KR102564470B1 - System for synonym extraction through semantic analysis of unregistered words based-on deep learning model and method thereof - Google Patents

Abstract

The present invention relates to a system for extracting synonyms through semantic analysis of unregistered words based on a deep learning model and a method for extracting synonyms. In a given document, a synonym is not recognized through morpheme analysis and is not properly separated, or is recognized through morpheme analysis but not registered in advance. Synonyms through semantic analysis of unregistered words based on a deep learning model that extracts unregistered words, immediately grasps the meaning of the extracted unregistered words through a deep learning model, and extracts and adds related synonyms so that they can be used for sentence analysis and search. It relates to an extraction system and its method.

Description

Synonym extraction system and method through semantic analysis of unregistered words based on deep learning model

The present invention relates to a synonym extraction system and method through semantic analysis of unregistered words based on a deep learning model, and more particularly, to a synonym extraction system and method for a given document that is not properly separated because it is not recognized through morpheme analysis in a given document, or is recognized through morpheme analysis but is not recognized through morpheme analysis. Meaning of unregistered words based on a deep learning model that extracts unregistered words that are not registered, immediately grasps the meaning of the extracted unregistered words through a deep learning model, and extracts and adds related synonyms to be used for sentence analysis and search. It relates to a synonym extraction system and method through analysis.

As we move into the digital age, many of the analog parts are being converted to digital. This epochal transformation is changing the environment of households, businesses and public institutions.

In particular, changes are most extreme in the office, where all documents in the office are digitized according to a paperless policy, and accordingly, internal payment and management are performed through digital documents.

At this time, in order to effectively manage and search digital documents, it is important to construct an effective database. To this end, there is a need for analysis models that automatically analyze documents, extract keywords, and extract summaries.

However, in the past, there was a problem in that the predictability of the analysis models was reduced because words that were not registered in the dictionary could not be managed and organized.

Therefore, in the present invention, unregistered words that are not properly separated because they are not recognized through morphological analysis of a given document, or words recognized through morphological analysis but not registered in the dictionary are extracted, and the meaning of the extracted unregistered words is immediately grasped to obtain related synonyms. By extracting and adding, we propose a method that can be easily used for sentence analysis and search.

Next, the prior inventions existing in the technical field of the present invention will be briefly described, and then the technical details to be achieved by the present invention to be differentiated from the prior inventions will be described.

First, Yonsei University Graduate School of Information Big Data master's thesis (Kim Han-jun, 2018.12) is an empirical study on the method of extracting new words and applying morpheme analysis based on domestic online community posts. It is a prior art related to a method of extracting a new word from text data through a discrimination model.

In addition, Korea Patent Publication No. 10-2022-0074572 (2022.06.03.) is not limited to pre-made morpheme dictionaries, but uses deep learning-based common pattern analysis of nouns such as noun positions in sentences, conjunctions, and endings to find new words. It is a prior invention related to a device capable of extracting and an operation method thereof.

In addition, Korean Patent Publication No. 10-2020-0141419 (2020.12.18.) analyzes the relationship between one or more words calculated through word embedding with respect to sentences collected from the community, and based on the analysis of the relationship, the target word and a synonym extraction method for determining a similar candidate word, calculating a similarity score for the target word and the similar candidate word, and determining a synonym relationship between the target word and the similar candidate word based on the similarity score. It is a preceding invention.

However, the present invention immediately grasps the meaning of unregistered words extracted from a given document, extracts and adds related synonyms, and utilizes them easily for sentence analysis and search. The dissertation, which extracts new words by removing words registered in , Korea Patent Publication No. 10-2022-0074572, which extracts new words by using noun positions in sentences, combined investigations, and endings, and calculation through word embedding Compared to Korean Patent Publication No. 10-2020-0141419, which determines a target word and a similar candidate word based on an association analysis of the words and extracts a synonym based on a similarity score, a significant structural difference there is

The present invention was created to solve the above problems, and a system and method for extracting non-registered words in a sentence, immediately grasping their meaning, and extracting and adding related synonyms so that they can be used for sentence analysis and search. is intended to provide

In addition, the present invention provides a system and method for adding unregistered words in a sentence to a user dictionary, generating a synonym group through word embedding, and then setting a representative word so that the words of the synonym group can be included in a search result for a different purpose

In addition, the present invention can be used for document search by removing unregistered words in a sentence, predicting words through a deep learning learning model, and then extracting synonyms by comparing similarities between sentences containing the predicted word and sentences containing unregistered words. It is another object to provide a system and method for enabling it.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problems described above, and other technical problems may exist.

A synonym extraction system through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention includes: a data collection unit for collecting documents; an unregistered word extraction unit extracting unregistered words from the collected documents; And extracting synonyms to replace the unregistered words by removing the extracted non-registered words, predicting words to be included in the part from which the unregistered words are removed, and comparing similarities between sentences containing the predicted words and sentences containing the unregistered words It includes a synonym extractor; and is characterized in that the synonym extracted according to the semantic analysis of the non-registered word is supported for use in sentence analysis and search.

In addition, the synonym extraction unit may include: an unregistered word removal unit that masks the extracted unregistered word; a word prediction unit that predicts at least one word to be included in the masked portion; a sentence similarity comparing unit for comparing a similarity between a sentence containing the at least one predicted word and a sentence containing the unregistered word; and an unregistered word synonym extraction unit for selecting and extracting any one word among words equal to or greater than a preset threshold as a result of the comparison as a synonym to replace the unregistered word.

At this time, the prediction is characterized by inputting the masked sentence to a pre-generated deep learning learning model for word prediction including BERT (Bidirectional Encoder Representations from Transformers), and predicting words to enter the masked part. do.

In addition, the system further includes an appearance frequency checking unit that determines whether the frequency of appearance of the extracted non-registered word is equal to or greater than a preset number of times, and as a result of checking by the frequency of appearance checking unit, the frequency of appearance of the extracted non-registered word is set in advance. If it is less than the number of times, it is characterized in that a synonym to replace the non-registered word is extracted through the synonym extraction unit.

In addition, the system may include: a morpheme analyzer user dictionary registration unit for registering the extracted unregistered word in the morpheme analyzer user dictionary when the appearance frequency of the extracted unregistered word is equal to or greater than a preset number of times as a result of checking by the appearance frequency checking unit; a word embedding and synonym group generation unit generating a synonym group by executing word embedding on all of the extracted words including unregistered words; and a representative word setter configured to set a representative word in the generated synonym group and to update information on a connection between the set representative word and the synonym in the user dictionary of the morpheme analyzer.

In addition, the unregistered word extraction unit may include a morpheme analysis unit separating words included in the document through a morpheme analyzer; a word separation confirmation unit checking whether word separation is normally performed through the morpheme analyzer; and an unregistered word confirmation unit for confirming whether the word is an unregistered word that is not registered in the morpheme analyzer user dictionary when word separation is normally performed through the morpheme analyzer as a result of the confirmation, wherein the word separation confirmation unit checks As a result, if the morpheme analyzer does not recognize it and is not divided into words, or if the non-registered word confirmation unit recognizes the morpheme as a word but is not registered in the user dictionary of the morpheme analyzer, it is characterized in that it is extracted as an unregistered word.

In addition, the non-registered word extraction unit may extract unregistered words from the collected documents using at least one of an LR tokenizer, a cohesion score, and branch entropy.

In addition, a synonym extraction method through semantic analysis of unregistered words based on a deep learning model according to an embodiment of the present invention includes a data collection step of collecting documents in a synonym extraction system through semantic analysis of unregistered words; an unregistered word extraction step of extracting unregistered words from the collected documents; And extracting synonyms to replace the unregistered words by removing the extracted non-registered words, predicting words to be included in the part from which the unregistered words are removed, and comparing similarities between sentences containing the predicted words and sentences containing the unregistered words A synonym extraction step is included, and the synonym extracted according to the semantic analysis of the non-registered word is supported for use in sentence analysis and search.

In addition, the synonym extraction step may include, in the synonym extraction system through the semantic analysis of the unregistered word, an unregistered word removal step of masking the unregistered word extracted in the unregistered word extraction step; a word prediction step of predicting at least one word to be included in the masked portion; a sentence similarity comparison step of comparing a similarity between a sentence containing the at least one predicted word and a sentence containing the unregistered word; And as a result of the comparison, an unregistered word synonym extraction step of selecting and extracting any one of the words equal to or greater than a preset threshold value as a synonym to replace the unregistered word; wherein the prediction is performed using previously generated words including BERT It is characterized in that a word to be entered into the masked part is predicted by inputting the masked sentence to a prediction deep learning model.

In addition, the method further includes an appearance frequency check step of determining whether the frequency of occurrence of the non-registered word extracted in the non-registered word extraction step is equal to or greater than a predetermined number of times in the synonym extraction system through the semantic analysis of the non-registered word, As a result of checking in the occurrence frequency checking step, if the appearance frequency of the extracted unregistered word is less than a preset number of times, a synonym to replace the unregistered word is extracted through the synonym extraction step.

In addition, in the synonym extraction system through semantic analysis of the non-registered word, as a result of checking in the occurrence frequency checking step, if the frequency of appearance of the extracted non-registered word is equal to or greater than a predetermined number of times, the extracted non-registered word is converted into a morpheme analyzer user dictionary. pre-registering a morpheme analyzer user to register in; a word embedding and synonym group generation step of generating a synonym group by executing word embedding for all words including the extracted non-registered words; and a representative word setting step of setting a representative word in the generated synonym group and updating information on a connection relationship between the set representative word and the synonym in the user dictionary of the morpheme analyzer.

In addition, the step of extracting unregistered words may include a morpheme analysis step of separating words included in the document through a morpheme analyzer; a word separation check step of checking whether word separation is normally performed through the morpheme analyzer; and an unregistered word check step of checking whether the word is a word not registered in the morpheme analyzer user dictionary, when the word split is normally performed through the morpheme analyzer as a result of the check, and as a result of the check in the word split check step If the morpheme analyzer does not recognize it and is not divided into words, or if the morpheme analyzer recognizes it as a word but is not registered in the user dictionary of the morpheme analyzer as a result of checking in the unregistered word confirmation step, it is characterized in that it is extracted as an unregistered word.

As described above, according to the synonym extraction system and method through semantic analysis of unregistered words based on a deep learning model of the present invention, words with low frequency of occurrence, technical terms, new words, product names, company names, etc. are not recognized through morphological analysis. By extracting non-registered words that are not properly separated because they are not properly separated or recognized through morphological analysis but not registered in the dictionary, and immediately grasping the meaning of the extracted non-registered words and extracting and adding related synonyms, it is easy to analyze and search sentences. It can be utilized and has the effect of increasing the accuracy of natural language analysis including morpheme analysis.

However, the effects of the present invention are not limited to the above-mentioned effects, and effects not mentioned will be clearly understood by those skilled in the art from this specification and the accompanying drawings.

1 is a conceptual diagram showing the overall configuration including a synonym extraction system through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention.
2 is a block diagram showing the configuration of a synonym extraction system through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention in more detail.
3 is a diagram showing the hardware structure of a synonym extraction system through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention.
4 is a flowchart illustrating in detail the operation process of a synonym extraction method through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention.

Hereinafter, specific embodiments of the present invention will be described in detail with reference to the drawings. However, the spirit of the present invention is not limited to the presented embodiments, and those skilled in the art who understand the spirit of the present invention may add, change, delete, etc. other elements within the scope of the same spirit, through other degenerative inventions or the present invention. Other embodiments included within the scope of the inventive idea can be easily proposed, but it will also be said to be included within the scope of the inventive concept.

In addition, components having the same function within the scope of the same idea appearing in the drawings of each embodiment are described using the same reference numerals.

1 is a conceptual diagram showing the overall configuration including a synonym extraction system through semantic analysis of non-registered words based on a deep learning model according to an embodiment of the present invention.

As shown in FIG. 1, the present invention includes a synonym extraction system through semantic analysis of unregistered words based on a deep learning model (100, hereinafter abbreviated as a synonym extraction system through semantic analysis of unregistered words), an external server 200, It is configured to include a morpheme analyzer user dictionary 300 and the like.

The synonym extraction system 100 through semantic analysis of the unregistered word may be connected to a plurality of external servers 200 through a network.

The network may be a core network integrated with a wired public network, a wireless mobile communication network, or a mobile Internet, etc., and various services existing in the TCP/IP protocol and its upper layer, that is, HTTP (Hyper Text Transfer Protocol), HTTPS (Hyper Text Transfer Protocol Secure) ), Telnet, FTP (File Transfer Protocol), DNS (Domain Name System), SMTP (Simple Mail Transfer Protocol), etc. It comprehensively means a data communication network capable of transmitting and receiving data.

The synonym extraction system 100 through semantic analysis of the unregistered word collects various documents from a plurality of external servers 200 through a network, extracts unregistered words from the collected documents, analyzes the meaning, and then analyzes the unregistered word. By adding a synonym to replace a word to the user dictionary 300 of the morpheme analyzer, it is possible to increase the accuracy of natural language analysis including morpheme analysis.

That is, the synonym extraction system 100 through semantic analysis of unregistered words forms a vocabulary set through similarity comparison with existing words, so that it can be utilized, such as searching for related documents in which unregistered words are not used.

In particular, most of the non-registered word extraction algorithms used the frequency count to prevent registration of one-time unregistered words due to misspellings, etc., but in the present invention, the meaning of an unregistered word appears immediately, and by replacing it with a similar word (i.e., a synonym), To improve search accuracy.

In addition, the synonym extraction system 100 through semantic analysis of the unregistered words processes word embedding of all words including unregistered words extracted from the collected documents, configures synonym groups having similar meanings, and then Set representative words and update information on the connection between representative words and synonyms so that they can be used for sentence analysis and search.

Here, the document is electronically created or managed within the institution, and may include, for example, internal approval documents and report documents in which business progress is accumulated. However, it is not limited to this, and the type of the document can be variously modified at a level obvious to those skilled in the art.

Electronic documents may be documents containing data such as various texts, and may be in various formats such as various administrative documents, reports, theses, and evaluations, and may include odt, pdf, ppt, pptx, xls, xlsx, doc, docx, hwp, and hwpx. etc., and may be stored in a database.

The unregistered word is a word that is not properly separated because the morpheme analyzer does not recognize it, such as a word with a low frequency of occurrence in the documents collected from the plurality of external servers 200, a technical term, a newly coined word, a product name, a company name, or the morpheme analyzer. A word that has been recognized but has not previously appeared, and is a word that is out of vocabulary (OOV) during natural language analysis.

On the other hand, the synonym extraction system 100 through semantic analysis of the unregistered word extracts the unregistered word through an LR tokenizer, cohesion score, branch entropy, etc. can be extracted.

In the LR tokenizer, nouns, verbs, adjectives, and adverbs can be located in the L part, and when the L part is recognized, the remaining parts become the R part.

In addition, the cohesion score is a measure of how closely a group of strings appear together.

Also, branch entropy is a measure of how many different next characters appear at a string boundary.

The external server 200 is a public institution server, a corporate server, and/or a personal server, and refers to a server that holds and collects documents or calculates processed information through data processing.

The external server 200 mentioned in the present invention may include other configurations for performing a server environment, and may include all types of devices.

For example, the external server 200 is a digital device, and may be a digital device equipped with a processor and memory and computing capability, such as a laptop computer, a notebook computer, a desktop computer, a web pad, and a mobile phone. there is.

The morpheme analyzer user dictionary 300 is a dictionary basically used in the morpheme analyzer, and can be used flexibly because items can be easily added and modified.

In this case, the morpheme analyzer user dictionary 300 may sort and store representative words and words (synonyms) similar to the representative word by a certain similarity criterion or more in a group unit. For example, when the representative word is 'airplane', synonyms may be 'aircraft', 'flying object', and 'aircraft'. In addition, if it is determined that the word does not exist in the dictionary, the word may be stored as an unregistered word in a separate storage space.

2 is a block diagram showing the configuration of a synonym extraction system through semantic analysis of non-registered words according to an embodiment of the present invention in more detail.

As shown in FIG. 2, the synonym extraction system 100 through semantic analysis of the unregistered word includes a data collection unit 100, an unregistered word extraction unit 120, an appearance frequency check unit 130, a synonym extraction unit ( 140), a morpheme analyzer user dictionary registration unit 150, a word embedding and synonym group generation unit 160, a representative word setting unit 170, and the like.

The data collection unit 110 collects various documents from a plurality of external servers 200 connected through a network.

The document may be a summary of several sentences in one document or may be partially extracted sentences.

In addition, the data collection unit 110 may include a communication unit (not shown). The communication unit may include, for example, a cellular module, a WiFi module, a Bluetooth module, a GNSS module, an NFC module, an RF module, a 5G module, an LTE module, a NB-IOT module, and/or a LoRa module.

The non-registered word extraction unit 120 performs a function of extracting unregistered words from documents collected from the external server 200 through the data collection unit 110, and includes a morpheme analysis unit 121 and a word separation confirmation unit. (122) and an unregistered word confirmation unit (123).

The morpheme analysis unit 121 separates words included in the document collected by the data collection unit 110 through the morpheme analyzer.

The word separation confirmation unit 122 checks whether the word is normally separated through the morpheme analyzer in the morpheme analyzer 121, and as a result of the check, the morpheme analyzer does not recognize and is not separated into words as an unregistered word. extract

The non-registered word checking unit 123 determines that if word separation is normally performed through the morpheme analyzer as a result of checking by the word separation check unit 122, the corresponding word is a word that is not registered in the morpheme analyzer user dictionary 300 ( OOV, out of vocabulary), and as a result of the confirmation, the morpheme analyzer recognizes it as a word, but as a word that has not previously appeared and is not registered in the morpheme analyzer user dictionary 300, it is extracted as an unregistered word.

In addition, if the non-registered word confirmation unit 123 confirms that the word is registered in the morpheme analyzer user dictionary 300 as a result of confirmation by the word separation confirmation unit 122, there is no need to perform synonym extraction for the unregistered word, It supports sentence analysis and search utilization.

Here, the meaning of sentence analysis and search utilization is that the meaning of an input sentence is analyzed and used for document search. may be utilized for

The appearance frequency checking unit 130 determines whether the frequency of appearance of the non-registered word extracted by the non-registered word extractor 120 is equal to or greater than a preset number of times, and as a result of the determination, if the frequency of appearance of the non-registered word is less than the preset number of times, the The synonym extraction unit 140 allows synonyms to be replaced with the non-registered words to be extracted.

That is, when the occurrence frequency of unregistered words does not reach a certain number of times, the meaning of the unregistered words extracted from the document is analyzed and synonyms to replace the unregistered words are extracted so that they can be used for sentence analysis and search.

In addition, if the appearance frequency of the non-registered word extracted by the non-registered word extractor 120 is equal to or greater than a preset number of occurrences, the appearance frequency checking unit 130 may include the morpheme analyzer user dictionary registration unit 150, word embedding and thesaurus group generator 160 and the representative word setting unit 170, the words are added to the user dictionary, a synonym group is created through word embedding, and a representative word is set so that words of the same synonym group can be included in the search results.

That is, when the frequency of occurrence of unregistered words exceeds a certain number of times set in advance, the non-registered words extracted from the document are added to the user dictionary, a synonym group is created through word embedding, and a representative word is set so that words in the same synonym group are added. It is to enable them to use it for sentence analysis and search.

The synonym extractor 140 predicts at least one or more words similar to the unregistered word extracted by the unregistered word extractor 120, and extracts a word most similar to the unregistered word among the predicted words as a synonym replacing the unregistered word. do.

That is, the synonym extractor 140 removes unregistered words extracted from the collected documents, predicts words to be included in the part from which the unregistered words are removed, through a previously generated deep learning learning model, and sentences containing the predicted words. By comparing the similarity of the unregistered word with the similarity of the sentence containing the unregistered word and extracting a synonym to replace the unregistered word, the synonym extracted according to the semantic analysis of the unregistered word is supported for sentence analysis and search.

The synonym extraction unit 140 includes an unregistered word removal unit 141, a word prediction unit 142, a sentence similarity comparison unit 143, and an unregistered word synonym extraction unit 144.

The non-registered word removal unit 141 masks the non-registered word extracted by the non-registered word extraction unit 120 and provides the masked sentence to the word prediction unit 142 . That is, the non-registered word part in the sentence is made blank and transmitted to the word prediction unit 142.

The word prediction unit 142 predicts at least one word to be included in the masked portion through the non-registered word removal unit 141, and transfers the sentence including the predicted word to the sentence similarity comparison unit 143. to provide.

At this time, the word prediction unit 142 inputs the sentence masked by the unregistered word removal unit 141 to a pre-generated deep learning learning model for word prediction including BERT (Bidirectional Encoder Representations from Transformers), and the mask processing Predict the words that will fit in a part.

The sentence similarity comparison unit 143 compares the similarity between a sentence including the word predicted by the word prediction unit 142 and a sentence including the unregistered word, and sends the comparison result to the unregistered word synonym extractor 144. to provide.

That is, similarity between a sentence containing a predicted word and a sentence containing an unregistered word is compared, and information on a word having a similarity equal to or higher than a predetermined threshold value is provided to the non-registered word synonym extractor 144 .

The non-registered word synonym extractor 144 selects and extracts any one word among words equal to or higher than a preset threshold value as a synonym to replace the unregistered word based on the comparison result from the sentence similarity comparison unit 143.

At this time, since the unregistered word synonym extractor 144 selects a word having the most similar meaning to the unregistered word as a synonym in consideration of the context, correction effects such as misspellings can be expected.

The morpheme analyzer user dictionary registration unit 150 converts the non-registered word extracted by the non-registered word extractor 120 to the morpheme analyzer when the appearance frequency of the non-registered word is equal to or greater than a predetermined predetermined number of times as determined by the appearance frequency confirmation unit 130. Register in the user dictionary 300.

The word embedding and synonym group generation unit 160 executes word embedding on all words including unregistered words extracted by the unregistered word extractor 120, and creates a synonym group according to the execution result.

The representative word setting unit 170 sets a representative word in the synonym group created by the word embedding and synonym group generation unit 160, and transmits information about a connection between the set representative word and the synonym to the morpheme analyzer user dictionary. Update to 300.

3 is a diagram showing the hardware structure of a synonym extraction system through semantic analysis of unregistered words according to an embodiment of the present invention.

As shown in FIG. 3, the hardware structure of the synonym extraction system 100 through semantic analysis of the unregistered word includes a central processing unit 1000, a memory 2000, a user interface 3000, and a database interface 4000. , a network interface 5000, a web server 6000, and the like.

The user interface 3000 provides an input and output interface to the user by using a graphical user interface (GUI).

The database interface 4000 provides an interface between a database and a hardware structure. The network interface 5000 provides a network connection between devices possessed by a user.

The web server 6000 provides a means for users to access hardware structures through a network. Most users can access the web server remotely and use the synonym extraction system 100 through semantic analysis of the non-registered word.

Each step of the above configuration or method may be implemented as computer readable code on a computer readable recording medium or transmitted through a transmission medium. A computer-readable recording medium is a data storage device capable of storing data that can be read by a computer system.

Examples of computer-readable recording media include, but are not limited to, databases, ROMs, RAMs, CD-ROMs, DVDs, magnetic tapes, floppy disks, and optical data storage devices. The transmission medium may include a carrier wave transmitted through the Internet or various types of communication channels. Also, the computer readable recording medium may be distributed over a network coupled computer system such that the computer readable code is stored and executed in a distributed manner.

In addition, at least one component applied to the present invention may include or be implemented by a processor such as a central processing unit (CPU) or a microprocessor that performs each function, and two or more of the components are one single All operations or functions for two or more components that are combined into components can be performed. In addition, some of at least one or more components applied to the present invention may be performed by other components among these components. Also, communication between the components may be performed through a bus (not shown).

Next, an embodiment of a synonym extraction method through semantic analysis of unregistered words according to the present invention configured as described above will be described in detail with reference to FIG. 4 . At this time, the order of each step according to the method of the present invention may be changed by a user environment or a person skilled in the art.

4 is a flowchart showing in detail the operation process of a synonym extraction method through semantic analysis of unregistered words according to an embodiment of the present invention.

As shown in FIG. 4 , the synonym extraction system 100 through semantic analysis of the unregistered word collects documents from a plurality of external servers 200 connected through a network (S102).

Subsequently, non-registered words are extracted from the documents collected through step S102.

That is, the synonym extraction system 100 through semantic analysis of the non-registered words performs morphological analysis of the collected documents through a morpheme analyzer to separate words included in the documents (S104), and then word separation is normally performed. It is determined whether or not (S106).

As a result of the determination, if word separation is normally performed through the morpheme analyzer as a result of the confirmation, it is determined whether the word is not registered in the morpheme analyzer user dictionary 300 (S108). That is, it is checked whether a normally separated word is out of vocabulary (OOV) through morpheme analysis.

As a result of the determination in step S108, if it does not correspond to OOV, the synonym extraction system 100 through semantic analysis of the non-registered word supports sentence analysis and search utilization (S110).

If it is determined in step S106 that the morpheme analyzer does not recognize and is not divided into words, or as a result of determination in step S108, the morpheme analyzer recognizes the word as a word but is not registered in the user dictionary 300 of the morpheme analyzer, it is not registered. It is extracted by fish (S112).

Subsequently, the synonym extraction system 100 through the semantic analysis of the unregistered word determines whether the frequency of occurrence of the unregistered word extracted in step S112 is equal to or greater than a preset number of times (S114).

As a result of the determination in step S114, if the frequency of appearance of the non-registered word is less than the preset number of times, the unregistered word extracted in step S112 is masked to remove the non-registered word part from the sentence (S116), and deep learning for word prediction including BERT At least one word to be included in the masked portion is predicted through the learning model (S118).

Then, the similarity between the sentence containing the word predicted in step S118 and the sentence containing the unregistered word is compared (S120), and as a result of the comparison, any one of the words above a predetermined threshold value is used as a synonym to replace the unregistered word. By selecting and extracting (S122), the synonym extracted according to the semantic analysis of the non-registered word is supported for use in sentence analysis and search (S124).

On the other hand, if the synonym extraction system 100 through semantic analysis of the unregistered word determines in step S114 that the frequency of appearance of the unregistered word is equal to or greater than the predetermined number of times, the non-registered word extracted in step S112 is converted to the morpheme analyzer user dictionary 300. Register in (S126).

Thereafter, word embedding is executed for all words including the non-registered words (S128), and a synonym group is generated according to the word embedding result (S130).

In addition, a representative word is set in the synonym group generated in step S130 (S132), and information on a connection relationship between the set representative word and synonym is updated in the user dictionary 300 of the morpheme analyzer, so that the sentence in step S124 Support analysis and search utilization.

As such, the present invention extracts unregistered words that are not properly separated because they are not recognized through morpheme analysis, or are recognized through morpheme analysis but are not registered in advance, and extract related synonyms by immediately grasping the meaning of the extracted unregistered words. And, since it is added, it can be easily used for sentence analysis and search, and the accuracy of natural language analysis including morpheme analysis can be increased.

In the accompanying drawings, in order to more clearly express the technical idea of the present invention, components that are not related to or detached from the technical idea of the present invention are briefly expressed or omitted.

In the above, the configuration and characteristics of the present invention have been described based on the embodiments according to the present invention, but the present invention is not limited thereto, and various changes or modifications can be made within the spirit and scope of the present invention. It is apparent to those skilled in the art, and therefore such changes or modifications are intended to fall within the scope of the appended claims.

100: Synonym extraction system through semantic analysis of unregistered words based on deep learning model
110: data collection unit 120: unregistered word extraction unit
121: morpheme analysis unit 122: word separation confirmation unit
123: Unregistered word confirmation unit 130: Appearance frequency confirmation unit
140: synonym extraction unit 141: unregistered word removal unit
142: word prediction unit 143: sentence similarity comparison unit
144: Unregistered word synonym extraction unit 150: Morphological analyzer user dictionary registration unit
160: word embedding and synonym group generation unit
170: representative word setting unit 200: external server
300: morpheme analyzer user dictionary

Claims (12)

a data collection unit that collects documents;
an unregistered word extraction unit extracting unregistered words from the collected documents; and
Includes a synonym extraction unit for extracting a synonym to replace the extracted unregistered word;
The synonym extraction unit,
an unregistered word removal unit that masks the extracted unregistered words;
a word prediction unit that predicts at least one word to be included in the masked portion;
a sentence similarity comparing unit for comparing a similarity between a sentence containing the at least one predicted word and a sentence containing the unregistered word; and
As a result of the comparison, an unregistered word synonym extraction unit for selecting and extracting any one word among words equal to or greater than a predetermined threshold value as a synonym to replace the unregistered word;
A synonym extraction system through semantic analysis of unregistered words based on a deep learning model, characterized in that to support the use of synonyms extracted according to the semantic analysis of the unregistered words for sentence analysis and search.
delete The method of claim 1,
The prediction is
Deep learning model-based non-registration characterized by predicting words to enter the masked part by inputting the masked sentence to a pre-generated deep learning learning model for word prediction including BERT (Bidirectional Encoder Representations from Transformers) A synonym extraction system through semantic analysis of words.
The method of claim 1,
The system,
It further includes; an appearance frequency check unit for determining whether the appearance frequency of the extracted non-registered word is equal to or greater than a preset number of times,
As a result of checking by the appearance frequency confirmation unit, if the frequency of appearance of the extracted unregistered word is less than a predetermined number of times, a synonym to replace the unregistered word is extracted through the synonym extraction unit Semantic analysis of the unregistered word based on a deep learning model Synonym extraction system through.
The method of claim 4,
The system,
As a result of checking by the appearance frequency confirmation unit, if the appearance frequency of the extracted non-registered word is more than a preset number of times,
a morpheme analyzer user dictionary registration unit registering the extracted unregistered word in a morpheme analyzer user dictionary;
a word embedding and synonym group generation unit generating a synonym group by executing word embedding on all of the extracted words including unregistered words; and
Based on the deep learning model, further comprising a representative word setting unit for setting a representative word in the generated synonym group and updating information on a connection between the set representative word and the synonym in the user dictionary of the morpheme analyzer. A synonym extraction system through semantic analysis of unregistered words.
The method of claim 1,
The unregistered word extraction unit,
a morpheme analysis unit separating words included in the document through a morpheme analyzer;
a word separation confirmation unit checking whether word separation is normally performed through the morpheme analyzer; and
When word separation is normally performed through the morpheme analyzer as a result of the check, an unregistered word check unit confirming whether the word is not registered in the user dictionary of the morpheme analyzer;
As a result of confirmation by the word separation confirmation unit, if the morpheme analyzer does not recognize and is not separated into words, or if the result of confirmation by the unregistered word confirmation unit is that the morpheme analyzer recognizes the word as a word but is not registered in the user dictionary of the morpheme analyzer, it is an unregistered word. A synonym extraction system through semantic analysis of unregistered words based on a deep learning model, characterized in that for extraction.
The method of claim 1,
The unregistered word extraction unit,
Semantic analysis of unregistered words based on a deep learning model, characterized by extracting unregistered words from the collected documents using at least one of an LR tokenizer, cohesion score, and branch entropy Synonym extraction system through.
A data collection step of collecting documents in a synonym extraction system through semantic analysis of unregistered words based on a deep learning model;
an unregistered word extraction step of extracting unregistered words from the collected documents; and
A synonym extraction step of extracting a synonym to replace the extracted non-registered word; includes,
In the synonym extraction step,
In the synonym extraction system through the semantic analysis of the non-registered word based on the deep learning model, a non-registered word removal step of masking the non-registered word extracted in the non-registered word extraction step;
a word prediction step of predicting at least one word to be included in the masked portion;
a sentence similarity comparison step of comparing a similarity between a sentence containing the at least one predicted word and a sentence containing the unregistered word; and
As a result of the comparison, an unregistered word synonym extraction step of selecting and extracting any one word among words equal to or greater than a preset threshold value as a synonym to replace the unregistered word;
A method of extracting synonyms through semantic analysis of unregistered words based on a deep learning model, characterized in that supporting the use of synonyms extracted according to the semantic analysis of the unregistered words for sentence analysis and search.
The method of claim 8,
The prediction is a deep learning model-based non-registered word, characterized in that by inputting the masked sentence into a pre-generated deep learning learning model for word prediction including BERT, and predicting words to enter the masked part Synonym extraction method through analysis.
The method of claim 8,
The method,
In the synonym extraction system through the semantic analysis of the non-registered word based on the deep learning model, an appearance frequency check step of determining whether the frequency of appearance of the non-registered word extracted in the non-registered word extraction step is equal to or greater than a predetermined number of times; further comprising,
As a result of checking in the occurrence frequency checking step, if the frequency of appearance of the extracted non-registered word is less than a preset number of times, a synonym to replace the non-registered word is extracted through the synonym extraction step. Meaning of non-registered word based on a deep learning model Synonym extraction method through analysis.
The method of claim 10,
The method,
In the synonym extraction system through semantic analysis of unregistered words based on the deep learning model, as a result of checking in the occurrence frequency checking step, if the appearance frequency of the extracted unregistered word is more than a preset number of times,
a morpheme analyzer user dictionary registration step of registering the extracted unregistered word in a morpheme analyzer user dictionary;
a word embedding and synonym group generation step of generating a synonym group by executing word embedding for all words including the extracted non-registered words; and
A representative word setting step of setting a representative word in the generated synonym group and updating information on a connection between the set representative word and the synonym in the user dictionary of the morpheme analyzer; based on a deep learning model further comprising Synonym extraction method through semantic analysis of unregistered words.
The method of claim 8,
In the step of extracting unregistered words,
a morpheme analysis step of separating words included in the document through a morpheme analyzer;
a word separation check step of checking whether word separation is normally performed through the morpheme analyzer; and
When word separation is normally performed through the morpheme analyzer as a result of the check, an unregistered word check step of confirming whether the word is not registered in a user dictionary of the morpheme analyzer;
As a result of checking in the word separation confirmation step, if the morpheme analyzer does not recognize it and is not separated into words, or if it is confirmed in the non-registered word check step that the morpheme analyzer recognizes it as a word but is not registered in the user dictionary of the morpheme analyzer, it is an unregistered word. Synonym extraction method through semantic analysis of unregistered words based on deep learning model, characterized in that for extraction.