KR20200057824A - Word spelling correction system - Google Patents

Abstract

Provided is a word correction system which performs word semantic embedding on typos. According to the present invention, the word correction system comprises: a user interface that receives a user input word from a user through a network; a processing data generation unit which separates words for learning and user input words by morphemes, generates processing data for analysis by separating the separated morphemes for each element, and generates an n-gram list by dividing the analysis processing data into n-grams; a word embedding learning unit that performs word embedding learning on a list of n-grams through a skip-gram model; and a semantic vector generator for generating a semantic vector from the n-gram list.

Description

Word spelling correction system

The present invention relates to a word correction system, and more particularly, to a system for correcting misspellings.

The present invention was derived from a study conducted by Saltlux Co., Ltd. as part of the SW Computing Industry Source Technology Development Project (SW) by the Ministry of Trade, Industry and Energy. [Research period: 2017.10.01. ~ 2018.09.30., Research institute: Korea Advanced Institute of Industrial Technology, Research project name: Development of a mobile-optimized multimodal question and answer framework, task identification number: N0001701]

Natural language refers to the language that people use everyday, and in computer science, natural language processing is a process that enables human language phenomena to be simulated using machines such as computers. Therefore, the most essential element in natural language processing is a technology that converts natural language into data that a computer can understand.

With the development of information technology, attempts have been made to apply this to various fields of natural language processing or natural language understanding by expressing the meaning of words through vector space using the word meaning embedding technology.

However, the existing word semantic embedding technique has a limitation in that it is possible to embed only words that exist in the document wordbook, and it is impossible to acquire vector information for words that do not exist in the wordbook, such as new words and typos. In particular, in Korean, there are a lot of typographical errors due to confusion of foreign language notation, or typographical errors due to difficulty in distinguishing overlays, and there is a limitation that existing word meaning embedding cannot be solved.

The technical problem of the present invention relates to a word correction system that performs word semantic embedding on typos.

A word correction system according to an aspect of the technical idea of the present invention for achieving the above technical problem includes: a user interface that receives a user input word from a user through a network; By separating the learning word and the user input word by morphemes and separating each separated morpheme by element, generating processing data for analysis, and separating the processing data for analysis into engrams (n-grams) to separate the engram list. A processing data generator for generating; A word embedding learning unit that performs word embedding learning on the engram list through a skip-gram model; And a semantic vector generator for generating a semantic vector from the engram list.

The processing data for analysis may include a symbol separating the start and end of a morpheme, and a syllable delimiter.

The processing data generation unit performs a morpheme analysis on the learning word and the user input word to separate morphemes by morphemes, a morpheme separation unit to separate each morpheme separated from the morpheme separation unit, and the grapheme separation An engram separation unit that separates into consonants and vowels separated from the unit may be included.

The engram separating unit may generate vowel removal processing data after removing vowels of syllables from the processing data for analysis, and then generate engrams by separating the vowel removal processing data.

The engram separator may generate separated engrams of 2 characters to a number of characters larger than 2 and smaller than x for x characters constituting each of the vowel removal processing data.

The number of the vowel removal processing data may be the same as the number of syllables in the morpheme used to generate the processing data for analysis.

The engram separating unit generates vowel removal processing data after removing vowels of syllables from the processing data for analysis, and final removal processing data after removing the finality of each syllable, and then removes the vowel removal processing data, and the Separation processing data can be separated to generate engrams.

An adjacent word list output unit comparing the semantic vector obtained from the learning word with the semantic vector obtained from the user input word, and selecting adjacent words for the user input word to provide a neighbor word list; A similarity calculation unit calculating a similarity between the semantic vector obtained from the user input word and the semantic vector of the selected adjacent words, and selecting a corrected word for the user input word; It may further include a correction word output unit that generates the selected corrected word in a natural language and provides it to the user through the user interface.

The semantic vector generator may generate the semantic vector by summing the vectors of all the engrams in the engram list and calculating the average.

The word correction system according to the present invention learns through word embedding learning using a list of words to be learned from various resources, a converted corpus, and word semantic data, and compares similarities between adjacent words, so that the original non-learning patterns such as typographical errors are It is possible to obtain vector information similar to a word.

In addition, since word embedding learning using semantic information of real morphemes is performed in a corpus, not conventional word embedding, word embedding learning is possible with less learning data and learning time may be reduced.

In addition, by learning based on the meaning of the word to be learned, rather than simply appearing word positional information, it is efficient for low-frequency word learning, and because it targets the angram of the word that appeared in the dictionary, it uses more words than conventional word embedding. It can be expressed as a vector.

In addition, a word vector is generated by processing words adjacent to a real morpheme and word semantic data (for example, a high-level word or a synonym) of a word adjacent to a real morpheme, rather than a conventional location-based word embedding, to generate a word vector, thereby making the relationship between words cosine similarity. Can be seen through.

In particular, the word correction system according to the present invention does not perform word embedding learning through a skip-gram model for words separated into morphemes, and uses a word list through a skip-gram model using an engram list, which is a list of separated engrams. Since embedding learning is performed, even in the case of the unlearned word misspeller, since the semantic vector can acquire the original word through the closest word, the misspeller can be corrected.

1 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention.
Fig. 2 is a block diagram illustrating an operation of a processing data generation unit of a word correction system according to an exemplary embodiment of the present invention.
Figure 3 is a flow chart for explaining the operation of the morpheme separation unit and the character separation unit of the word correction system according to an exemplary embodiment of the present invention.
Figure 4 is a flow chart for explaining the operation of the engram separation unit of the word correction system according to an exemplary embodiment of the present invention.
Fig. 5 is a block diagram illustrating an operation of a processing data generating unit of a word correction system according to an exemplary embodiment of the present invention.
Fig. 6 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention.
FIG. 7 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The embodiments of the present invention are provided to more fully describe the present invention to those skilled in the art. The present invention can be applied to various changes and may have various forms, and specific embodiments will be illustrated in the drawings and described in detail. However, this is not intended to limit the present invention to a specific disclosure form, and it should be understood that it includes all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. In describing each drawing, similar reference numerals are used for similar components. In the accompanying drawings, the dimensions of the structures are enlarged or reduced than actual ones for clarity of the present invention.

Terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, terms such as “include” or “have” are intended to indicate the presence of features, numbers, steps, actions, elements, parts or combinations thereof described in the specification, but one or more other features. It should be understood that the existence or addition possibilities of fields or numbers, steps, actions, components, parts or combinations thereof are not excluded in advance.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms, such as those defined in the commonly used dictionary, should be interpreted as having meanings consistent with meanings in the context of related technologies, and are not to be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. .

In the following drawings and descriptions, one block, for example, a component represented or described as '~ unit' or '~ module' may be a hardware block or a software block. For example, each of the components may be an independent hardware block that communicates with each other, or may be a software block executed in one processor.

In order to fully understand the configuration and effects of the present invention, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

1 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention.

Referring to FIG. 1, the word correction system 1 is a processing data generation unit 100 that receives a learning word from the processing data storage unit 1000 and generates an n-gram list 140, processing data It includes a word embedding learning unit 200 for performing word embedding learning based on the engram list 140 generated by the generator 100, and a semantic vector generator 300. The word correction system 1 may further include a learning word semantic vector storage 350 that stores the learning word semantic vector generated by the semantic vector generator 300.

The processing data storage unit 1000 may have a learning word that is processing data for generating the engram list 140 in the processing data generation unit 100. The processing data generation unit 100 may be, for example, a space capable of storing data in any form, such as NoSQL, relational database, and file system. The processing data generating unit 100 may be a storage unit logically divided, a division unit logically dividing one or a plurality of storage units, a storage unit physically divided, or a division unit logically divided. It can be a part. In some embodiments, the processing data storage unit 1000 may be a space that can store learning words included in the word correction system 1.

In some other embodiments, the processing data storage unit 1000 may be a space or a system that stores a learning word outside the word correction system 1, and in this case, the word correction system 1 may process data through a network. It may be connected to the storage unit 1000. The processing data storage unit 1000 is, for example, a vocabulary, which is a language material that collects text in a computer-readable form for language research, or a vocabulary that is a knowledge base that shows a large vocabulary as an association or semantic relationship between words. It can be a semantic network. For example, the processing data storage unit 1000 may be a single corpus, or a vocabulary semantic network, but is not limited thereto, and may be a plurality of corpuses separately constructed or a vocabulary semantic network.

The processing data generation unit 100 may include a morpheme separation unit 110, a grapheme separation unit 120, and an engram separation unit 130.

The morpheme separating unit 110 may perform morpheme analysis on a learning word received from the processing data storage unit 1000 to separate morphemes.

The grapheme separation unit 120 may separate each morpheme separated from the morpheme separation unit 110 for each grapheme. In some embodiments, the grapheme separation unit 120 may separate each morpheme into consonants and vowels. In some other embodiments, the grapheme separating part 120 may be divided into consonants and vowels to divide each morpheme into super, neutral, and vertical (base).

In some embodiments, the phoneme separator 120 may insert syllable separators between the separated phonemes.

The engram separation unit 130 separates each morpheme into an n-gram using consonants and vowels separated from the grapheme separation unit 120. The engram separation unit 130 may perform a removal process for each vowel included in each morpheme and / or each species before separation into engrams. In this case, instead of the removed vowel and / or species, Can be given a removal symbol.

For example, when one morpheme consists of x characters, the engram of the morpheme may consist of 2 to x-1 characters. At this time, the x characters constituting the morpheme, in addition to the consonants and vowels, the delimiter given by the morpheme separating part 110 and / or the semetic part separating part 120, and / or the removal given by the engram separating part 130 Symbols and the like may be further included.

The list of engrams separated from the engram separator 130 may be generated as the engram list 140.

The word embedding learning unit 200 may perform word embedding learning through a skip-gram model using the processing data received from the processing data generation unit 100, for example, an engram list.

For example, the word embedding learning unit 200 places words to be learned in the engram list, which is a list of engrams for morphemes, in the input layer of the skip-gram in word embedding, and words to be learned through word embedding learning. You can learn to generate word vectors. The word embedding learning unit 200 performs word embedding learning through a feed forward process and a back propagation process of word embedding learning, and processing data of words to be learned in the back propagation process. The weight value associated with the word to be learned can be changed without changing the connected weight value.

The semantic vector generation unit 300 may generate a semantic vector of words learned through the word embedding learning unit 200. In some embodiments, the semantic vector generator 300 may generate the n words having the highest vector similarity in descending order by using the learned word vector information, thereby providing a correction word for the misspeller.

The learning word semantic vector storage 350 may store the semantic vector of the learned word. The semantic vector of the learned word stored in the learning word semantic vector storage 350 may be used to provide a correction word for a misspelled word.

The learning word meaning vector storage 350 may be, for example, a space for storing data in any form, such as NoSQL, a relational database, and a file system. The learning word meaning vector storage 350 is a storage unit logically divided, a division unit that logically divides one or a plurality of storage units, a storage unit that is physically divided, or a division unit that is logically divided. It can be part of.

The word correction system 1 according to the present invention learns through word embedding learning using a list of words to be learned from various resources (eg, corpus, vocabulary semantic network), transformed corpus, and word semantic data and similarity between adjacent words By comparison, vector information similar to the original word can be obtained for an unlearned pattern such as a typo.

In addition, since word embedding learning using semantic information of real morphemes is performed in a corpus, not conventional word embedding, word embedding learning is possible with less learning data and learning time may be reduced.

In addition, by learning based on the meaning of the word to be learned rather than simply appearing word positional information, it is also effective for low-frequency word learning, and because it targets angram of words that appear in the dictionary, vector more words than conventional word embedding Can be expressed as

In addition, a word vector is generated by processing words adjacent to a real morpheme and word semantic data (for example, a high-level word or a synonym) of a word adjacent to a real morpheme, rather than a conventional location-based word embedding, to generate a word vector, thereby making the relationship between words cosine similarity. Can be seen through.

In particular, the word correction system 1 according to the present invention does not perform word embedding learning through a skip-gram model on words separated into morphemes, and embeds words through a skip-gram model using a list of separated engrams. Since learning is performed, even in the case of a misspelled word, the original word can be acquired through the word having the closest semantic vector, so that the misspelled word can be corrected.

2 is a block diagram for explaining the operation of the processing data generation unit of the word correction system according to the exemplary embodiment of the present invention, and FIG. 3 is a morpheme separation unit and grapheme of the word correction system according to the exemplary embodiment of the present invention. 4 is a flowchart for explaining the operation of the separation unit, and FIG. 4 is a flowchart for explaining the operation of the engram separation unit of the word correction system according to the exemplary embodiment of the present invention.

Referring to FIGS. 2 to 4 together, the processing data generation unit 100 includes a morpheme separation unit 110, a grapheme separation unit 120, and an engram separation unit 130. The list of engrams separated from the engram separator 130 may be generated as the engram list 140.

The morpheme separating unit 110 may perform morpheme analysis on a learning word provided from the processing data storage unit (1000 in FIG. 1), and may separate the morphemes. For example, when the processing data storage unit 1000 has language data such as 'orange is delicious' (S100), the processing data storage unit 1000 morphologically separates the learning phrases 'orange' and 'delicious'. It may be provided to the unit 110 (S200).

The morpheme separation unit 110 may include a morpheme separation module 112 and a morpheme separator insertion module 114. The morpheme separating module 112 may separate morphemes for the phrases 'orange' and 'delicious' to obtain separated morphemes such as 'orange', 'silver', 'delicious', and 'da' (S300) ). The morpheme separator insertion module 114 may insert a symbol that separates the start and end of a morpheme.

The grapheme separation unit 120 may include a grapheme separation module 122 and a syllable separator insertion module 124. The phoneme separation module 122 may separate each morpheme into consonants and vowels. For example, the grapheme separation module 122 has 19 morphemes of each Korean morpheme (ㄱ, ㄴ, ㄷ, ㄹ, ㅁ, ㅂ, ㅅ, ㅇ, ㅈ, ㅊ, ㅋ, ㅌ, ㅍ, ㅎ, ㄲ , ㄸ, ㅃ, ㅆ, ㅉ) and 21 vowels (ㅏ, ㅑ, ㅓ, ㅕ, ㅗ, ㅛ, ㅜ, ㅠ, ㅡ, ㅣ, ㅐ, ㅒ, ㅔ, ㅖ, ㅘ, ㅙ, ㅚ, ㅝ, ㅞ, ㅟ, ㅢ) can be decomposed to a total of 40 characters.

The grapheme separation module 122 first separates syllables into 'o', 'ren', and 'ji' with respect to the morpheme 'orange' (S400), then 'ㅇ', 'ㅗ', 'd', 'ㅔ Separation can be performed with ',' ㄴ ',' ㅈ ',' ㅣ '(S500). The syllable separator insertion module 124 may insert a syllable separator between each syllable.

'를 삽입할 수 있다. 따라서, 형태소 분리부(110)와 자소 분리부(120)를 통하여, '오렌지'라는 형태소는 '<ㅇㅗ

ㄹㅔㄴ

ㅈㅣ>'와 같은 분석용 가공 데이터로 가공될 수 있다(S600). 자소가 분리되고, 형태소 구분 기호와 음절 구분 기호가 삽입되도록 가공된 형태소를 분석용 가공 데이터라 호칭할 수 있다. For example, in the morpheme separator insertion module 114, the symbols '<' and '>' that separate the start and end of the morpheme are inserted at the start and end of the morpheme, and the syllable separator insertion module 124 Between each syllable is the syllable separator '

'Can be inserted. Therefore, through the morpheme separating part 110 and the morpheme separating part 120, the morpheme called 'orange'is'<ㅇ

ㄹ ㅔ ㄴ

It can be processed into processing data for analysis such as ㅈ ㅣ>'(S600). The morpheme processed so that the grapheme is separated and the morpheme separator and syllable separator are inserted may be referred to as processing data for analysis.

In FIG. 3, the morpheme separator and the syllable separator are illustrated as being inserted in the final step (S600), but this is for illustrating mainly the processes of morpheme, syllable, and phoneme separation, but is not limited thereto. For example, the morpheme separator may be inserted in step S300 in which the morpheme is separated, and the syllable separator may be inserted in step S400 in which the syllable is separated.

The engram separation unit 130 separates into engrams using processing data for analysis of each morpheme. The engram separating unit 130, before separating into engrams, the vowel removing module 132 that performs the removal process for each vowel included in each morpheme, and the engram list generating module 136 that separates into engrams ).

ㄹㅔㄴ

ㅈㅣ>'라는 분석용 가공 데이터가 제공되면(S600), 모음 제거 모듈(132)은 음절 각각의 모음을 제거하여, '<ㅇ

ㄹㅔㄴ

ㅈㅣ>', '<ㅇㅗ

ㄹㄴ

ㅈㅣ>', '<ㅇㅗ

ㄹㅔㄴ

ㅈ>'와 같은 모음 제거 가공 데이터들을 생성할 수 있다(S700). 모음 제거 가공 데이터들의 개수는 분석용 가공 데이터를 생성하는 데에 사용된 형태소가 가지는 음절수와 동일할 수 있다. About the morpheme 'orange', '<ㅇㅗ

ㄹ ㅔ ㄴ

When the processing data for analysis called ㅈ ㅣ>'is provided (S600), the vowel removal module 132 removes each vowel of syllables, and the'<ㅇ

ㄹ ㅔ ㄴ

ㅈ ㅣ>','<ㅇㅗ

ㄹㄴ

ㅈ ㅣ>','<ㅇㅗ

ㄹ ㅔ ㄴ

Collection data such as ㅈ>'may be generated (S700). The number of vowel removal processing data may be the same as the number of syllables in the morpheme used to generate the processing data for analysis.

ㄹㅔㄴ

ㅈㅣ>', '<ㅇㅗ

ㄹㄴ

ㅈㅣ>', '<ㅇㅗ

ㄹㅔㄴ

ㅈ>'와 같이 모음 제거 가공 데이터들이 10개의 문자로 이루어지는 경우, 엔그램 리스트 생성 모듈(136)은 2개 내지 9개의 문자로 이루어지는 분리된 엔그램을 생성할 수 있다. The engram list generation module 136 may separate each of the vowel removal processing data into an engram. For example, when the vowel removal processing data for one morpheme is composed of x characters, the engram list generation module 136 generates a separate engram composed of 2 to x-1 characters for the morpheme. can do. That is, the engram range may be 2 to x-1. '<ㅇ

ㄹ ㅔ ㄴ

ㅈ ㅣ>','<ㅇㅗ

ㄹㄴ

ㅈ ㅣ>','<ㅇㅗ

ㄹ ㅔ ㄴ

When the vowel removal processing data such as ㅈ>'is composed of 10 characters, the engram list generation module 136 may generate a separated engram composed of 2 to 9 characters.

ㄹㅔㄴ

ㅈㅣ>'에 대해서 3개로 이루어지는 엔그램을 분리하여, '<ㅇ

', 'ㅇ

ㄹ', '

ㄹㅔ', 'ㄹㅔㄴ', 'ㅔㄴ

', 'ㄴ

ㅈ', '

ㅈㅣ', 'ㅈㅣ>'와 같은 엔그램을 분리할 수 있고, '<ㅇㅗ

ㄹㄴ

ㅈㅣ>'에 대해서 3개로 이루어지는 엔그램을 분리하여, '<ㅇㅗ', 'ㅇㅗ

', 'ㅗ

ㄹ', '

ㄹㄴ', 'ㄹㄴ

', 'ㄴ

ㅈ', '

ㅈㅣ', 'ㅈㅣ>'와 같은 엔그램을 분리할 수 있고, , '<ㅇㅗ

ㄹㅔㄴ

ㅈ>'에 대하여 3개로 이루어지는 엔그램을 분리하여, '<ㅇㅗ', 'ㅇㅗ

', 'ㅗ

ㄹ', '

ㄹㅔ', 'ㄹㅔㄴ', 'ㅔㄴ

', 'ㄴ

ㅈ', '

ㅈ>'와 같은 엔그램을 분리할 수 있다(S800). Specifically, '<ㅇ

ㄹ ㅔ ㄴ

For ㅈ ㅣ>', separate the three-gram engram,'<ㅇ

',' ㅇ

L', '

ㄹ ㅔ ',' ㄹ ㅔ ㄴ ',' ㅔ ㄴ

', 'N

ㅈ ','

Engrams like ㅈ ㅣ ',' ㅈ ㅣ>'can be separated, and'<ㅇㅗ

ㄹㄴ

For ㅈ ㅣ>', separate three engrams,'<ㅇ<',' ㅇㅗ

',' ㅗ

L', '

ㄹㄴ ',' ㄹㄴ

', 'N

ㅈ ','

Engrams such as ㅈ ㅣ ',' ㅈ ㅣ>'can be separated, and'<ㅇㅗ

ㄹ ㅔ ㄴ

For ㅈ>', separate the three-gram engrams,'<ㅇㅗ',' ㅇㅗ

',' ㅗ

L', '

ㄹ ㅔ ',' ㄹ ㅔ ㄴ ',' ㅔ ㄴ

', 'N

ㅈ ','

Engrams such as ㅈ>'can be separated (S800).

In some embodiments, the engram list generation module 136 may generate a number of characters larger than 2 to 2 and smaller than x-1, for example, 2 to 6 separated engrams. That is, the engram range may be a number less than 2 to x-1, for example, 2 to 6.

', 'ㅇ

ㄹ', '

ㄹㅔ', 'ㄹㅔㄴ', 'ㅔㄴ

', 'ㄴ

ㅈ', '

ㅈㅣ', 'ㅈㅣ>', '<ㅇㅗ', 'ㅇㅗ

', 'ㅗ

ㄹ', '

ㄹㄴ', 'ㄹㄴ

', '

ㅈ>'로 생성될 수 있다(S900). 도 4에는 예시적으로, 3개로 이루어지는 분리된 엔그램을 도시하였으며, 이에 대해여 설명하였으나, 생성된 엔그램 리스트(140)는 2개, 및 4개 이상의 분리된 엔그램을 더 포함할 수 있다. As such, a list of separated engrams generated in the engram list generation module 136 may be generated as an engram list 140. For example, the engram list 140 removes overlapping engrams among the separated engrams.

',' ㅇ

L', '

ㄹ ㅔ ',' ㄹ ㅔ ㄴ ',' ㅔ ㄴ

', 'N

ㅈ ','

ㅈ ㅣ ',' ㅈ ㅣ>','<ㅇㅗ',' ㅇㅗ

',' ㅗ

L', '

ㄹㄴ ',' ㄹㄴ

','

It can be created as ㅈ>'(S900). For example, in FIG. 4, three separate engrams are illustrated, which have been described, but the generated engram list 140 may further include two and four or more separated engrams. .

Fig. 5 is a block diagram illustrating an operation of a processing data generating unit of a word correction system according to an exemplary embodiment of the present invention.

Referring to FIG. 5, the processing data generation unit 100 includes a morpheme separation unit 110, a grapheme separation unit 120, and an engram separation unit 130. The list of engrams separated from the engram separator 130 may be generated as the engram list 140.

')를 삽입할 수 있다. 따라서, 자소 분리부(120)는 '오렌지'라는 형태소에 대하여 '<ㅇㅗ

ㄹㅔㄴ

ㅈㅣ

>'라는 분석용 가공 데이터를 제공할 수 있다. 다른 일부 실시 예에서, 자소 분리 모듈(120)은 종성 미포함 기호를 삽입하지 않을 수 있고, 이 경우, 음절 구분 기호인 '

' 앞에 모음이 오는 경우, 종성이 포함되지 않는 것으로 판단할 수 있다. The grapheme separation unit 120 may include a grapheme separation module 122 and a syllable separator insertion module 124. Elemental separation module 122, in some embodiments, the elemental separation module 122 may be separated into consonants and vowels to divide each morpheme into super, neutral, and longitudinal (support). For example, the grapheme separation module 120 has 19 consonants for each morpheme (ㄱ, ㄴ, ㄷ, ㄹ, ㅁ, ㅂ, ㅅ, ㅇ, ㅈ, ㅊ, ㅋ, ㅌ, ㅍ, ㅎ, ㄲ, ㄸ, ㅃ, ㅆ, ㅉ), 21 neutral vowels (ㅏ, ㅑ, ㅓ, ㅕ, ㅗ, ㅛ, ㅜ, ㅠ, ㅡ, ㅣ, ㅐ, ㅒ, ㅔ, ㅖ, ㅘ, ㅙ, ㅚ, ㅝ , ㅞ, ㅟ, ㅢ), and 27 bases (ㄱ, ㄲ, ㄳ, ㄴ, ㄵ, ㄶ, ㄷ, ㄹ, ㄺ, ㄻ, ㄼ, ㄽ, ㄾ, ㄿ, ㅀ, ㅁ, ㅂ, ㅄ, ㅅ , ㅆ, ㅇ, ㅈ, ㅊ, ㅋ, ㅌ, ㅍ, ㅎ) You can perform the process of decomposing a total of 67 characters. In some embodiments, the grapheme separation module 120 does not have a seed (base), and does not have a seed (eg, '

') Can be inserted. Therefore, the grapheme separation unit 120 is referred to as'<ㅇㅗ

ㄹ ㅔ ㄴ

ㅈ ㅣ

>'Can provide processing data for analysis. In some other embodiments, the grapheme separation module 120 may not insert a symbol without a species, in this case, a syllable separator '

'If a vowel comes before, it can be judged that the species is not included.

The engram separation unit 130 separates into engrams using processing data for analysis of each morpheme. The engram separation unit 130 may include a vowel removal module 132, a finality removal module 134, and an engram list generation module 136 that separates into engrams.

ㄹㅔㄴ

ㅈㅣ

>''라는 분석용 가공 데이터가 제공되면, 모음 제거 모듈(132)은 음절 각각의 모음을 제거하여, '<ㅇ

ㄹㅔㄴ

ㅈㅣ

>', '<ㅇㅗ

ㄹㄴ

ㅈㅣ

>', '<ㅇㅗ

ㄹㅔㄴ

ㅈ

>'와 같은 모음 제거 가공 데이터들을 생성할 수 있고, 종성 제거 모듈(134)은 음절 각각의 종성을 제거하여, '<ㅇㅗ

ㄹㅔㄴ

ㅈㅣ

>', '<ㅇㅗ

ㄹㅔ

ㅈㅣ

>', '<ㅇㅗ

ㄹㅔㄴ

ㅈㅣ>'와 같은 종성 제거 가공 데이터들을 생성할 수 있다. 종성 제거 가공 데이터들의 개수는 분석용 가공 데이터를 생성하는 데에 사용된 형태소가 가지는 음절수와 동일할 수 있다. About the morpheme 'orange', '<ㅇㅗ

ㄹ ㅔ ㄴ

ㅈ ㅣ

>'' When the processing data for analysis is provided, the vowel removing module 132 removes each vowel of syllables, so that '<ㅇ

ㄹ ㅔ ㄴ

ㅈ ㅣ

>','<ㅇㅗ

ㄹㄴ

ㅈ ㅣ

>','<ㅇㅗ

ㄹ ㅔ ㄴ

ㅈ

>'Can be generated vowel removal processing data, and the deconciliation removing module 134 removes the disposition of each syllable, so that'<ㅇㅗ

ㄹ ㅔ ㄴ

ㅈ ㅣ

>','<ㅇㅗ

ㄹ ㅔ

ㅈ ㅣ

>','<ㅇㅗ

ㄹ ㅔ ㄴ

It is possible to generate de-scaling data such as ㅈ ㅣ>'. The number of species removal processing data may be the same as the number of syllables in the morpheme used to generate the processing data for analysis.

The engram list generation module 136 may separate each of the vowel removal processing data and the seed removal processing data into engrams, and the list of the separated engrams generated in the engram list generation module 136 may be yen. It can be generated as a gram list 140.

Fig. 6 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention.

Referring to FIG. 6, the word correction system 2 includes a user interface (UI, 20) that receives words from the user 10 through the network 50 and provides corrected words.

The network 50 includes wired Internet service, local area network (LAN), broadband communication network (WAN), intranet, wireless Internet service, mobile computing service, wireless data communication service, wireless Internet access service, satellite communication service, wireless LAN, Bluetooth, etc. It can include anything that can send and receive data over wired or wireless. When the network 50 is connected to a smartphone or tablet, the network 50 may be wireless data communication services such as 3G, 4G, 5G, wireless LAN such as Wi-Fi, Bluetooth, or the like.

The user interface 50 may provide an interface for accessing the word correction system 2 through a terminal or the like used by the user 10. The user 10 may input and transmit words to the word correction system 2 through the user interface 20, and receive corrected words provided by the word correction system 2 through the user interface 20. have.

The word correction system 2 includes a processing data generation unit 100, a word embedding learning unit 200, a semantic vector generation unit 300, and a learning word semantic vector storage 350. The processing data generation unit 100, the word embedding learning unit 200, and the semantic vector generation unit 300 of the word correction system 2 include the processing data generation unit 100 and words described with reference to FIGS. 1 to 5. It is substantially the same as the embedding learning unit 200 and the semantic vector generator 300, and when the word entered by the user 10 through the user interface 50 is a misspelling that requires correction, the misspelled word (user Input words) are performed through the processes described with reference to FIGS. 1 to 5. The learning word semantic vector storage 350 may store a semantic vector of learning words for a correct morpheme (a learning word, or a learning word) rather than a typo described through FIGS. 1 to 5.

The word embedding learning unit 200 and the semantic vector generator 300 may obtain engram vectors by performing engram embedding on the engram list 140. The semantic vector of a word can be obtained by summing all the engram vectors in the engram list and calculating the average, which is expressed as follows.

Here, Vw represents the semantic vector value of the input word w, i represents the index of the engram list, k represents the total number of engram lists, and Vi represents the vector of the i th engram.

The word correction system 2 further includes an adjacent word list output unit 400, a similarity calculation unit 500, and a correction word output unit 600.

When the semantic vector generator 300 generates a semantic vector for the user input word, the adjacent word list output unit 400 refers to the semantic vector of the learned word stored in the learning word semantic vector store 350, and inputs the user. A list of adjacent words for a word can be selected. For example, the adjacent word list output unit 400 compares the engram list of the learned word with the engram list of the user input word, and the ratio of the engram overlapping the engram list of the user input word with a constant value. The learned words having the above engram list may be selected as adjacent words, and the list may be provided. For example, the adjacent word list output unit 400 compares the similarity between the engram vector of the learned word and the engram vector of the user input word, and has an engram vector having a high similarity with the engram vector of the user input word. The learned words may be selected as adjacent words, and a list may be provided.

The similarity calculator 500 may calculate a similarity between the semantic vector of the user input word and the semantic vector of the selected adjacent word, and select the corrected word for the misspeller. For example, the similarity calculator 500 may select words with adjacent vector values as corrected words by comparing the cosine similarity between the semantic vector of the user input word and the semantic vectors of the selected adjacent words. In some embodiments, the similarity calculator 500 may select a plurality of words having high similarity of the semantic vector in descending order.

The correction word output unit 600 may generate the corrected word selected through the user interface 20 as a natural language and provide it to the user 10. In some embodiments, the correction word output unit 600 may generate words having high similarity of the semantic vector in natural language and provide the user 10 with a plurality in descending order.

FIG. 7 is a schematic block diagram of a word correction system according to an exemplary embodiment of the present invention. For example, the word correction system 1 shown in FIG. 1 is a system that obtains a semantic vector for a word that is a correct morpheme, not a typographical error, and stores it in the learning word semantic vector storage 350 to correct the word. The word correction system 2 shown in FIG. 7 is a system for obtaining a corrected word for a misspeller requiring word correction by using the learning word semantic vector storage 350 in which vector information about the correct morphological word is already stored. The visible word correction system 1 obtains the semantic vector for the correct morphological word and stores it in the learning word semantic vector storage 350, and corrects for the misspeller requiring word correction using the learning word semantic vector storage 350 It may be a system in which finding a word that has been made is done together. Therefore, contents overlapping with those described in FIGS. 1 to 6 may be omitted.

Referring to FIG. 7, the correction word providing system 3 includes a user interface 20, a processing data generating unit 100, a word embedding learning unit 200, a semantic vector generating unit 300, and a learning word semantic vector storage ( 350), an adjacent word list output unit 400, a similarity calculation unit 500, and a correction word output unit 600.

The word correction system 3 receives a learning word from the processing data storage unit 1000, generates an engram list 140 for the learning word in the processing data generation unit 100, and a word embedding learning unit 200 After performing the word embedding learning in, the semantic vector generator 300 generates a learning word semantic vector and stores it in the learning word semantic vector storage 350.

In addition, the word correction system 3 receives a user input word that is a misspelling that needs to be corrected from the user 10 through the network 50, and the engram list 140 for the user input word in the processing data generation unit 100 After generating the word embedding learning in the word embedding learning unit 200, the semantic vector generating unit 300 generates a semantic vector of the user input word.

The adjacent word list output unit 400 compares the engram list of the learning word and the user input word stored in the learning word semantic vector storage 350, or compares the similarity of the engram vector to compare adjacent words to the user input word. It can be selected and provided.

Subsequently, the similarity calculating unit 500 may calculate the similarity between the semantic vector of the user input word and the semantic vector of the selected adjacent word, to select a corrected word for the user input word that is a misspelling, and the correction word output unit ( The 600 may generate the corrected word selected through the user interface 20 as a natural language and provide it to the user 10.

As described above, the present invention has been described in detail with reference to preferred embodiments, but the present invention is not limited to the above embodiments, and various modifications and alterations by those skilled in the art within the technical spirit and scope of the present invention This is possible.

1, 2, 3: word correction system, 10: user, 20: user interface, 50: network, 100: processing data generation unit, 200: word embedding learning unit, 300: semantic vector generator, 350: learning word semantic vector Repository, 400: adjacent word list output unit, 500: similarity calculation unit, 600: correction word output unit, 1000: processing data storage unit

Claims (9)

A user interface that receives a user input word from a user through a network;
By separating the learning word and the user input word by morphemes and separating each separated morpheme by element, generating processing data for analysis and separating the processing data for analysis into engrams (n-grams) to separate the engram list. A processing data generator for generating;
A word embedding learning unit that performs word embedding learning on the engram list through a skip-gram model; And
And a semantic vector generator for generating a semantic vector from the engram list. According to claim 1,
The processing data for analysis, the word correction system, characterized in that it includes a symbol for distinguishing the morpheme time and end, and a syllable separator. According to claim 2,
The processing data generation unit,
A morpheme separation unit that separates each morpheme separated from the morpheme separation unit by morphemes by performing a morpheme analysis on the learning word and the user input word, and a consonant separated from the morpheme separation unit. A word correction system comprising an engram separator that separates into engrams using vowels. According to claim 3,
The engram separating unit, after generating vowel removal processing data with each syllable removed from the analysis processing data, separating the vowel removal processing data to generate engrams. According to claim 4,
The engram separating unit, for the x characters constituting each of the vowel removal processing data, a word correction system characterized in that for generating the separated engrams consisting of a number of characters greater than 2 and less than x 2 characters. . According to claim 4,
The number correction processing number of the word correction system, characterized in that the same as the number of syllables in the morpheme used to generate the processing data for analysis. According to claim 3,
The engram separating unit generates vowel removal processing data after removing vowels of syllables from the processing data for analysis, and vocal removal processing data after removing the finality of each syllable, and then removes the vowel removal processing data, and the A word correction system characterized by separating de-scaling processing data to generate engrams. According to claim 1,
An adjacent word list output unit comparing the semantic vector obtained from the learning word with the semantic vector obtained from the user input word, and selecting adjacent words for the user input word to provide a neighbor word list;
A similarity calculation unit calculating a similarity between the semantic vector obtained from the user input word and the semantic vector of the selected adjacent words, and selecting a corrected word for the user input word;
And a correction word output unit that generates the selected corrected word in natural language and provides it to the user through the user interface. According to claim 1,
The semantic vector generation unit, a word correction system, characterized in that by summing the vectors of all the engrams in the engram list, and then obtaining the average, generating the semantic vector.

Images