KR20220081009A - Keyword extraction apparatus, control method thereof and keyword extraction program - Google Patents

Abstract

The present invention relates to an apparatus and a control method for extracting a main keyword from a document composed of text data. More specifically, the present invention extracts a keyword corresponding to a predetermined morpheme pattern from a stored extraction target document as a candidate, and performs first filtering on the extracted keyword candidate in consideration of at least one of grammaticality and frequency of appearance, A first duplicate removal is performed in consideration of whether a character string is common among the first filtered keyword candidates, a rank on the extraction target document is calculated for each of the first removed duplicate keyword candidates, and the calculated rank is A second filtering is performed on the first de-duplicated keyword candidate based on the second filtering, a second de-duplication is removed in consideration of semantic duplication between the second filtered keyword candidates, and a second de-duplication is extracted from the second de-duplicated keyword candidate. Keywords are extracted by selecting keywords.

Description

Main keyword extraction device, its control method, and main keyword extraction program

The present invention relates to an apparatus, a control method, and a program for extracting a main keyword, and more particularly, to an apparatus and control for extracting a keyword that can be represented from text data consisting of a plurality of words such as a document and providing it to a user It relates to methods and programs.

Keyword extraction refers to extracting major keywords that can be represented from text data including a plurality of words, such as a document. The extracted keyword can be used to easily understand the contents of the document or to perform a search. In some cases, one keyword representing a document is extracted, but it is common to extract and provide 2 to 5 keywords with high importance so that the contents of the document can be easily understood. In the keyword extraction technique, it is important to accurately determine the importance in order to provide a keyword with high importance, but it is also important to minimize overlap between a plurality of main keywords.

Existing technologies are attempting to extract keywords through various approaches. For example, techniques such as the 'two-step method' of generating candidates using heuristic definitions and the 'Graph-based ranking method' of expressing the relationship between candidates in a graph in a document are commonly used keyword extraction techniques.

With the recent development of speech recognition technology, various applications for transcribing a human voice into text exist. In order to schematically check the contents of the recorded voice, techniques for extracting main keywords from the transcribed text are widely used. However, due to the characteristics of speech recognition technology, misrecognized characters inevitably exist, and since there are multiple speakers, it frequently occurs that words with the same meaning are expressed differently. For example, person A can be expressed as 'mobile phone', but person B can be expressed as 'cell phone'. 'Cell phone' and 'cell phone' are different words by themselves, but the implied meaning (the intended meaning) will be effectively the same. Only when these words are treated as a single word in the keyword extraction process, the accuracy of the extracted keywords can be increased and duplication between keywords can be minimized.

Existing techniques for extracting major keywords from a document have a problem in that they do not properly reflect the characteristics of the speech recognition technique as described above. For example, there may be a problem in that 'mobile phone' and 'cell phone' are recognized as different keywords and provided as main keywords.

Accordingly, there is a need for research on a keyword extraction technology optimized from a document in which a recognized voice is transcribed.

SUMMARY OF THE INVENTION An object of the present invention is to provide an apparatus, a control method, and a program for extracting main keywords from text data such as a document.

Another object to be solved by the present invention is to provide an apparatus for extracting keywords, a control method, and a program optimized for an audio transcription document.

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

According to one aspect of the present invention in order to solve the above or other problems, extracting a keyword corresponding to a predetermined morpheme pattern from the extraction target document as a candidate; performing first filtering on the keyword candidate in consideration of at least one of grammaticality and frequency of appearance; calculating a rank on the extraction target document for the keyword candidate; performing second filtering on the keyword candidates based on the calculated rank; and selecting a keyword to be extracted.

The method may further include performing the first duplicate removal in consideration of whether a character string is common among the keyword candidates.

The performing of the first duplicate removal may include: clustering into a plurality of groups in consideration of whether a character string is common among the keyword candidates; and removing the remaining keyword candidates except for the representative keyword of each of the plurality of clustered groups.

The clustering may include: calculating a Longest Common Subsequence (LCS) between the keyword candidates; and grouping keyword candidates having the calculated LCS length equal to or greater than a predetermined value into the same group.

The performing of the first filtering may include calculating the importance of the keyword candidate in consideration of at least one of grammaticality and frequency of appearance; and filtering the calculated keyword candidates having low importance.

The calculating of the importance may include: receiving a plurality of news contents in consideration of the grammar; learning a grammatical language model based on the received news content; and calculating the importance level based on the learned grammatical language model.

Calculating the importance may include: in considering the frequency of appearance, learning an informational language model based on the extraction target document; and calculating the importance level based on the learned informational language model.

The method may further include removing a second duplication in consideration of semantic duplication between the keyword candidates.

The calculating of the rank may include calculating the rank in consideration of at least one of a count value of a predetermined keyword and a count value of a sentence including the predetermined keyword on the extraction target document.

The performing of the second filtering may be performed in such a way that a keyword having a low calculated rank is removed from among the keyword candidates.

According to another aspect of the present invention to solve the above or other problems, a memory for storing an extraction target document; and a control unit for extracting a keyword from the stored extraction target document, wherein the control unit extracts a keyword corresponding to a predetermined morpheme pattern from the extraction target document as a candidate, and for the keyword candidate, one of grammatical properties and frequency of appearance In consideration of the above, the first filtering is performed, the rank on the extraction target document is calculated for the keyword candidate, the second filtering is performed on the keyword candidate based on the calculated rank, and the keyword to be extracted is selected. It provides a keyword extraction device, characterized in that the selection.

The controller may perform the first duplicate removal in consideration of whether a character string is common among the keyword candidates.

When the control unit performs the first duplicate removal, it is clustered into a plurality of groups in consideration of whether the character string is common among the keyword candidates, and the remaining keyword candidates except for the representative keyword of each of the clustered groups are removed. can

In the clustering, the controller may calculate a Longest Common Subsequence (LCS) between the keyword candidates, and group keyword candidates having the calculated LCS length equal to or greater than a predetermined value into the same group.

When the control unit performs the first filtering, the importance of the keyword candidate is calculated in consideration of at least one of grammaticality and frequency of appearance, and the calculated keyword candidates having low importance may be filtered.

When the control unit considers the grammatical property, it is possible to receive a plurality of news contents, learn a grammatical language model based on the received news contents, and calculate an importance based on the learned grammatical language model. have.

When the control unit considers the frequency of appearance, an informational language model may be learned based on the extraction target document, and an importance may be calculated based on the learned informational language model.

The controller may perform the second duplication removal in consideration of semantic duplication between the keyword candidates.

When calculating the rank, the controller may calculate the rank in consideration of at least one of a count value of a predetermined keyword and a count value of a sentence including the predetermined keyword on the extraction target document.

When the control unit performs the second filtering, it may be performed in such a way that a keyword having a low calculated rank is removed from among the keyword candidates.

The main keyword extraction apparatus according to the present invention, the control method thereof, and the effects of the program will be described as follows.

According to at least one of the embodiments of the present invention, there is an advantage in that it is possible to extract a main keyword for representing a document with high accuracy while simultaneously minimizing duplication between extracted keywords.

In addition, according to at least one of the embodiments of the present invention, there is an advantage in that it is possible to minimize overlap between keywords extracted from a document recognizing the voices of several speakers.

Further scope of applicability of the present invention will become apparent from the following detailed description. However, it should be understood that the detailed description and specific embodiments such as preferred embodiments of the present invention are given by way of example only, since various changes and modifications within the spirit and scope of the present invention may be clearly understood by those skilled in the art.

1 is a diagram illustrating a typical application example of a general speech recognition technology.
2 is a diagram showing a block diagram of the main keyword extraction apparatus 100 according to an embodiment of the present invention.
3 is a diagram illustrating a control flowchart of the keyword extraction apparatus 100 according to an embodiment of the present invention.
4 is an example of outputting keywords extracted according to an embodiment of the present invention.
5 is a diagram illustrating a detailed flowchart of keyword extraction according to an embodiment of the present invention.
6 is a diagram illustrating a learning conceptual diagram of two types of language models according to an embodiment of the present invention.
7 is a diagram illustrating a control flowchart for removing duplicate keywords based on a common character string (step S504) according to an embodiment of the present invention.
8 shows an example of a concept of removing duplicate keywords based on a common character string.
9 shows an example of calculating a second weight according to an embodiment of the present invention.
10 is a diagram illustrating a conceptual diagram of removing keyword duplication in consideration of semantic duplication according to an embodiment of the present invention.
11 is a diagram illustrating a conceptual diagram of filtering nonsense words according to an embodiment of the present invention.

Hereinafter, the embodiments disclosed in the present specification will be described in detail with reference to the accompanying drawings, but the same or similar components are assigned the same reference numbers regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "part" for components used in the following description are given or mixed in consideration of only the ease of writing the specification, and do not have distinct meanings or roles by themselves. In addition, in describing the embodiments disclosed in the present specification, if it is determined that detailed descriptions of related known technologies may obscure the gist of the embodiments disclosed in this specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in the present specification, and the technical idea disclosed herein is not limited by the accompanying drawings, and all changes included in the spirit and scope of the present invention , should be understood to include equivalents or substitutes.

Terms including an ordinal number such as 1st, 2nd, etc. may be used to describe various elements, but the elements are not limited by the terms. The above terms are used only for the purpose of distinguishing one component from another.

When an element is referred to as being “connected” or “connected” to another element, it is understood that it may be directly connected or connected to the other element, but other elements may exist in between. it should be On the other hand, when it is said that a certain element is "directly connected" or "directly connected" to another element, it should be understood that the other element does not exist in the middle.

The singular expression includes the plural expression unless the context clearly dictates otherwise.

In the present application, terms such as “comprises” or “have” are intended to designate that a feature, number, step, operation, component, part, or combination thereof described in the specification exists, but one or more other features It should be understood that this does not preclude the existence or addition of numbers, steps, operations, components, parts, or combinations thereof.

Although the embodiments of the present invention described below relate to keyword extraction techniques optimized for voice transcription documents, it is obvious that the present invention is not limited to voice transcription documents and can be extended and applied to general documents. In addition, the keyword extracted in the present invention may be a single word, but may also mean a combination of two or more words (hereinafter, a combination of two or more words is expressed as a word sequence). For example, the keyword extracted according to an embodiment of the present invention may be a combination of two words, such as 'search result', or a combination of three words, such as 'underwater penetration possibility'.

1 is a diagram illustrating a typical application example of a general speech recognition technology.

When a voice input is received from the plurality of speakers 111-1 to 111-3, the voice recognition apparatus 112 may recognize the input voice and generate the voice transcription document 200 transcribed into text. For example, when it is desired to make a meeting record with a recording file of a multi-party meeting, it may be converted into a voice transcription document 200 through the voice recognition device 112 . Furthermore, if the voice recognition apparatus 112 identifies a speaker during voice recognition, as shown in FIG. 1 , the voice transcribed into the voice transcription document 200 may be divided and displayed 201 for each speaker. This is because, in the case of minutes, it can be important to distinguish the speakers.

In the case of a general document that is not an audio transcription document, terms having a specific meaning are unified into one. However, in the case of voice, terms having the same meaning are often expressed in various ways. This is because there are many cases where one person expresses it in various ways, but there are many cases where each person uses their own expression. For example, a person named A may be expressed as a 'car', but a person named B may be expressed as a 'car' or a 'vehicle'. If keywords are extracted without reflecting various expressions in this way, the reliability of the extracted results is inevitably lowered. Accordingly, in an embodiment of the present invention described below, a method for extracting a main keyword with higher accuracy is proposed.

2 is a diagram illustrating a block diagram of a main keyword extraction apparatus 100 (hereinafter referred to as a keyword extraction apparatus) according to an embodiment of the present invention.

According to an embodiment of the present invention, the keyword extraction device 100 is a mobile phone, a cellular phone, a smart phone, a personal computer, a laptop, a notebook computer, a netbook or tablet, a personal digital assistant (PDA), a digital camera, Game console, MP3 player, personal multimedia player (PMP), electronic book (E-Book), navigation, disk player, set-top box, home appliance, communication device, display device, or other electronic device It may be built into or interact with the device. In addition, the keyword extraction device 100 may be embedded in or interact with a smart home appliance, an intelligent vehicle, an autonomous driving device, a smart home environment, a smart building environment, a smart office environment, a smart electronic security system, or the like.

The keyword extraction apparatus 100 may include a microphone 101 , a communication unit 102 , a control unit 103 , an output unit 104 , a memory 105 , and a power supply unit 106 . The components shown in FIG. 2 are not essential in implementing the keyword extraction apparatus 100, so the keyword extraction apparatus 100 described in this specification is more or less than the components listed above. can have

The microphone 101 processes an external sound signal as electrical voice data. The processed voice data may be variously utilized according to a function (or a running application program) being performed by the keyword extraction apparatus 100 . Meanwhile, various noise removal algorithms for removing noise generated in the process of receiving an external sound signal may be implemented in the microphone 101 .

The control unit 103 generally controls the overall operation of the keyword extraction apparatus 100 . The controller 103 may provide or process appropriate information or functions to the user by processing signals, data, information, etc. input or output through the above-described components.

In particular, the controller 103 according to an embodiment of the present invention may recognize a voice from an audio signal input through the microphone 101 and extract a main keyword from the recognized voice. Alternatively, the main keyword may be extracted from the document to be extracted stored in the memory 105 . A specific extraction method of the main keyword will be described later in detail.

In addition, the memory 105 stores data supporting various functions of the keyword extraction apparatus 100 . The memory 105 may store data and commands for the operation of the keyword extraction apparatus 100 . In addition, the memory 105 may store an 'extraction target document' to be described later.

The communication unit 102, one or more modules that enable wired/wireless communication between the keyword extraction device 100 and a wired/wireless communication system, between the keyword extraction device 100 and a mobile terminal, or between the keyword extraction device 100 and an external server may include In addition, the communication unit 102 may include one or more modules for connecting the keyword extraction device 100 to one or more networks. The voice recognition result and/or the keyword extraction result may be transmitted to another mobile terminal or an external server through the communication unit 102 .

The output unit 104 is configured to output the voice transcription result and/or the extracted keyword, and may include a display. The voice recognition result may be output through the display, but may also be output in the form of text data. When output as text data, the output unit 104 may store the text data in the above-described memory 105 and provide a voice transcription result in a form in which the user confirms the stored text data.

Meanwhile, the voice transcription result and/or the extracted keyword may be provided through a display of a mobile terminal connected through the communication unit 102 .

The power supply unit 106 receives external power and internal power under the control of the control unit 103 to supply power to each component included in the keyword extraction device 100 .

3 is a diagram illustrating a control flowchart of the keyword extraction apparatus 100 according to an embodiment of the present invention.

In step S1100, the keyword extraction apparatus 100 receives a voice input. Although the keyword extraction device 100 has its own microphone 101 and may receive a voice input through it, the voice input received through the microphone of the external mobile terminal may be received in the form of being transmitted through the communication unit 102 . it might be

In step S1101, the controller 103 of the keyword extraction apparatus 100 recognizes a voice from the received voice input. The recognized voice may be text data composed of a plurality of words. And in step S1102, the control unit 103 of the keyword extraction apparatus 100 extracts a main keyword from the recognized voice (ie, text data). In step S1103 , the control unit 103 of the keyword extraction apparatus 100 may output the extracted main keyword through the output unit 104 . An example in which the extracted main keyword is provided will be described below with reference to FIG. 4 .

4 is an example of outputting keywords extracted according to an embodiment of the present invention. The extracted keyword in this example may be output from the keyword extraction device 100 itself, but the result of voice recognition and/or keyword extraction performed in the keyword extraction device 100 is transmitted to an external mobile terminal, and the mobile It may be displayed on the terminal.

The output screen in the illustrated example includes a voice recognition output area 402 and an extracted keyword area 401 . When a voice input in the form of an audio signal is received through the microphone 101 , the controller 103 may recognize a voice from the audio signal and output the recognized result to the voice recognition output area 402 . And when a main keyword is extracted according to an embodiment of the present invention, at least one main keyword may be output on the extracted keyword area 401 and provided to the user.

Hereinafter, a specific method (step S1102) of extracting a main keyword from text data will be described.

5 is a diagram illustrating a detailed flowchart of keyword extraction according to an embodiment of the present invention.

In step S501, the control unit 103 analyzes the morpheme of text data, which is a target for extracting a keyword. Here, the text data may consist of a plurality of words, and may be a result of voice recognition in step S1101. Hereinafter, the keyword extraction target text data will be simply referred to as an 'extraction target document'.

In step S502, the controller 103 extracts word(s) corresponding to a predetermined morpheme pattern as keyword candidates. For example, even if the word(s) corresponding to the morpheme pattern of 'NNG XSN NNG' is grown, it can be extracted as a candidate. Table 1 below is a part of the morpheme code. That is, in the present invention, pattern(s) of morphemes generally used as keywords are stored in advance, and a keyword having the corresponding morpheme pattern is selected as a candidate.

morpheme code Explanation morpheme code Explanation NNG common noun IC interjection NNP proper noun JKS nominative investigation NNB dependent noun JKC supplemental investigation NR Investigation JKG tubular case study NP pronoun JKO objective investigation VV verb JKB sub-firing investigation VA adjective JKV assault investigation VX auxiliary verb JKQ Quotation Investigation VCP affirmative adjective JX assistant VCN negative designator JC connection investigation MM detective EP fresh fish mother MAG common adverb EF terminating ending MAJ conjunctive adverb SN number ...

In one embodiment of the present invention, it is proposed to learn the morpheme pattern using highly reliable text data. In particular, in one embodiment of the present invention, it is proposed to use a news article as text data with high reliability. Furthermore, only the headline excluding the main body of the news article may be used. This is because, in general, keywords with high importance in the news body are placed in news titles, so they are optimal for learning morpheme patterns for extracting keyword candidates.

The keyword extraction apparatus 100 according to an embodiment of the present invention may convert the learned morpheme pattern into a database and store it in the memory 105 .

Subsequently, in step S503, the controller 103 may perform the first filtering on the extracted keyword candidate in consideration of one or more of grammaticality and frequency of appearance.

Filtering in consideration of grammatical properties means removing word(s) that do not fit the grammar from among keyword candidates.

The frequency of appearance may mean the frequency of appearance in the 'extraction target document'. Word(s) that do not appear frequently may be filtered to be removed from keyword candidates by determining that their importance is not high.

The first filtering as described above may serve to roughly filter out keywords that are not suitable as keyword candidates in a simple way.

In particular, according to an embodiment of the present invention, in step S503, it is proposed to use a language model (LM). The language model means a probability distribution for at least two consecutive word sequences (word sequences). In other words, it refers to a model that assigns probabilities to word sequences to model the phenomenon of language. The language model calculates the probability that a particular sequence of words will be used. For example, it can be used to calculate the probability that the word sequence 'coronavirus' is used.

과 단어

로 이루어지는 단어 시퀀스에 대해 아래 수학식 1과 같이 표현될 수 있다.A language model is, as an example, a word

and words

It can be expressed as in Equation 1 below for a word sequence consisting of .

은 분석소스 상에서 단어

를 카운트 한 개수,

는 분석소스 상에서 단어 시퀀스

를 카운트 한 개수를 의미한다.here

is a word in the analysis source

the number of counts,

is a word sequence on the analysis source

means the number of counts.

For example, when keywords extracted from an analysis source (eg news content) were counted, the word 'MERS' was 1,346,709, the word sequence 'MERS incubation period' was 1,038, and the word 'MERS symptom' Assume there are 8,018 sequences. When calculating the probabilities for each of the 'MERS incubation period' and the 'MERS symptom' through the language model, it is calculated as shown in Equations 2 and 3 below.

That is, the language model means the ratio of using the word 'A' as a sequence of words 'A+B' compared to the case where the word 'A' is used alone.

In particular, in applying a language model according to an embodiment of the present invention, it is proposed to use a combination of two types of language models. One of the two types is a 'grammar language model' that learns from a reliable source in order to focus on grammatical characteristics considering grammar, and the other is a 'grammar language model' for reflecting the contents of the 'extraction target document'. informational language model. Two types of language models will be described with reference to FIG. 6 .

6 is a diagram illustrating a learning conceptual diagram of two types of language models according to an embodiment of the present invention.

Since the content of the 'extraction target document' itself may be a text transcription of the user's voice input, the grammatical reliability may not be high. In addition, even in the case of a document written in a general manner, high reliability cannot be expected in terms of grammar if the document has no credibility (eg, simple meeting minutes or simple memos). However, in the case of news content, high reliability can be expected, and further, higher reliability can be expected from the news body.

Therefore, in one embodiment of the present invention, in order to analyze the grammar of the word sequence, it is proposed to utilize the 'grammar language model 603' learned based on the news content 601 (that is, the language model The source of analysis is news content). Furthermore, according to an embodiment of the present invention, the 'grammar language model 603' can be learned only from the news body among the news content 601 . When the probability of a specific word sequence is calculated through the learned 'grammar language model 603', the calculated probability may have high grammatical reliability.

To this end, the keyword extraction apparatus 100 according to an embodiment of the present invention extracts ngrams (unigram, bigram, etc.) from a plurality of news contents 601, and based on the extracted contents, a 'grammar language model 603' )' can be learned.

The 'information language model 604' is a model that reflects the contents in the 'extraction target document 602'. This is to calculate a probability of how often a specific word sequence is used in the 'extraction target document 602'. That is, the analysis source of the 'information language model 604' will be the 'extraction target document 602'.

To this end, the keyword extraction apparatus 100 according to an embodiment of the present invention may extract ngrams from the 'extraction target document 602' and learn the 'information language model 604' therefrom. Simply, the 'information language model 604' refers to a model learned based on the 'extraction target document 602'.

The keyword extraction apparatus 100 according to an embodiment of the present invention performs the first filtering (step S503) in consideration of both the 'grammar language model 603' and the 'information language model 604'. can When both types of models are considered together, filtering that considers both grammatical and appearance frequency may be performed. In one embodiment of the present invention, in order to use two types of models together, it is proposed to use 'pointwise KL (kullback) divergence'. However, the use of 'pointwise KL (kullback) divergence' is only an example, and the present invention is not limited to this method.

Returning to FIG. 5 , in step S504 , the control unit 103 removes the keyword duplication based on a character string common to the first filtering performance result. Hereinafter, a control method for removing duplicate keywords based on a common character string will be described in more detail with reference to FIGS. 7 and 8 .

7 is a diagram illustrating a control flowchart for removing duplicate keywords based on a common character string (step S504) according to an embodiment of the present invention. 8 shows an example of a concept of removing duplicate keywords based on a common character string. It will be described with reference to FIGS. 7 and 8 together.

Clustering is performed based on a common character string on the result of performing the first filtering (S504-1). The common character string means a character string that two keywords to be judged have in common. That is, it is to group keywords having a common character string into one group.

More specifically, in the present invention, a Longest Common Subsequence (LCS) algorithm may be used to calculate whether there is a common character string and the length of the common character string. The control unit 103 may calculate an LCS among the first filtered candidates, and if the calculated LCS length is equal to or greater than a predetermined value, it may be grouped into a single group. In the present invention, for efficient iteration, the minimum LCS value may be increased from 3 to 1 each time the number of clustering increases.

In the example shown in FIG. 8 , keyword candidates 801 including 'Christmas', 'Christmas holiday', 'Christmas event' and the like are extracted. The keyword candidate 801 may be a result of the filtering performed in step S503. Since the three keyword candidates 'Christmas', 'Christmas holiday', and 'Christmas event' commonly include the string 'Christmas', they may be grouped into one first group 802-1. Similarly, keywords including the string 'Android' in common may be grouped into the second group 802-2.

Subsequently, the first duplicate removal may be performed (step S504) in a manner of removing the remaining keywords except for the representative keyword of each grouped group (step S504-2). A first deduplication is performed so that a representative keyword 'Christmas' is selected from the first group 802-1 shown in FIG. 8 and the remaining 'Christmas holidays' and 'Christmas events' are removed.

The keyword extraction apparatus 100 according to an embodiment of the present invention may consider the above-mentioned 'grammar language model 603' and 'information language model 604' together in order to select a representative keyword within a group. have. For example, through 'pointwise KL (kullback) divergence', a probability score from two language models may be calculated, and a keyword with the highest probability score may be selected as a representative keyword. That is, the representative keyword may be a word frequently used on the 'extraction target document' while having high grammatical reliability.

Returning to FIG. 5 again, it is possible to calculate a rank for each of the keyword candidates for which duplication has been removed in S504 (step S505). Here, the rank is determined based on a weight of the corresponding keyword on the 'extraction target document', and may mean the importance of the keyword on the 'extraction target document'. That is, the weight means an importance score for determining a relative ranking among keyword candidates.

The keyword extraction apparatus 100 according to an embodiment of the present invention proposes to obtain a comprehensive weight by combining weights calculated by various approaches.

* First weight considering the frequency of appearance (NPweight)

The first weight is simply calculated through a relative comparison of the frequency of appearance, and can be calculated through Equation 4 below.

는 '추출대상 문서' 상에서 전체 키워드에 대한 카운트 합을 의미한다.keyword _i means a keyword to be weighted, n is the total number of keyword candidates (here, all keyword candidates for determining the ranking);

denotes the sum of counts for all keywords on the 'extraction target document'.

* Second weight considering the importance of sentences and keywords (TF-IDF _weight )

TF-IDF (Term Frequency - Inverse Document Frequency) is a weight used in information retrieval and text mining, and is a statistical number indicating how important a word is within a specific document when there is a document group consisting of several documents. It can be used for purposes such as extracting key words from documents, determining the ranking of search results in a search engine, or finding the degree of similarity between documents.

9 shows an example of calculating a second weight according to an embodiment of the present invention.

When the 'extraction target document' consists of several sentences, the second weight is a weight for considering both the count value of the keyword and the count value of the sentence including the keyword on the entire 'extraction target document'.

The second weight may be calculated as in Equation 5 below.

keyword _i is a keyword for weight calculation, sentence _i is a sentence containing a keyword for weight calculation, TF (Term Frequency) is a count value of a specific keyword (keyword for weight calculation) on the 'extraction target document', and IDF (Inverse) Document Frequency) means a value obtained by dividing the count value of sentences containing a specific keyword by the total number of sentences and taking the log.

Referring to the illustrated example, TF and IDF may be calculated for each keyword candidate, and a second weight may be obtained by multiplying the TF and IDF. Taking the keyword 'result' in FIG. 9 as an example, it can be confirmed that TF is 3 because 3 is obtained when the keyword 'result' is counted from the whole. And since the number of sentences including the keyword 'result' is 2 and the total number of sentences is 3, IDF is log(2/3).

* Third weight considering the similarity between sentences (Sentence _weight )

The third weight is a weight value for a sentence including a corresponding keyword, and the similarity between sentences may be reflected by using a Textrank algorithm. The third weight may be calculated by Equation 6 below. It means the importance of the sentence in the 'extraction target document'.

The TextRank algorithm is a well-known algorithm and a detailed description thereof will be omitted.

In order to comprehensively consider the first to third weights, the total weights may be determined as in Equation 7 below.

Here, the values of a, b, and c are constants for experimentally selecting and setting a value with the highest accuracy. For example, a=0.5, b=0.25 and c=0.25 can be set.

That is, in order for the keyword extraction apparatus 100 to calculate the rank in step S505, the comprehensive weight in Equation 7 may be used.

Returning to the flowchart of FIG. 5 , in step S506, second filtering is performed based on the calculated rank. This is to exclude keyword candidates with a low rank. As a result, according to the second filtering, keyword candidates with low importance may be excluded.

When the number of keywords to be finally selected is N, the keyword extraction apparatus 100 according to an embodiment of the present invention may perform second filtering to leave the keywords of N*(N-1) and exclude the rest. This is because semantic duplication must be removed for the remaining keyword candidates, and the amount of computation required to consider semantic duplication is large. That is, in order to efficiently manage the amount of computation, keyword candidates with low importance are removed in the second filtering step.

Subsequently, in step S507, the duplicate keywords are removed in consideration of the semantic overlap between the remaining keyword candidates. That is, with respect to duplicate keywords, only representative keywords are left and removed.

As described above, when there are a plurality of speakers in a transcription document, it is frequently honored to express the same meaning differently. There will be a need to remove semantic redundancy not only in the phonetic transcription document, but also in the general document. Therefore, in one embodiment of the present invention, it is determined whether they are semantically identical/similar (duplicate), and it is proposed to exclude overlapping keywords from candidates according to the determination result.

For this, an embodiment of the present invention proposes to use an embedding vector (eg, sent2vec) learned from a large number of documents. Since it is an embedding vector learned from a large number of documents, an embedding vector calculated using the embedding vector may contain the meaning of a keyword. After the embedding vectors are calculated for each keyword candidate, if the similarity between the embedding vectors is high, it may be determined that they are semantically overlapped.

For example, the similarity between two keyword candidates may be calculated through cosine similarity.

10 is a diagram illustrating a conceptual diagram of removing keyword duplication in consideration of semantic duplication according to an embodiment of the present invention.

An embedding vector (1002-1 to 1002-3) is calculated for each of the three keywords (1001-1 to 1001-3). The calculated embedding vectors 1002-1 to 1002-3 may be vectors containing the meaning of each keyword. Then, the similarity between the calculated embedding vectors 1002-1 to 1002-3 is calculated, and when the calculated similarity is equal to or greater than a preset value, it can be determined that the meanings of the two keywords are the same. For example, cosine similarity (1003-1 to 1003-3) is calculated for two keywords to be compared, and when the calculated similarity (1003-1 to 1003-3) is 0.6 or more, it can be determined that they have the same meaning.

In the example of FIG. 10 , the first embedding vector 1002-1 for 'laundry (1001-1)' and the second embedding vector 1002-2 for 'Ssangyong Motor (1001-2)' are first When the cosine similarity (1003-1) was calculated, a value of 0.57 was calculated. Since the first cosine similarity (1003-1) is a value lower than 0.6, the two keywords 'layoff (1001-1)' and 'Ssangyong Motor (1001-2)' are keywords with different meanings (that is, semantic It is not a duplicate).

Similarly, if the same process is repeated for 'Ssangyong Motor 1001-2' and 'Ssangyong Motor 1001-3', a third cosine similarity 1003-3 of 0.86 is calculated. Since this is a higher number than 0.6, it is determined that 'Ssangyong Motor (1001-2)' and 'Ssangyong Motor (1001-3)' are keywords with the same meaning, and either one can be removed.

Finally, the keyword extraction apparatus 100 may select all or some of the remaining keyword candidates as the main keywords for the 'extraction target document' after the steps S501 to S507.

On the other hand, the keyword extraction apparatus 100 according to an embodiment of the present invention may perform the forbidden word filtering step together with the steps S501 to S507. Kinsoku filtering means excluding keywords that are difficult to be selected as keywords as well as profanity or non-grammatical keywords.

11 is a diagram illustrating a conceptual diagram of filtering nonsense words according to an embodiment of the present invention.

For this purpose, the forbidden word dictionary 1101 according to an embodiment of the present invention may include at least one of a morpheme analyzer dictionary 1101-1, an abstract expression dictionary 1101-2, and a user-specified forbidden word dictionary 1101-3. have. The keyword extracting apparatus 100 according to an embodiment of the present invention determines whether there is a keyword included in the forbidden word dictionary 1101 among the keyword candidates, and if there is such a keyword, performs non-restrictive word filtering to remove it from the keyword candidate. You can do it.

The morpheme analyzer dictionary 1101-1 is a database that stores expressions having adverbs or ambiguities. For example, expressions such as 'on the other hand' or 'each' can be stored. Since such an expression is not suitable for extraction as a keyword, it may be desirable to exclude it from keyword candidates.

The abstract expression dictionary 1101-2 is a database in which abstract expressions are arranged. For example, expressions such as 'this day', 'this time', 'next year', 'remaining' are stored. These expressions can be referred to in 'Sejong Corpus'.

In addition, expressions inappropriate to be selected as keywords may be stored in the user-designated forbidden word dictionary 1101-3. For example, expressions such as 'but' or 'now' may be used.

This forbidden word filtering step may be performed before or after the main keyword extraction step (S1102), but preferably it may be performed on the keyword candidates extracted in the step S502.

Furthermore, in the present invention, it is proposed to minimize duplication between keywords by treating keywords extracted from the same location as one.

12 is a diagram illustrating a conceptual diagram of a deduplication method in consideration of an extracted location according to an embodiment of the present invention.

Referring to the illustrated example, a keyword candidate is extracted from a predetermined sentence 1200 of the 'extraction target document'. In this case, the extracted keyword candidates 1201-1 to 1201-3 may be words corresponding to the preset morpheme patterns of Table 2 in the predetermined sentence 1200 . The first to third keyword candidates 1201-1 to 1201-3 were extracted from the word sequence 'underwater penetration possibility' included in the predetermined sentence 1200 .

preset morpheme pattern Extracted keyword candidates NNG NNG underwater penetration NNG NNG NNG Possibility of underwater penetration NNG NNG Penetration potential

When a plurality of word sequences including a word at a specific position on a keyword candidate exist, the plurality of word sequences are highly likely to be overlapping keywords, and thus it is proposed to treat them as one. The first to third keyword candidates 1201-1 to 1201-3 extracted in the example of FIG. 12 are a plurality of word sequences including the word 'penetration' in a specific position, and each of them needs to be viewed as an individual keyword candidate. there will be no For example, only the second keyword 1201-2 (possibility of underwater penetration) may be left in the keyword candidates, and the remaining first and third keywords 1201-1 and 1201-3 may be excluded from the keyword candidates.

Although the embodiments of the main keyword extraction apparatus and the control method thereof have been described above according to the present invention, these are described as at least one embodiment, and the technical spirit of the present invention and its configuration and operation are not limited thereby. That is, the scope of the technical idea of the present invention is not limited / limited by the drawings or the description with reference to the drawings. In addition, the concepts and embodiments of the present invention presented in the present invention can be used by those of ordinary skill in the art as a basis for modifying or designing other structures in order to perform the same purpose of the present invention. , an equivalent structure modified or changed by a person of ordinary skill in the art to which the present invention belongs is bound by the technical scope of the present invention described in the claims, and does not depart from the spirit or scope of the invention described in the claims Various changes, substitutions and changes are possible within the limits.

Claims (21)

In the control method of a device for extracting a keyword on an extraction target document,
extracting a keyword corresponding to a predetermined morpheme pattern from the extraction target document as a candidate;
performing first filtering on the keyword candidate in consideration of at least one of grammaticality and frequency of appearance;
calculating a rank on the extraction target document for the keyword candidate;
performing second filtering on the keyword candidates based on the calculated rank; and
characterized in that it comprises the step of selecting a keyword to be extracted,
Control method of keyword extraction device. The method of claim 1,
The method further comprising the step of performing a first duplicate removal in consideration of whether a character string is common among the keyword candidates,
Control method of keyword extraction device. The method of claim 2, wherein performing the first deduplication comprises:
clustering into a plurality of groups in consideration of whether a character string is common among the keyword candidates; and
It characterized in that it comprises the step of removing the remaining keyword candidates except for the representative keyword of each of the plurality of clustered groups,
Control method of keyword extraction device. The method of claim 3, wherein the clustering comprises:
calculating a Longest Common Subsequence (LCS) between the keyword candidates; and
and bundling keyword candidates whose length of the calculated LCS is greater than or equal to a predetermined value into the same group,
Control method of keyword extraction device. The method of claim 1, wherein performing the first filtering comprises:
calculating the importance of the keyword candidate in consideration of at least one of grammaticality and frequency of appearance; and
It characterized in that it comprises the step of filtering the calculated keyword candidates with low importance,
Control method of keyword extraction device. The method of claim 5, wherein the calculating of the importance comprises:
receiving a plurality of news contents;
learning a grammatical language model based on the received news content; and
It characterized in that it comprises the step of calculating the importance based on the learned grammatical language model,
Control method of keyword extraction device. The method of claim 5, wherein the calculating of the importance comprises:
learning an informational language model based on the extraction target document; and
It characterized in that it comprises the step of calculating the importance based on the learned informational language model,
Control method of keyword extraction device. The method of claim 1,
The method further comprising the step of removing a second duplication in consideration of semantic duplication between the keyword candidates,
Control method of keyword extraction device. The method of claim 1, wherein calculating the rank comprises:
characterized in that the rank is calculated in consideration of at least one of a count value of a predetermined keyword and a count value of a sentence including the predetermined keyword on the extraction target document;
Control method of keyword extraction device. The method of claim 9, wherein performing the second filtering comprises:
characterized in that it is made in such a way as to remove the keyword having the calculated low rank from among the keyword candidates,
Control method of keyword extraction device. a memory for storing the document to be extracted; and
Including a control unit for extracting a keyword from the stored extraction target document,
The control unit is
extracting a keyword corresponding to a predetermined morpheme pattern from the extraction target document as a candidate;
First filtering is performed on the keyword candidates in consideration of at least one of grammaticality and frequency of appearance,
calculating a rank on the extraction target document for the keyword candidate;
performing a second filtering on the keyword candidates based on the calculated rank;
Characterized in selecting a keyword to be extracted,
keyword extraction device. 12. The method of claim 11, wherein the control unit,
characterized in that the first duplicate removal is performed in consideration of whether the character string is common among the keyword candidates,
keyword extraction device. The method of claim 12, wherein when the control unit performs the first deduplication,
Clustering into a plurality of groups in consideration of whether the character string is common among the keyword candidates,
characterized in that the remaining keyword candidates except for the representative keyword of each of the clustered plurality of groups are removed,
keyword extraction device. The method of claim 13, wherein in the clustering by the control unit,
Calculating a Longest Common Subsequence (LCS) between the keyword candidates,
Characterized in grouping keyword candidates having the calculated LCS length equal to or greater than a predetermined value into the same group,
keyword extraction device. The method of claim 11, wherein when the control unit performs the first filtering,
Calculating the importance of the keyword candidate in consideration of at least one of grammaticality and frequency of appearance,
Characterized in filtering the calculated keyword candidates with low importance,
keyword extraction device. The method of claim 15, wherein the control unit considers the grammar,
receive input from a plurality of news content,
Learning a grammatical language model based on the received news content,
Characterized in calculating the importance based on the learned grammatical language model,
keyword extraction device. The method according to claim 15, wherein the control unit considers the frequency of appearance,
Learning an informational language model based on the extraction target document,
Characterized in calculating the importance based on the learned informational language model,
keyword extraction device. 12. The method of claim 11, wherein the control unit,
characterized in that the second duplication removal is performed in consideration of the semantic duplication between the keyword candidates,
keyword extraction device. 12. The method of claim 11, wherein when the control unit calculates the rank,
characterized in that the rank is calculated in consideration of at least one of a count value of a predetermined keyword and a count value of a sentence including the predetermined keyword on the extraction target document;
keyword extraction device. The method of claim 19, wherein when the control unit performs the second filtering,
characterized in that it is made in such a way as to remove the keyword having the calculated low rank from among the keyword candidates,
keyword extraction device. A computer program stored in a medium for executing the method of any one of claims 1 to 10 in combination with hardware.

Images