based prediction model 400 may be a transform encoder-based prediction model disclosed in FIG. 3 . There is a possibility of more than one tag conversion occurring between a pair of source/erroneous words. Accordingly, an iterative tagging method may be applied in which the prediction result value of the conversion model is input back to the prediction model.

(2) knowledge-based prediction model (420)

The knowledge-based prediction model 420 can perform error detection and error correction based on knowledge of error rules that are not recognized by the transformation-based prediction model.

Also, the GED/GEC service engine may include a word recommendation system. A word recommendation system is a word recommendation system that proposes an optimal word by measuring the semantic similarity of words through transformer-based sentence embedding using a collected raw dataset. A word recommendation system may be implemented to suggest N most similar words.

Users can access the GED/GEC service engine through the REST API and receive correction services on the Korean correction service platform provider.

5 is a conceptual diagram illustrating the operation of a data set generation module according to an embodiment of the present invention.

5 discloses a method for the data set generation module to generate an error sentence according to an error generation rule.

Referring to FIG. 5 , a second dataset generation module is implemented to generate an error sentence by applying a self-defined error generation rule to a given sentence.

The second dataset generation module must define error generation rules for grammatical errors that users may generate in order to apply the error generation rules.

The error generation rule defined by the second dataset generation module is defined as n separate error generation rules, and each of the n error generation rules sets a plurality of error generation levels according to the degree of error in the error sentence. can cause

Each of the n error generating rules may be set for each country. When learning Korean in different countries, errors may occur differently due to grammatical habits of each language in different countries. Therefore, each of the n error generation rules may be defined differently in consideration of the language region.

For example, the first error generating rule is an error generating rule for Koreans (Koreans), the second error generating rule is an error generating rule for learning Korean for Chinese people, and the third error generating rule is for Japanese people learning Korean. This is an error generating rule, and the fourth error generating rule may be an error generating rule for learning Korean in the United States.

After the n-th error generation rule is defined in this way, n error generation levels may be hierarchically defined by adjusting the degree of errors generated in each of the n-th error generation rules.

For example, the first error generation rule for Koreans is defined as 10 first error occurrence levels to 10th error occurrence levels, and smaller errors occur from the first error occurrence level to the 10th error occurrence level. It can be. That is, if the error sentence generated based on the first error occurrence level is the level of an error sentence that a first grader of elementary school can generate, the error sentence generated based on the tenth error occurrence level is an error sentence that a college student can generate. may be the level of

In this way, error generation rules and error occurrence levels are determined based on user information, and based on this, different education data according to the user's education level can be generated and provided for Korean language learning.

In order to determine these error generation rules and error generation levels, error sentences generated based on user information (nationality, education, etc.) are input, and based on these, error generation rules and error generation levels may be individually set.

In addition, a faster Korean language correction service can be provided based on these error generation rules and error generation levels. More specifically, based on the user's education level, error detection is performed first for errors that occur frequently in the user's education level, and errors that are highly unlikely to occur at a specific education level are judged in a later order to provide a faster error detection service. may be provided.

That is, when the Korean language correction service is provided, error detection rules verified according to the user's education level may be set to different priorities and verified. Based on this, error detection can be performed more quickly.

Also, in the third data set generating module for applying spacing errors to given sentences according to the spacing rules of the present invention, n error generation rules and n' error generation levels may be separately defined.

6 is a conceptual diagram illustrating the operation of a data set generation module according to an embodiment of the present invention.

6 discloses an integration method between error generation levels and error generation rules in the data set generation module.

Referring to FIG. 6 , an integrated processing method through comparison between n error generation rules and n error occurrence levels in each of a second data set generation module and a third data set generation module is disclosed. For convenience of description, the second data set generation module is assumed, but the same may be applied to the third data set generation module.

Rule integration and/or level integration may be performed when the number of common error rules between similar levels is greater than or equal to a threshold ratio. For example, when the similarity between the plurality of error generating rules included in the first error generating rule 610 and the plurality of error generating rules included in the fifth error generating rule 650 is greater than or equal to the threshold similarity, the error generating rules are integrated. this can be done

After firstly integrating between error generating rules, integration between error generating levels may be secondarily performed based on similarities between error generating levels. For example, error generation levels having similar error generation rules between error generation levels may be integrated through comparison between a plurality of error generation levels of different error generation rules.

Specifically, when the first error generation rule 610 and the fifth error generation rule 650 are integrated, the plurality of error generation levels included in the first error generation rule 610 and the fifth error generation rule 650 A comparison between a plurality of included error occurrence levels may be performed.

Integration between error generation levels may be performed based on a comparison between the plurality of error generation levels included in the first error generation rule 610 and the plurality of error generation levels included in the fifth error generation rule 650 . For example, integration between the second error generation level 620 of the first error generation rule 610 and the third error generation level 630 of the fifth error generation rule 650 is performed and one error generation level is formed. can be integrated and managed.

In this way, by integrating and managing error generation rules and error occurrence levels, it is possible to easily secure learning data and to easily generate educational data for Korean language education corresponding to such error occurrence levels.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. The computer readable recording medium may include program instructions, data files, data structures, etc. alone or in combination. Program instructions recorded on the computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. medium), and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language codes that can be executed by a computer using an interpreter or the like as well as machine language codes generated by a compiler. A hardware device may be modified with one or more software modules to perform processing according to the present invention and vice versa.

Although the present invention has been described above with specific details such as specific components and limited embodiments and drawings, these are only provided to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments, and the present invention Those with ordinary knowledge in the technical field to which the invention belongs may seek various modifications and changes from these descriptions.

Therefore, the spirit of the present invention should not be limited to the above-described embodiments and should not be determined, and all scopes equivalent to or equivalently changed from the claims as well as the claims described below are within the scope of the spirit of the present invention. will be said to belong to

Claims (6)

How to provide Korean language correction service,
generating a training data set by a data center collecting original texts and generating erroneous sentences for the original texts;
performing, by a grammar error detection/grammar error correction (GED/GEC) service engine, grammar correction learning to extract grammatical errors based on the training data set;
A Korean correction service platform providing unit providing a grammatical error correction service, a semantic word recommendation service, a typo correction service, or a punctuation mark correction service based on the GED / GEC service engine,
The GED / GEC service engine includes a data noise unit,
The data noising unit includes a plurality of data set generating modules,
Each of the plurality of data set generating modules is implemented to generate the error sentence for the original text based on the original text,
The learning data set is generated based on a first learning data set generation method and a second learning data set generation method;
The first learning data set generation method is a method of defining raw data as error data and then generating error correction data;
The second learning data set generation method is a method of generating error data after defining raw data as error-free data,
The first learning data set generation method and the second learning data set generation method are set differently according to data sources,
For the data source crawled by the crawling engine unit, if the error rate is greater than or equal to a critical percentage based on the existing determination result of the GED/GEC service engine, the data source generates a learning data set using the first learning data set generation method. And if the error rate is less than the critical percentage, the data source generates a learning data set using the second learning data set generation method. According to claim 1,
Selection of the first learning data set generation method or the second training data set generation method for the data source is performed when the data of the data source is accumulated to a threshold amount of data or more,
When the amount of data accumulated in the data source is less than the threshold data amount, the first training data set generation method and the second training data set generation method are simultaneously applied to the same data to create a data set;
When the first learning data set generation method and the second learning data set generation method are simultaneously used, a first error rate for a determination result of the GED/GEC service engine based on the first learning data set generation method and A second error rate for the determination result of the GED/GEC service engine is determined based on the second learning data set generation method,
When the difference between the first error rate and the second error rate exceeds a threshold value, the first training data set generation method or the second training data set generation method having a smaller error rate is selected, and then the selected data set is generated A method for providing Korean language correction service, characterized in that the method is determined by the main generation method. According to claim 2,
When generating an erroneous sentence, the data noise unit determines the number of errors based on the length of the sentence and the type of grammatical parts of speech located in the sentence;
The data noise unit groups sentences to determine an error occurrence position,
The data noise unit relatively increases the number of errors occurring as the length of the sentence is relatively long,
The data noise unit causes an error by setting a relatively high priority for parts of speech that have a lot of errors among parts of speech,
The data noising unit groups a plurality of parts of speech constituting a sentence into sub-sentence units based on nouns and verbs in order to determine the location of the error occurrence,
The method of claim 1 , wherein the data noise unit determines a sub-sentence group including a relatively large number of part-of-speech errors among the grouped sub-sentence groups as the location of occurrence of the error. The Korean proofreading system, which provides Korean proofreading services,
a data center configured to collect original texts and generate erroneous sentences for the original texts to generate a training data set;
a grammar error detection/grammar error correction (GED/GEC) service engine implemented to perform grammar correction learning for extracting grammatical errors based on the training data set;
A Korean correction service platform providing unit implemented to provide a grammatical error correction service, a semantic word recommendation service, a typo correction service, or a punctuation mark correction service based on the GED / GEC service engine,
The GED / GEC service engine includes a data noise unit,
The data noising unit includes a plurality of data set generating modules,
Each of the plurality of data set generating modules is implemented to generate the error sentence for the original text based on the original text,
The learning data set is generated based on a first learning data set generation method and a second learning data set generation method;
The first learning data set generation method is a method of defining raw data as error data and then generating error correction data;
The second learning data set generation method is a method of generating error data after defining raw data as error-free data,
The first learning data set generation method and the second learning data set generation method are set differently according to data sources,
For the data source crawled by the crawling engine unit, if the error rate is greater than or equal to a critical percentage based on the existing determination result of the GED/GEC service engine, the data source generates a learning data set using the first learning data set generation method. And when the error rate is less than the critical percentage, the data source generates a learning data set using the second learning data set generation method. According to claim 4,
Selection of the first learning data set generation method or the second training data set generation method for the data source is performed when the data of the data source is accumulated to a threshold amount of data or more,
When the amount of data accumulated in the data source is less than the threshold data amount, the first training data set generation method and the second training data set generation method are simultaneously applied to the same data to create a data set;
When the first learning data set generation method and the second learning data set generation method are simultaneously used, a first error rate for a determination result of the GED/GEC service engine based on the first learning data set generation method and A second error rate for the determination result of the GED/GEC service engine is determined based on the second learning data set generation method,
When the difference between the first error rate and the second error rate exceeds a threshold value, the first training data set generation method or the second training data set generation method having a smaller error rate is selected, and then the selected data set is generated A Korean correction system, characterized in that the method is determined by the main generation method. According to claim 5,
When generating an erroneous sentence, the data noise unit determines the number of errors based on the length of the sentence and the type of grammatical parts of speech located in the sentence;
The data noise unit groups sentences to determine an error occurrence position,
The data noise unit relatively increases the number of errors occurring as the length of the sentence is relatively long,
The data noise unit causes an error by setting a relatively high priority for parts of speech that have a lot of errors among parts of speech,
The data noising unit groups a plurality of parts of speech constituting a sentence into sub-sentence units based on nouns and verbs in order to determine the location of the error occurrence,
The Korean language correction system, characterized in that the data noise unit determines a sub-sentence group containing a relatively large number of parts of speech in which errors occur among the grouped sub-sentence groups as the error occurrence position.

Images