KR20240047736A - Method for domain-specific keyword extraction and system thereof - Google Patents

Abstract

This disclosure relates to a method for extracting domain core keywords and a system to which the method is applied. The domain core keyword extraction method according to the present disclosure includes obtaining first domain reference keywords extracted from a document set of a first domain, and using a previously learned embedding model to obtain a first embedding vector for the target document. obtaining second embedding vectors for some of the keywords extracted from the target document using the embedding model, and using the embedding model to obtain second embedding vectors for each of the first domain reference keywords. Obtaining third embedding vectors, and among the second embedding vectors and the third embedding vectors, a keyword of an embedding vector having a similarity of more than a reference value to the first embedding vector is used for the first domain of the target document. It may include the step of outputting key keywords.

Description

Domain key keyword extraction method and system {METHOD FOR DOMAIN-SPECIFIC KEYWORD EXTRACTION AND SYSTEM THEREOF}

This disclosure relates to a method for extracting domain core keywords and a system to which the method is applied. More specifically, it relates to a target document, a method for extracting important keywords in a specific field not included in the target document, and a system to which the method is applied.

Technologies for extracting key keywords included in a specific document set have been provided in the past through techniques such as artificial intelligence or text mining. However, these prior technologies only present a method of extracting words that frequently appear in previously collected documents as key keywords.

However, even if the previously collected document does not correspond to a specific field, the document may contain many keywords related to terms commonly used in the specific field. For example, there may be many keywords related to the technology commercialization field in a document in the technology field, but the user cannot extract key keywords (terms) in the technology commercialization field from the technology field document simply by using the above-mentioned prior technology.

In addition, the prior art extracts key keywords from a specific document set using text mining techniques such as TF-IDF (Term Frequency-Inverse Document Frequency), but this simply extracts keywords based on the frequency of their occurrence in the document. Because it is an extraction method, there are frequent errors in extracting even keywords that are not related to the field to which the document belongs, which requires post-editing by the user.

Therefore, there is a need to provide a method to automatically extract key keywords in the same field that do not appear in a specific document set and the specific document, and a system to which the method is applied, but existing keyword extraction methods do not provide such a function.

The technical task to be achieved through some embodiments of the present disclosure is related to key keywords of previously collected documents and related fields of the documents, and a method for extracting key keywords that are not included in the previously collected documents and the same. The goal is to provide a system to which the method has been applied.

Another technical task to be achieved through some embodiments of the present disclosure is to provide a method of efficiently performing keyword extraction within a set keyword extraction limit.

Another technical task to be achieved through some embodiments of the present disclosure is to provide a method of training an artificial intelligence language model to be specialized in a specific field.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below.

The domain core keyword extraction method according to an embodiment of the present disclosure to solve the above technical problem includes obtaining first domain reference keywords extracted from a document set of the first domain, and using a previously learned embedding model. Thus, obtaining a first embedding vector for the target document, using the embedding model, obtaining second embedding vectors for some of the keywords extracted from the target document, and using the embedding model Thus, obtaining third embedding vectors for each of the first domain reference keywords, and among the second embedding vectors and the third embedding vectors, an embedding vector having a similarity greater than a reference value with the first embedding vector. It may include outputting a keyword as a core keyword for the first domain of the target document.

In some embodiments, the target document may not be included in the document set of the first domain.

In some embodiments, outputting keywords as core keywords for the first domain may include outputting keywords of some of the third embedding vectors as core keywords for the first domain of the target document. there is.

In some embodiments, the method for extracting domain core keywords includes key keywords for the first domain having a similarity of the first embedding vector below a standard value, and core keywords for another first domain having a similarity of more than the standard value. A step of excluding keywords having a keyword from core keywords for the first domain may be further included.

In some embodiments, keywords extracted from the target document may include single-word keywords, double-word keywords, and triple-word keywords.

A domain core keyword extraction system according to another embodiment of the present disclosure for solving the above technical problem includes a network interface for receiving a plurality of documents, a memory on which a domain core keyword extraction program is loaded, and the domain core keyword extraction program. It may include one or more processors to execute, and the domain core keyword extraction program uses instructions for obtaining first domain reference keywords extracted from a document set of the first domain and a previously learned embedding model. , an instruction for obtaining a first embedding vector for the target document, an instruction for obtaining second embedding vectors for some of the keywords extracted from the target document using the embedding model, and using the embedding model , an instruction for obtaining third embedding vectors for each of the first domain reference keywords, and a keyword of an embedding vector having a similarity of more than a reference value to the first embedding vector among the second embedding vectors and the third embedding vectors. It may include an instruction to output as a key keyword for the first domain of the target document.

In some embodiments, the target document may not be included in the document set of the first domain.

In some embodiments, the instruction for outputting keywords as core keywords for the first domain may include instructions for outputting keywords of some of the third embedding vectors as core keywords for the first domain of the target document. there is.

In some embodiments, the domain core keyword extraction program selects core keywords for the first domain that have a similarity of the first embedding vector or less than a standard value and to core keywords for another first domain that have a similarity of more than the standard value. It may further include an instruction for excluding the keyword having a keyword from the core keyword for the first domain.

In some embodiments, keywords extracted from the target document may include single-word keywords, double-word keywords, and triple-word keywords.

1 is an exemplary block diagram illustrating a domain core keyword extraction system according to an embodiment of the present disclosure.
FIG. 2 is a diagram illustrating the differences between an embedding model according to some embodiments of the present disclosure and a conventional embedding model.
Figure 3 is a diagram for explaining an embedding model used in some embodiments of the present disclosure.
FIG. 4 is a diagram illustrating the number of learning losses that change as a result of learning an embedding model according to some embodiments of the present disclosure.
Figure 5 is a flowchart of a method for extracting domain core keywords according to another embodiment of the present disclosure.
FIG. 6 is a diagram illustrating the steps of obtaining a first embedding vector and a second embedding vector that can be performed in some embodiments of the present disclosure.
FIG. 7 is a diagram illustrating a method of obtaining a third embedding vector that can be performed in some embodiments of the present disclosure.
Figure 8 is a hardware configuration diagram of a domain core keyword extraction system according to another embodiment of the present disclosure.

Hereinafter, preferred embodiments of the present disclosure will be described in detail with reference to the attached drawings. The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings. However, the technical idea of the present invention is not limited to the following embodiments and may be implemented in various different forms. The following examples are merely intended to complete the technical idea of the present invention and to be used in the technical field to which the present invention pertains. It is provided to fully inform those skilled in the art of the scope of the present invention, and the technical idea of the present invention is only defined by the scope of the claims.

Before describing various embodiments of the present disclosure, terms used in the following embodiments will be clarified.

In the following embodiments, ‘Domain’ may refer to a field that is the target of keyword extraction. For example, in the case of 'extracting keywords for technology commercialization domain', this may mean extracting terms in the technology commercialization field.

In the following embodiments, 'Embedding model' may refer to an artificial intelligence model that converts natural language used by humans into a vector, a numerical form that machines can understand. In the technical field, the embedding model may be used interchangeably with terms such as 'language model'.

In the following embodiments, an 'embedding vector' may refer to a vector derived as a result of converting a specific keyword into a vector by the embedding model.

In describing the present disclosure, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description will be omitted.

Hereinafter, several embodiments of the present disclosure will be described with reference to the drawings.

FIG. 1 is a block diagram illustrating the structure of a domain core keyword extraction system 1000 according to an embodiment of the present disclosure.

Referring to FIG. 1, a domain core keyword extraction system 1000 according to some embodiments of the present disclosure may include a document pre-processing unit 110. In some embodiments, the document preprocessing unit 110 may perform an operation of preprocessing documents input to the domain core keyword extraction system 1000 according to specific rules.

In some embodiments of the present disclosure, the document preprocessor 110 may perform normalization on text data included in a document input to the domain core keyword extraction system 1000.

Here, the normalization may mean integrating words with different expression methods but the same actual meaning into a single expression. For example, 'USA' and 'US' correspond to texts with different expressions, but actually correspond to the same word meaning the United States, so they can be unified as 'USA' by the document preprocessing unit 110. . More preferably, just as unification into the 'US' expression requires less storage space than the 'USA' expression, the document pre-processing unit 110 selects substantially the same words among the plurality of expressions signified by the words. It can be integrated into a representation that requires the least amount of storage space.

In some other embodiments of the present disclosure, the document pre-processing unit 110 may integrate uppercase and lowercase distinction of English text data included in a document input to the domain core keyword extraction system 1000.

In some other embodiments of the present disclosure, the document pre-processing unit 1100 may remove words whose frequency of appearance is less than the reference value from the document input to the domain core keyword extraction system 1000. According to this embodiment, the domain core keyword extraction system 1000 extracts specific keywords from the document by deleting data that does not affect the analysis of sentences included in the input document but occupies unnecessary storage space. The effect of saving resources consumed in operation can be achieved.

In some other embodiments of the present disclosure, the document pre-processing unit 1100 may delete stopwords included in text data of a document input to the domain core keyword extraction system 1000. Here, the stop word may mean an exclamation point or the like that is included in text data but has no actual meaning.

The domain core keyword extraction system 1000 according to some embodiments of the present disclosure may include an embedding model learning unit 130. In some embodiments, the embedding model learning unit 130 may perform an operation of training an embedding model that outputs an embedding vector as a result of inputting specific text data. Hereinafter, the description will be made with reference to FIG. 2.

In some embodiments of the present disclosure, the embedding model learning unit 130 inputs data about a specific domain into the previously learned Korean-based embedding model 21 and trains it to improve analysis performance for text related to the specific domain. An additional learned embedding model (65) can be constructed. Here, the specific domain may be a predefined target domain on which the domain core keyword extraction system 1000 performs the core keyword extraction operation, but the domain core keyword extraction system 1000 receives input to perform the core keyword extraction operation. It may be automatically identified as a domain related to the target document.

However, the fact that the conventional embedding model 21 is a Korean-based embedding model 21 learned to analyze the Korean language is only an example to help understand the present disclosure, and is not limited to the language analyzed by the embedding model 21. No. Additionally, in some other embodiments, the embedding model 21 is not previously learned and may be a model constructed as a result of learning by the embedding model learning unit 13.

In some other embodiments of the present disclosure, the embedding model 21 may be a natural language processing model based on an artificial neural network. That is, the embedding model 21 may be understood as one of a T5 model, a BERT (Bidirectional Encoder Representations from Transformers) model, and a GPT model, but is not limited thereto.

More preferably, in order for the domain core keyword extraction system 1000 to optimize the performance of extracting specific core keywords from documents of a specific domain, the embedding model 21 may be a BERT model. Here, the BERT model can analyze the meaning of each keyword included in the sentence by referring to information about keywords other than the keywords included in the sentence. 3, when the words [CLS] (32) and I, Love, You (33) are input together into the embedding model (65), the [CLS] (32) input is a single word, but the output The information included in [CLS] (31) may include information on the words I, Love, You (33).

In some other embodiments of the present disclosure, referring to FIG. 4, the embedding model learning unit 130 may construct an additionally learned embedding model 65 by inputting data about a specific domain into the embedding model 21. However, it can be understood that the number of times the additionally learned embedding model 65 learns the entire learning target data and the number of learning losses occur are inversely proportional.

The domain core keyword extraction system 1000 according to some embodiments of the present disclosure may include a core keyword extraction unit 140. In some embodiments, the core keyword extraction unit 140 may perform an operation of extracting core keywords from a document input to the domain core keyword extraction system 1000 based on predefined rules.

In some embodiments of the present disclosure, the core keyword extraction unit 140 may measure the frequency with which each keyword is included in the document from the text included in the document input to the domain core keyword extraction system 1000. Here, keywords with a frequency greater than the reference value can be extracted as reference keywords for the document.

In some other embodiments of the present disclosure, the core keyword extraction unit 140 is operated by the domain core keyword extraction system 1000 based on the calculation result of the similarity comparison unit 150 included in the domain core keyword extraction system 1000. The operation of extracting key keywords from the input document can be performed, but this will be described later.

The domain core keyword extraction system 1000 according to some embodiments of the present disclosure may include a keyword purification unit 120. The keyword refinement unit 120 may perform an operation of deleting some of the core keywords that are determined to be unnecessary based on predefined rules from the keywords extracted as a result of the core keyword extraction operation of the core keyword extraction unit 140. You can.

In some embodiments of the present disclosure, the keyword refinement unit 120 compares each vectorized core keyword with a document input to the vectorized domain core keyword extraction system 1000 to determine the similarity between the vectorized core keyword and the document. The result of measuring can be received from the similarity comparison unit 150. Additionally, key keywords whose similarity to the document is below the standard value may be determined to be unnecessary and deleted. Here, the key keywords may not be included in the document.

In some other embodiments of the present disclosure, the keyword refinement unit 120 measures the similarity between each vectorized core keyword, and determines the keyword of the core keyword vector that has a similarity greater than the standard value with other core keyword vectors except itself as unnecessary. You can judge and delete it. However, the similarity between each of the vectorized core keywords can be understood by those skilled in the art to mean the distance between each other when the vector of each of the core keywords is expressed in n-dimensional coordinates. .

In some other embodiments of the present disclosure, the keyword refinement unit 120 assigns the similarity between the vectorized core keyword and the document as a first score to each core keyword extracted by the core keyword extraction unit 140. Similarity with other core keyword vectors except , can be assigned as a second score. Additionally, if there is a core keyword for which the sum of the first score and the second score is greater than or equal to a standard value, the core keyword may be determined to be unnecessary and deleted.

The domain core keyword extraction system 1000 according to some embodiments of the present disclosure may include a similarity comparison unit 150. In some embodiments, the similarity comparison unit 150 determines the similarity between a plurality of documents input to the domain core keyword extraction system 1000 and the core keyword extraction unit 140 in the documents input to the domain core keyword extraction system 1000. The similarity between core keywords extracted through the core keyword extraction operation can be measured.

In some embodiments of the present disclosure, the similarity comparison unit may compare each vectorized core keyword with a document input to the vectorized domain core keyword extraction system 1000 to measure the similarity between the vectorized core keyword and the document. . Additionally, the similarity comparison unit 150 may transmit the measured similarity to the keyword refining unit 120.

In some embodiments of the present disclosure, the similarity comparison unit 150 measures the similarity between a core keyword vector obtained by vectorizing core keywords extracted from a document input to the domain core keyword extraction system 1000 and a document vector obtained by vectorizing the document. can do. Additionally, the similarity comparison unit 150 may transmit the measured similarity to the key keyword extraction unit 140.

So far, the configuration and operation of the domain core keyword extraction system 1000 and an exemplary environment to which the domain core keyword extraction system 1000 can be applied have been described with reference to FIGS. 1 to 4. In addition, various embodiments related to the operation of the domain key keyword extraction system 1000 are mentioned. It should be noted that the above embodiments should be understood as illustrative only and are not limiting.

Hereinafter, a method for extracting domain core keywords according to another embodiment of the present disclosure will be described in detail with reference to FIGS. 5 to 7.

Figure 5 is a flowchart of a method for extracting domain core keywords according to another embodiment of the present disclosure. The steps described in the flowchart below can be understood as being performed by a domain core keyword extraction system unless otherwise specified.

In step S100 shown in FIG. 5, the domain core keyword extraction system may obtain a first domain reference keyword. Here, it should be noted that the first domain illustratively represents the domain that is the target of the core keyword extraction operation performed by the domain core keyword extraction system, and is not limited.

In some embodiments of the present disclosure, the first domain may be determined by the domain core keyword extraction system according to the content of the input document.

In some embodiments related to step S100, the domain core keyword extraction system calculates, in the text included in the document set related to the input first domain, the frequency with which each word included in the text appears in the document, and the appearance Words with a frequency equal to or higher than the standard value may be identified as first domain reference keywords.

In some other embodiments related to step S100, the domain core keyword extraction system calculates the frequency of occurrence of each word included in the target document in the text included in the input target document, and the frequency of appearance is Words that are above the standard can be identified as target document reference keywords. Here, the target document may be a document corresponding to the same domain as the first domain, but may not be included in the document set of the first domain received by the domain key keyword extraction system.

Additionally, identifying the target document reference keyword may include identifying a single-word reference keyword, identifying a double-word reference keyword, and identifying a triple-word reference keyword. Here, a single-word reference keyword may be understood as a reference keyword consisting of one word, a double-word reference keyword may be understood as a reference keyword consisting of two words, and a triple-word reference keyword may be understood as a reference keyword consisting of three words.

Each of the single-word reference keywords, double-word reference keywords, and triple-word reference keywords refers to 'Unigram reference keywords', 'Bigram reference keywords', and 'Trigram reference keywords' in the art. It can be understood in terms.

For example, assuming that the target document received by the domain core keyword extraction system includes the sentence 'Application of automatic mathematical proposition generation problem for computable numbers', the above sentence is used when identifying single word reference keywords. Separated into 'computation', 'possible', 'number', 'about', 'mathematical', 'proposition', 'automatic', 'generation', 'problem', 'application', etc., but double word reference keywords In the step of identifying 'computable', 'possible number', 'about number', 'about mathematical', 'numerical proposition', 'proposition automatic', 'automatically generated', 'generating problem', 'problem' It can be separated into 'application of', etc.

Additionally, single-word reference keywords, double-word reference keywords, and triple-word reference keywords can be stored as separate reference keywords. More preferably, the domain core keyword extraction system may separately configure each group of single-word reference keywords, double-word reference keywords, and triple-word reference keywords.

As mentioned above, the domain core keyword extraction system acquires reference keywords for the input document based on the frequency with which each word included in the text appears in the document. This is only an example to help understand the present disclosure, and the reference keyword is a specific There are no limitations to the applicable rules.

Next, in step S200, the domain core keyword extraction system may obtain a first embedding vector for the input target document using an embedding model additionally learned by the embedding model learning unit of the domain core keyword extraction system. Meanwhile, the additionally learned embedding model may be an embedding model that is additionally learned using a document related to the first domain as learning data. Hereinafter, it will be described with reference to FIG. 6.

The first embedding vector 66 may be in the form of a vector of the target document 61 output as a result of the additionally learned embedding model 65 receiving text data of the target document 61. The method by which the target document 61 is converted into text data by the additionally learned embedding model 65 may differ depending on the type of embedding model. For example, in the case of the BERT embedding model according to a preferred embodiment of the present disclosure, the target document 61 is a 768-dimensional first embedding containing information about the association relationship between each word included in the target document 61. It can be converted to vector (66).

In some embodiments related to step S200, the domain key keyword extraction system may preprocess the target document 61 before inputting it into the additional learned embedding model 65. The operation of preprocessing the target document 61 may be any one of normalization, integration of uppercase and lowercase letters, and deletion of stop words.

In step S300, referring to FIG. 6, the domain core keyword extraction system inputs information about some keywords (62, 63, 64) extracted from the target document (61) into the additionally learned embedding model (65) to create a second Embedding vectors (67, 68, 69) can be obtained.

In some embodiments related to step S300, some keywords 62, 63, and 64 extracted from the target document 61 are included among words included in the target document 61 at a frequency greater than the reference value in the target document 61. It could be a word.

In some embodiments related to step S300, the domain key keyword extraction system inputs the single word reference keyword 62 of the target document 61 into the additional learned embedding model 65 to generate a second embedding for the single word reference keyword. Vector 67 can be obtained.

In some other embodiments related to step S300, the domain core keyword extraction system inputs the double-word reference keywords 63 of the target document 61 into the additional learned embedding model 65 to generate a second search for the double-word reference keywords. The embedding vector 68 can be obtained.

In some other embodiments related to step S300, the domain core keyword extraction system inputs the triple-word reference keywords 64 of the target document 61 into the additionally trained embedding model 65 to generate a search engine for the triple-word reference keywords. 2 Embedding vector 69 can be obtained.

In some embodiments of the present disclosure, the domain core keyword extraction system includes second embedding vectors 67, 68, 69 for single-word reference keywords, double-word reference keywords, and triple-word reference keywords obtained by the domain core keyword extraction system. ) Similarity measurement with the first embedding vector 66 obtained in step S200 may be performed for each, but this will be described later.

In step S400, the domain core keyword extraction system may obtain third embedding vectors for each of the first domain reference keywords obtained in step S100. In order to explain the step of acquiring the third embedding vectors in more detail, it will be described with reference to FIG. 7.

In some embodiments related to step S400, referring to FIG. 7, the domain core keyword extraction system inputs the first domain reference keyword 82 pre-extracted in step S100 into the additionally learned embedding model 65 to obtain the first domain reference keyword 82. Third embedding vectors 83 corresponding to each domain reference keyword 82 can be obtained.

In some embodiments of the present disclosure, the domain core keyword extraction system measures the similarity of each third embedding vector 83 obtained by the domain core keyword extraction system with the first embedding vector 66 obtained in step S200. can be performed, but this will be described later.

Next, in step S500, the domain core keyword extraction system may extract the core keyword of the first domain based on the embedding vectors obtained in steps S200 to S400. For a more detailed description of the step of extracting the first domain core keyword, please refer to FIG. 7.

In some embodiments related to step S500, the domain core keyword extraction system measures the similarity with the first embedding vector 66 for each of the second embedding vectors 67 of the single word reference keyword 62 obtained in step S300; , the single-word reference keyword 62 of the second embedding vector 67 whose similarity is greater than or equal to the reference value may be determined as the single-word core keyword 91 for the target document 61. However, the single word core keyword 91 may be any one of the keywords included in the target document 61.

Here, the similarity between the second embedding vector 67 and the first embedding vector 66 of the domain core keyword extraction system may be measured by the cosine similarity technique or the Euclidean distance technique. , the present disclosure is not limited thereto, and is not limited to any common method in the field of natural language processing that measures similarity between vectors. In addition, it goes without saying that the method of measuring the similarity between the above-described embedding vectors is also applied to other embodiments in which the domain core keyword extraction system, which will be described later, measures the similarity between a plurality of embedding vectors.

In some other embodiments related to step S500, the domain core keyword extraction system measures the similarity with the first embedding vector 66 for each of the second embedding vectors 68 of the double word reference keyword 63 obtained in step S300. And, the double-word reference keyword 63 of the second embedding vector 68 whose similarity is greater than or equal to the reference value may be determined as the double-word key keyword 92 for the target document 61. However, the double-word core keyword 92 may be any one of compound text consisting of two words included in the target document 61.

In some other embodiments related to step S500, the domain core keyword extraction system determines the similarity with the first embedding vector 66 for each of the second embedding vectors 69 of the triple word reference keyword 64 obtained in step S300. measurement, and the triple word reference keyword 64 of the second embedding vector 69 whose similarity is greater than or equal to the reference value can be determined as the triple word core keyword 93 for the target document 61. However, the triple word core keyword 93 may be any one of the compound text consisting of three words included in the target document 61.

In some other embodiments related to step S500, the domain core keyword extraction system determines the similarity with the first embedding vector 66 for each of the third embedding vectors 83 of the first domain reference keyword 82 obtained in step S400. can be measured, and the first domain reference keyword 82 of the third embedding vector 83 whose similarity is greater than or equal to the reference value can be determined as the first domain key keyword 94 for the target document 61. However, the first domain core keyword 94 may be included in the target document 61.

In some other embodiments related to step S500, the domain core keyword extraction system extracts single word core keywords (91), double word core keywords (92), triple word core keywords (93), and first domain core keywords (94). It may also be included in the first domain final core keyword group 160 shown in FIG. 7.

Conventionally, a method of extracting core keywords from a target document was recommended based on the frequency of occurrence of words included in the target document, but in such cases, meaningless core keywords that the user does not want are extracted or even though they are related to the domain the user is trying to extract. However, there was a problem with keywords not being extracted due to their low frequency of appearance in the target document. According to this embodiment, the domain core keyword extraction system can not only accurately extract core keywords related to the first domain from the target document, but also has the effect of extracting core keywords for the first domain that are not included in the target document. It can also be achieved.

Next, in step S600, the domain core keyword extraction system may refine keywords included in the first domain final core keyword group. In some embodiments of the present disclosure, the operation of refining the keyword may mean an operation of leaving only one of a plurality of keywords that have substantially the same meaning but different expressions and deleting the rest. Hereinafter, description will be made with reference to FIG. 7.

For example, in the first domain final core keyword group 160, a plurality of words such as 'graphene', 'graphene application case', 'graphene application case proximity', 'biosensor', 'biosensor application case', etc. Assuming that there is a keyword, 'graphene application examples' and 'graphene application examples close to' all contain additional meanings to 'graphene', and the meaning as keywords is substantially the same, so they can be deleted. will be. However, the criteria for deleting keywords will be described later.

In some embodiments related to step S600, the domain core keyword extraction system may retain only one of the plurality of identical keywords included in the first domain final core keyword group 160 and delete the rest.

For example, the keyword ‘graphene’ included in the single-word core keyword (91) extracted from the target document (61) and the keyword ‘graffin’ included in the first domain core keyword (94) extracted from the first domain document set (81). If all 'pin' keywords are included in the first domain final core keyword group 160, the domain core keyword extraction system determines that two or more of the same keywords exist, and the 'pin' keyword extracted from the first domain document set 81 The keyword ‘graphene’ can be deleted.

In some other embodiments related to step S600, the domain core keyword extraction system includes an embedding vector (67, 68, 69, 83) and a first embedding vector (66) for each of the keywords included in the first domain final core keyword group (160). ) can be measured, and keywords corresponding to the embedding vectors (67, 68, 69, 83) whose similarity is below the standard value can be excluded from the first domain final core keyword group (160).

For example, there are the 'graphene' keyword and the 'application case' keyword extracted from the first domain document set 81 and included in the first domain final core keyword group 160, and the 'application case' keyword Based on the fact that the embedding vector has a similarity below the standard value for the first embedding vector 66, the domain core keyword extraction system determines that the 'Application Case' keyword is an inappropriate keyword to represent the target document 61, and , the 'Application Case' keyword can be removed from the first domain final core keyword group 160.

In some other embodiments related to step S600, the domain core keyword extraction system measures the similarity between the embedding vectors 67, 68, 69, and 83 of each of the keywords included in the first domain final core keyword group 160, Keywords corresponding to the embedding vectors (67, 68, 69, 83) whose similarity is higher than the standard value may be excluded from the first domain final core keyword group (160).

For example, the 'graphene' keyword extracted from the first domain document set 81 and included in the first domain final core keyword group 160, and the double word reference keyword 92 from the target document 61. Extracted, the 'graphene application case' keyword included in the first domain final core keyword group 160, and the single word reference keyword 91 from the target document 61 are extracted, and the first domain final core keyword group ( When the 'biosensor' keyword included in 160) exists, the 'graphene application example' keyword may be excluded from the first domain final core keyword group (160) because it has a similarity higher than the standard value with the 'graphene' keyword. 'Biosensor' may not be excluded because it does not have a similarity higher than the standard value for any keyword.

In some further embodiments related to step S600, the domain core keyword extraction system may include a first embedding vector 66 in each of the embedding vectors 67, 68, 69, and 83, respectively. The similarity to may be assigned as a first score, and the similarity to other embedding vectors 67, 68, 69, and 83 may be assigned to each of the embedding vectors 67, 68, 69, and 83 as a second score. In addition, the domain core keyword extraction system may exclude keywords of the embedding vectors 67, 68, 69, and 83 in which the sum of the first score and the second score is greater than the standard value from the first domain final core keyword group 160. .

In some other embodiments related to step S600, when the domain core keyword extraction system refines the first domain final core keyword based on the sum of the first score and the second score, the domain core keyword extraction system may use a predefined The score agreement standard may vary depending on the limited number of keywords extracted.

For example, the first domain final core keyword group 160 includes the keywords 'graphene', 'graphene application cases' keywords, 'graphene application case proximity' keywords, 'biosensor' keywords, and 'bio Assuming that the 'sensor application case' keyword exists, the domain core keyword extraction system can delete the rest except for the 'graphene' and 'biosensor' keywords when the predefined limited number of keywords extracted is 2, but When the limited number of keywords extracted is 3, the rest except for the 'graphene' keyword, 'biosensor' keyword, and 'biosensor application case' keyword can be deleted.

In some other embodiments related to step S600, the domain key keyword extraction system extracts the embedding vectors 67, 68, 69, and 83 that are produced as a result of performing a decision point process operation on each of the embedding vectors 67, 68, 69, and 83. ) Based on the probability that each is included in the first domain final core keyword group 160, it can be determined whether each of the embedding vectors 67, 68, 69, and 83 includes the first domain final core keyword group 160.

Here, the probability that each of the embedding vectors (67, 68, 69, and 83) is included in the first domain final core keyword group (160) is determined by the determinant value of each of the embedding vectors (67, 68, 69, and 83). It may be proportional. Additionally, the determinants of each of the embedding vectors 67, 68, 69, and 83 can be calculated as follows.

For example, assuming that a specific embedding vector S exists, the determinant of S can be calculated as follows. However, in the above formula, () may mean an arbitrary Fourier feature value for a specific embedding vector, ) may mean the similarity between a specific embedding vector and the first embedding vector 66.

So far, the domain core keyword extraction method according to an embodiment of the present disclosure has been described in detail.

Figure 8 is a hardware configuration diagram of a domain core keyword extraction system according to some embodiments of the present disclosure. For example, the domain core keyword extraction system 1000 shown in FIG. 8 may refer to the domain core keyword extraction system 1000 described with reference to FIG. 1 .

The domain key keyword extraction system 1000 includes one or more processors 1100, a system bus 1600, a communication interface 1200, and a memory 1400 that loads a computer program 1500 executed by the processor 1100. ) and a storage 1300 that stores the computer program 1500.

The processor 1100 controls the overall operation of each component of the domain key keyword extraction system 1000. The processor 1100 may perform operations on at least one application or program to execute methods/operations according to various embodiments of the present disclosure. The memory 1400 stores various data, commands and/or information. The memory 1400 may load one or more computer programs 1500 from the storage 1300 to execute methods/operations according to various embodiments of the present disclosure.

The bus 1600 provides a communication function between components of the domain key keyword extraction system 1000.

The communication interface 1200 supports Internet communication of the domain key keyword extraction system 1000.

Storage 1300 may non-temporarily store one or more computer programs 1500.

The computer program 1500 may include one or more instructions implementing methods/operations according to various embodiments of the present disclosure. When the computer program 1500 is loaded into the memory 1400, the processor 1100 can perform methods/operations according to various embodiments of the present disclosure by executing the one or more instructions.

For example, the computer program 1500 operates to obtain first domain reference keywords extracted from a document set of the first domain and obtains a first embedding vector for the target document using a previously learned embedding model. An operation of obtaining second embedding vectors for some of the keywords extracted from the target document using the embedding model, and using the embedding model to obtain second embedding vectors for each of the first domain reference keywords. An operation of acquiring third embedding vectors, and among the second and third embedding vectors, a keyword of an embedding vector having a similarity level of more than a reference value to the first embedding vector is used for the first domain of the target document. As key keywords, it may include instructions that perform output operations.

In some embodiments, the domain key keyword extraction system 1000 may be configured using one or more physical servers included in a server farm based on cloud technology such as a virtual machine.

So far, various embodiments of the present disclosure and effects according to the embodiments have been mentioned with reference to FIGS. 1 to 8 . The effects according to the technical idea of the present disclosure are not limited to the effects mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the description below.

The technical ideas of the present disclosure described so far can be implemented as computer-readable code on a computer-readable medium. The computer program recorded on the computer-readable recording medium can be transmitted to another computing device through a network such as the Internet, installed on the other computing device, and thus used on the other computing device.

Although operations are shown in the drawings in a specific order, it should not be understood that the operations must be performed in the specific order shown or sequential order or that all illustrated operations must be performed to obtain the desired results. In certain situations, multitasking and parallel processing may be advantageous. Although embodiments of the present disclosure have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing the technical idea or essential features. I can understand that there is. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive. The scope of protection of the present invention should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be construed as being included in the scope of rights of the technical ideas defined by this disclosure.

Claims (12)

In a method of extracting domain core keywords from a target document on a computing device,
Obtaining first domain reference keywords extracted from a document set of the first domain;
Obtaining a first embedding vector for a target document using a previously learned embedding model;
Obtaining second embedding vectors for some of the keywords extracted from the target document using the embedding model;
Obtaining third embedding vectors for each of the first domain reference keywords using the embedding model; and
Outputting, among the second and third embedding vectors, a keyword of an embedding vector having a similarity level greater than or equal to a reference value with the first embedding vector as a key keyword for the first domain of the target document. doing,
How to extract domain core keywords. According to claim 1,
The target document is not included in the document set of the first domain,
How to extract domain core keywords. According to claim 1,
The step of outputting as a core keyword for the first domain is,
Including, outputting keywords of some of the third embedding vectors as core keywords for the first domain of the target document.
How to extract domain core keywords. According to claim 1,
Among the core keywords for the first domain, keywords that have a similarity of less than a standard value to the first embedding vector and have a similarity of more than a standard value to core keywords of other first domains are excluded from the core keywords of the first domain. further comprising steps;
How to extract domain core keywords. According to claim 1,
Keywords extracted from the target document are:
including single word keywords, double word keywords and triple word keywords,
How to extract domain core keywords. According to claim 1,
Based on the determinant value of each vector of the core keywords for the first domain, which is calculated as a result of performing a decision point process operation for each vector of the core keywords for the first domain, the core keywords for the first domain Further comprising the step of excluding some of the keywords from core keywords for the first domain,
How to extract domain core keywords. In computing systems,
a network interface for receiving a plurality of documents;
Memory into which the domain key keyword extraction program is loaded; and
Including one or more processors executing the domain key keyword extraction program,
The domain core keyword extraction program is,
Instructions for obtaining first domain reference keywords extracted from a document set of the first domain; (instruction);
Instructions for obtaining a first embedding vector for a target document using a previously learned embedding model;
Instructions for obtaining second embedding vectors for some of the keywords extracted from the target document using the embedding model;
Instructions for obtaining third embedding vectors for each of the first domain reference keywords using the embedding model; and
Among the second and third embedding vectors, an instruction for outputting a keyword of an embedding vector having a similarity greater than or equal to a reference value with the first embedding vector as a key keyword for the first domain of the target document; includes; doing,
Domain core keyword extraction system. According to clause 7,
The target document is not included in the document set of the first domain,
Domain core keyword extraction system. According to clause 7,
The instructions output as key keywords for the first domain are:
Instructions for outputting keywords of some of the third embedding vectors as core keywords for the first domain of the target document; including,
Domain core keyword extraction system. According to clause 7,
The domain core keyword extraction program is,
Among the core keywords for the first domain, keywords that have a similarity of less than a standard value to the first embedding vector and have a similarity of more than a standard value to core keywords of other first domains are excluded from the core keywords of the first domain. Further including instructions;
Domain core keyword extraction system. According to clause 7,
Keywords extracted from the target document are:
including single word keywords, double word keywords and triple word keywords,
Domain core keyword extraction system. According to clause 7,
The domain core keyword extraction program is,
Based on the determinant value of each vector of the core keywords for the first domain, which is calculated as a result of performing a decision point process operation for each vector of the core keywords for the first domain, the core keywords for the first domain Further comprising instructions for excluding some of the keywords from core keywords for the first domain,
Domain core keyword extraction system.

Images