KR102168504B1 - Aparatus for coherence analyzing between each sentence in a text document and method thereof - Google Patents

Abstract

The present invention relates to an apparatus and a method for analyzing consistency between sentences in a text document. A text document in a natural language in a document unit is received as an input and decomposed into paragraphs and sentences, and the decomposed paragraphs and sentences are morpheme analyzed to provide text A first consistency analysis method that quantitatively calculates consistency by analyzing the keywords of the sentence excluding the first sentence of the paragraph and the keywords of the preceding sentence for the preprocessor outputting the document and the text document output through the morpheme analysis from the preprocessor , After extracting the feature of each sentence and expressing it as a vector, the second consistency analysis method, which calculates the similarity of each sentence and quantitatively calculates the consistency using this, and the machine learning method randomly generates inconsistent sentences. Each unit sentence maintains consistency in the context of the entire text document by using at least one consistency analysis method among the third consistency analysis methods that learn using a deep learning-based convolutional neural network and quantitatively calculate consistency through the combined results. By including a text consistency analysis unit that quantitatively analyzes whether or not it is, it is possible to effectively analyze the consistency of each document for quality evaluation and analysis of unstructured text data.

Description

A device for analyzing consistency between sentences in a text document and its method {APARATUS FOR COHERENCE ANALYZING BETWEEN EACH SENTENCE IN A TEXT DOCUMENT AND METHOD THEREOF}

The present invention relates to a consistency analysis apparatus and method for automatically evaluating the consistency between sentences in a text document in a natural language.

In general, coherence refers to the close order between sentences.

In other words, coherence is a broader concept than unity that describes the semantic connection between contents, and is used in a similar meaning to coherence.

In particular, consistency in text-based language is a concept that includes not only the repetition of important phrases, thesaurus, and pronouns between successive sentences, but also the closeness between the content and the content as it has completeness.

In this way, consistency is a principle applied syntactically between sentence and sentence, content (or paragraph) and content.

On the other hand, natural language processing is a field of artificial intelligence, a technology that has the purpose of computers to understand human language. Mainly, syntaxic analysis such as morphological analysis and parsing, semantic analysis such as named entity recognition, semantic role labeling, document summarization and correlation Includes discourse analysis, such as coreference resolution.

While the natural language processing problems described above are relatively well defined and are being actively studied, studies on consistency analysis have been insufficient to date.

Korean Patent Publication No. 10-2017-0030297 (published on March 17, 2017)

The present invention was devised to solve the above-described problem, and an object of the present invention is to quantitatively measure whether each unit sentence maintains consistency in the context of the entire text by decomposing it into sentences and paragraphs by receiving text in a document unit as an input. By doing so, it is to provide an apparatus and a method for analyzing consistency between sentences in a text document so that the consistency of each document can be effectively analyzed for quality evaluation and analysis of unstructured text data.

In order to achieve the above object, the first aspect of the present invention is to receive a text document in a natural language in a document unit as an input and decompose it into paragraphs and sentences, and morpheme analysis of the decomposed paragraphs and sentences to output as a text document. Pretreatment unit; And a first consistency analysis method of quantitatively calculating consistency by analyzing keywords of sentences other than the first sentence of the paragraph and keywords of the preceding sentence with respect to the text document output through morpheme analysis from the preprocessor, and the features of each sentence ( Feature) is extracted and expressed as a vector, and then the second consistency analysis method, which calculates the similarity of each sentence, and uses it to quantitatively calculate the consistency, randomly generates inconsistent sentences using a machine learning method, and creates a convolution based on deep learning. Text that quantitatively analyzes whether each unit sentence maintains the contextual consistency of the entire text document by using at least one of the third consistency analysis methods, which learns using a neural network and calculates consistency quantitatively through the combined results. A device for analyzing consistency between sentences in a text document including a consistency analysis unit is provided.

Here, it is preferable that the first consistency analysis method compares the morphemes of each sentence in order to determine how much adjacent sentences are semantically related, and calculates the consistency quantitatively using Equation 1 below.

(Equation 1)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence, and at least one cross-reference relationship to two adjacent sentences. If there is (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise it is defined as '0'.

Preferably, the second consistency analysis method includes word embedding and sentence embedding in which each sentence is expressed as a feature vector through dimension reduction and abstraction by learning context information of words and sentences of text. Embedding) is used to express the sentences constituting the text document as each sentence vector, calculate the similarity between each sentence vector through each sentence vector expressed, and then use the calculated similarity between each sentence vector to the corresponding text The semantic consistency of the document can be calculated quantitatively.

Preferably, the sentence embedding may be defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.

Preferably, the semantic consistency of the text document through the second consistency analysis method can be calculated by Equation 2 below by using the average of the similarity between all sentences constituting the text document.

(Equation 2)

라고 할 때 문장 간의 유사도를 계산한 것이다.Here, N is the total number of sentences in the document, and Sim(s _i , s _j ) is a text document consisting of N sentences using Cosine Similarity D=(s ₁ , s ₂ ,... The sentence vector corresponding to the sentence s _i for ,s _N }

When said, the degree of similarity between sentences was calculated.

Preferably, the third consistency analysis method performs the convolutional neural network learning using each sentence constituting the text document, the preceding sentence of each sentence, and the back sentence of each sentence in order to measure the consistency of the corresponding text document. However, from the result of morphological analysis of the input sentence for use as an input of the convolutional neural network, one morpheme is assumed as a word, and each sentence is expressed by converting it into a word vector using word embedding. It can be used as a sentence matrix that connects vectors of words that compose a sentence matrix (Sentence Matrix).

Preferably, as input of the convolutional neural network, a total of three sentences consisting of each sentence constituting the corresponding text document, the front sentence of each sentence, and the back sentence of each sentence are used, but N (the number of total sentences in the document Consistent document consisting of) sentences D=(s ₁ ,s ₂ ,... For ,s _N }, three sentences are defined as one set (q) to form the input data of the model, and the middle sentence among the three sentences in one set (q) is replaced with another random inconsistent sentence. Error learning data is generated, and'y _q =1' is set for a set of three sentences, which is one learning data, for sentences constituting a consistent document, and the middle sentence among the three sentences is For error training data generated by replacing with other inconsistent sentences, the model is trained by setting'y _q = 0', and the consistency can be quantitatively calculated through the combined results.

Preferably, a consistent document consisting of N sentences through the combined result of the convolutional neural network D={s ₁ ,s ₂ ,... , s _N} consistency (S _D) for a can be calculated quantitatively by the following equation 3.

(Equation 3)

인 경우이고, p는 상기 합성곱 신경망의 합습된 결과로서 3개의 문장 세트에 대한 일관성 확률이다.Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.

Preferably, according to the control of the text consistency analysis unit, a display unit for displaying a consistency score between each sentence quantitatively analyzed in a corresponding text document may be further included on the display screen so that the user can visually view the consistency score.

Preferably, according to the control of the text consistency analyzer, a storage unit for storing the consistency information data between each sentence quantitatively analyzed in the corresponding text document into a database (DB) for each text document or each sentence may be further included.

Preferably, a communication unit for transmitting the consistency information data between each sentence quantitatively analyzed in a corresponding text document to an external user terminal by wire or wireless under the control of the text consistency analysis unit may be further included.

A second aspect of the present invention is a method of analyzing the consistency between sentences in a text document using a device including a preprocessor and a text consistency analysis unit, comprising: (a) a text document in a document-level natural language through the preprocessor. Receiving an input and decomposing it into paragraphs and sentences, analyzing the decomposed paragraphs and sentences in morphemes, and outputting them as text documents; And (b) quantitatively calculating consistency by analyzing keywords of sentences other than the first sentence of the paragraph and keywords of the preceding sentence with respect to the text document that is morphologically analyzed and output in step (a) through the text consistency analysis unit. The first consistency analysis method, the second consistency analysis method, which extracts the features of each sentence and expresses it as a vector, then calculates the similarity of each sentence and uses it to quantitatively calculate the consistency. Using at least one of the third consistency analysis methods, which randomly generates and learns it using a deep learning-based convolutional neural network, and quantitatively calculates consistency through the combined results, each unit sentence is the context of the entire text document. It is to provide a method for analyzing consistency between sentences in a text document, comprising the step of quantitatively analyzing whether image consistency is maintained.

Preferably, in step (b), the first consistency analysis method compares the morphemes of each sentence in order to determine how semantically related adjacent sentences are, and calculates the consistency quantitatively using Equation 4 below. I can.

(Equation 4)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence, and at least one cross-reference relationship to two adjacent sentences. If there is (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise it is defined as '0'.

Preferably, in the step (b), the second coherence analysis method is a word embedding (Word) representing each sentence as a feature vector through dimension reduction and abstraction by learning the context information of words and sentences in the text. Embedding) and sentence embedding (Sentence Embedding) to express the sentences constituting the text document as each sentence vector, calculate the similarity between each sentence vector through each expressed sentence vector, and then calculated each sentence vector The semantic consistency of the text document can be calculated quantitatively by using the similarity between the two.

Preferably, the sentence embedding may be defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.

Preferably, the semantic consistency of the text document through the second consistency analysis method can be calculated by Equation 5 below using the average of the similarities between all sentences constituting the text document.

(Equation 5)

라고 할 때 문장 간의 유사도를 계산한 것이다.Here, N is the total number of sentences in the document, and Sim(s _i , s _j ) is a text document consisting of N sentences using Cosine Similarity D=(s ₁ , s ₂ ,... The sentence vector corresponding to the sentence s _i for ,s _N }

When said, the degree of similarity between sentences was calculated.

Preferably, in the step (b), the third consistency analysis method uses each sentence constituting the text document, the front sentence of each sentence, and the back sentence of each sentence in order to measure the consistency of the text document. Then, the convolutional neural network is trained, but from the result of morphological analysis of the input sentence for use as the input of the convolutional neural network, one morpheme is assumed as a word, and a word vector using word embedding After converting to, it can be used as a sentence matrix that connects vectors of words constituting each sentence.

Preferably, as input of the convolutional neural network, a total of three sentences consisting of each sentence constituting the corresponding text document, the front sentence of each sentence, and the back sentence of each sentence are used, but N (the number of total sentences in the document Consistent document consisting of) sentences D=(s ₁ ,s ₂ ,... For ,s _N }, three sentences are defined as one set (q) to form the input data of the model, and the middle sentence among the three sentences in one set (q) is replaced with another random inconsistent sentence. Error learning data is generated, and'y _q =1' is set for a set of three sentences, which is one learning data, for sentences constituting a consistent document, and the middle sentence among the three sentences is For error training data generated by replacing with other inconsistent sentences, the model is trained by setting'y _q = 0', and the consistency can be quantitatively calculated through the combined results.

Preferably, a consistent document consisting of N sentences through the combined result of the convolutional neural network D={s ₁ ,s ₂ ,... , s _N} consistency (S _D) for a can be calculated quantitatively by the following formula 6.

(Equation 6)

인 경우이고, p는 상기 합성곱 신경망의 합습된 결과로서 3개의 문장 세트에 대한 일관성 확률이다.Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.

Preferably, after step (b), the step of displaying the consistency score between each sentence quantitatively analyzed in step (b) on a display screen of a separate display unit so that the user can visually see it. .

Preferably, after the step (b), the step of converting the consistency information data between each sentence quantitatively analyzed in the step (b) into a database (DB) for each text document or for each sentence and storing it in a separate storage unit. It may contain more.

Preferably, after the step (b), the step of transmitting the consistency information data between each sentence quantitatively analyzed in the step (b) to an external user terminal by wire or wirelessly through a separate communication unit may be further included. have.

A third aspect of the present invention provides a computer-readable recording medium in which a program capable of executing a method for analyzing consistency between sentences in the above-described text document is recorded.

The method for analyzing consistency between sentences in a text document according to the present invention may be implemented as a computer-readable code on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system.

For example, computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, hard disk, floppy disk, removable storage device, and non-volatile memory. And optical data storage devices.

According to the apparatus and method for analyzing consistency between sentences in the text document of the present invention as described above, the text in the document unit is received as input and decomposed into sentences and paragraphs, so that each unit sentence maintains consistency in the context of the entire text. By quantitatively measuring whether or not, there is an advantage of being able to effectively analyze the consistency of the document unit for quality evaluation and analysis of unstructured text data.

In addition, according to the present invention, there is an advantage of being able to effectively measure text quality by analyzing document consistency.

In addition, according to the present invention, in analyzing the meaning of the document summary and the text document, there is an advantage that can be used to filter documents below a specific threshold.

In addition, according to the present invention, when writing, there is an advantage that it is possible to easily measure whether or not the sentence is well written in a consistent sentence through, for example, a computer device.

In addition, according to the present invention, there is an advantage that it can be used for determining the suitability of not only humans but also natural language generation generated by machines (or artificial intelligence).

1 is an overall block diagram illustrating an apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention.
2 is an overall flowchart illustrating a method for analyzing consistency between sentences in a text document according to an embodiment of the present invention.
3 is a diagram illustrating a cross-referencing relationship between sentences in a matrix in a first consistency analysis method applied to an embodiment of the present invention.
4 is a conceptual diagram illustrating a sentence embedding model based on a convolutional neural network among the third coherence analysis method applied to an embodiment of the present invention.
5 is a conceptual diagram illustrating a convolutional neural network-based model structure for measuring document consistency among the third consistency analysis methods applied to an embodiment of the present invention.
6 is a diagram of an example in which input data of a document consistency measurement model is displayed in a table form among a third consistency analysis method applied to an embodiment of the present invention.

The above-described objects, features, and advantages will be described later in detail with reference to the accompanying drawings, and accordingly, one of ordinary skill in the art to which the present invention pertains will be able to easily implement the technical idea of the present invention. In describing the present invention, if it is determined that a detailed description of known technologies related to the present invention may unnecessarily obscure the subject matter of the present invention, a detailed description will be omitted.

Terms including ordinal numbers, such as first and second, may be used to describe various elements, but the elements are not limited by the terms. These terms are used only for the purpose of distinguishing one component from another component. For example, without departing from the scope of the present invention, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element. The terms used in the present application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise.

The terms used in the present invention have been selected from general terms that are currently widely used while considering functions in the present invention, but this may vary depending on the intention or precedent of a technician working in the field, the emergence of new technologies, and the like. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning of the terms will be described in detail in the description of the corresponding invention. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall contents of the present invention, not a simple name of the term.

When a part of the specification is said to "include" a certain component, it means that other components may be further included rather than excluding other components unless otherwise stated. In addition, terms such as "... unit" and "module" described in the specification mean units that process at least one function or operation, which may be implemented as hardware or software, or as a combination of hardware and software. .

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. However, the embodiments of the present invention exemplified below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below. The embodiments of the present invention are provided to more completely describe the present invention to those of ordinary skill in the art.

Combinations of each block of the attached block diagram and each step of the flowchart may be executed by computer program instructions (execution engine), and these computer program instructions are executed on the processor of a general purpose computer, special purpose computer or other programmable data processing equipment. As it may be mounted, its instructions executed by the processor of a computer or other programmable data processing equipment generate means for performing the functions described in each block of the block diagram or each step of the flowchart. These computer program instructions may also be stored in computer-usable or computer-readable memory that can be directed to a computer or other programmable data processing equipment to implement a function in a particular manner, so that the computer-usable or computer-readable memory It is also possible to produce an article of manufacture containing instruction means for performing the functions described in each block of the block diagram or each step of the flow chart.

In addition, since computer program instructions can be mounted on a computer or other programmable data processing equipment, a series of operation steps are performed on a computer or other programmable data processing equipment to create a process that is executed by a computer, It is also possible for the instructions to perform possible data processing equipment to provide steps for executing the functions described in each block of the block diagram and each step of the flowchart.

In addition, each block or each step may represent a module, segment, or part of code containing one or more executable instructions for executing specified logical functions, and in some alternative embodiments mentioned in the blocks or steps. It should be noted that it is also possible for functions to occur out of order. For example, two blocks or steps shown in succession may in fact be performed substantially simultaneously, and the blocks or steps may be performed in the reverse order of a corresponding function as necessary.

1 is an overall block diagram illustrating an apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention.

Referring to FIG. 1, an apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention largely includes a preprocessor 100 and a text consistency analysis unit 200. In addition, the apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention may further include a display unit 300, a storage unit 400, and a communication unit 500. Meanwhile, since the elements shown in FIG. 1 are not essential, the apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention may have more elements or fewer elements.

Hereinafter, components of an apparatus for analyzing consistency between sentences in a text document according to an embodiment of the present invention will be described in detail.

The preprocessor 100 receives unstructured data, that is, a text document in a document-based natural language, decomposes it into paragraphs and sentences, analyzes the decomposed paragraphs and sentences morphologically, and outputs a text document. do.

In this case, it is preferable to implement the text document in the natural language of the document unit to be inputted through a separate user input unit (not shown). The user input unit performs a function of outputting a specific input signal of a corresponding user according to a user's request or a user's manipulation, and is generally preferably made of a mouse and/or a keyboard, but is not limited thereto. May be made of a remote control or a touch screen.

For example, the touch screen may employ a resistive film method, a capacitive method, an infrared method, an ultrasonic method, and the like, and the capacitive method is most preferably applied in order to minimize the thickness.

The capacitive touch screen is typically an ITO (Indium Tin Oxide) whose structure is made of a conductive transparent plate, an electrode part formed with silver powder paint on the edge of the ITO, and an insulating coating that insulates the lower part of the electrode. It can be composed of wealth. Meanwhile, the ITO may be composed of an ITO film made of a light-transmitting resin, and an ITO coating layer in which a conductive material is coated on a lower portion of the ITO film.

In the capacitive touch screen as described above, when the upper surface of the ITO is touched with a finger, each electrode provided on the four sides according to the change in capacitance through the finger senses the touch position.

In addition, the text document in the natural language of the document unit may be implemented to receive wired/wireless input through a separate user terminal (not shown). In this case, the user terminal is preferably made of at least one mobile terminal device among a smart phone, a smart pad, or a smart note communicating through wireless Internet or portable Internet. Access to the preprocessor 100 such as personal PC, notebook PC, Palm PC, mobile play-station, DMB (Digital Multimedia Broadcasting) phone with communication function, tablet PC, iPad, etc. It can mean all wired/wireless home appliances/communications devices having a user interface for comprehensively.

In addition, the text consistency analysis unit 200 performs overall analysis and control of the apparatus for analyzing consistency between sentences in the text document according to an embodiment of the present invention, and in particular, the output from the preprocessor 100 through morpheme analysis. For text documents, the first consistency analysis method, which quantitatively calculates consistency by analyzing the keywords of the sentence excluding the first sentence of the paragraph and the keywords of the preceding sentence, extracts the features of each sentence and expresses them as vectors. The second consistency analysis method, which calculates the similarity of each sentence and uses it to quantitatively calculate the consistency, randomly generates inconsistent sentences using a deep learning-based convolutional neural network, and learns it using a convolutional neural network based on deep learning, and uses the combined results. It performs a function of quantitatively analyzing whether each unit sentence maintains consistency in the context of the entire text document by using at least one consistency analysis method among the third consistency analysis methods that quantitatively calculates consistency.

Here, it is preferable that the first consistency analysis method compares the morphemes of each sentence in order to determine how much adjacent sentences are semantically related, and calculates the consistency quantitatively using Equation 1 below.

(Equation 1)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence, and at least one cross-reference relationship to two adjacent sentences. If there is (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise it is defined as '0'.

In addition, the second coherence analysis method includes word embedding and sentence embedding in which each sentence is expressed as a feature vector through dimension reduction and abstraction by learning context information of words and sentences in text. ) To express the sentences constituting the text document as sentence vectors, calculate the similarity between each sentence vector through each sentence vector expressed, and then use the similarity between the calculated sentence vectors to the corresponding text document The semantic consistency of can be calculated quantitatively.

In this case, the sentence embedding is preferably defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.

In addition, the semantic consistency of the text document through the second consistency analysis method can be calculated by Equation 2 below using the average of the similarities between all sentences constituting the text document.

(Equation 2)

라고 할 때 문장 간의 유사도를 계산한 것이다.Here, N is the total number of sentences in the document, and Sim(s _i , s _j ) is a text document consisting of N sentences using Cosine Similarity D=(s ₁ , s ₂ ,... The sentence vector corresponding to the sentence s _i for ,s _N }

When said, the degree of similarity between sentences was calculated.

In addition, the third consistency analysis method performs the convolutional neural network learning using each sentence constituting the text document, the preceding sentence of each sentence, and the back sentence of each sentence in order to measure the consistency of the corresponding text document. However, from the result of morphological analysis of the input sentence to be used as an input of the convolutional neural network, one morpheme is assumed as a word, and each sentence is expressed by converting it into a word vector using word embedding. It is preferable to use the vectors of the words that are composed of them by replacing them with a sentence matrix.

At this time, a total of three sentences consisting of each sentence constituting the text document, the front sentence of each sentence, and the back sentence of each sentence are used as the input of the convolutional neural network, and N (the number of total sentences in the document) Consistent document D={s ₁ ,s ₂ ,… For ,s _N }, three sentences are defined as one set (q) to form the input data of the model, and the middle sentence among the three sentences in one set (q) is replaced with another random inconsistent sentence. Error learning data is generated, and'y _q =1' is set for a set of three sentences, which is one learning data, for sentences constituting a consistent document, and the middle sentence among the three sentences is For error training data generated by replacing with other inconsistent sentences, the model is trained by setting'y _q = 0', and the consistency can be quantitatively calculated through the combined results.

And, through the combined result of the convolutional neural network, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... , s _N} consistency (S _D) for a can be calculated quantitatively by the following equation 3.

(Equation 3)

인 경우이고, p는 상기 합성곱 신경망의 합습된 결과로서 3개의 문장 세트에 대한 일관성 확률이다.Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.

Additionally, the display unit 300 performs a function of displaying a consistency score between each sentence quantitatively analyzed in a corresponding text document on the display screen so that the user can visually view it under the control of the text consistency analysis unit 200.

Such display unit 300 is, for example, a liquid crystal display (LCD), a light emitting diode (LED), a thin film transistor liquid crystal display (TFT LCD), an organic light emitting diode. (Organic Light Emitting Diode, OLED), Flexible Display, Plasma Display Panel (PDP), Surface Alternate Lighting (ALiS), Digital Light Source Treatment (DLP), Silicon Liquid Crystal (LCoS), Surface Conduction Electron-emitting device display (SED), field emission display (FED), laser TV (quantum dot laser, liquid crystal laser), optoelectric liquid display (FLD), interferometric modulator display (iMoD), thick film dielectric electricity (TDEL), It may include at least one of a quantum dot display (QD-LED), a telescopic pixel display (TPD), an organic light emitting transistor (OLET), a laser fluorescent display (LPD), and a 3D display, but is not limited thereto. Anything that can display numbers, characters, or figures can be included.

Furthermore, the storage unit 400 stores the consistency information data between each sentence quantitatively analyzed in the text document under the control of the text consistency analysis unit 200 into a database (DB) for each text document and/or for each sentence. Performs the function of

Such a storage unit 400 is, for example, a flash memory type, a hard disk type, a multimedia card micro type, a card type memory (eg, SD or XD memory). Etc.), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), It may include at least one type of storage medium among magnetic memory, magnetic disk, and optical disk.

In addition, the communication unit 500 transmits the consistency information data between each sentence quantitatively analyzed in the text document under the control of the text consistency analysis unit 200 to an external user terminal (not shown) by wire and/or wirelessly. Functions.

At this time, the external user terminal receives the consistency information data between each sentence quantitatively analyzed in the text document transmitted from the communication unit 500 through the previously installed sentence consistency analysis application service, and the corresponding user can visually view it based on this. The display screen of the user terminal may be displayed in the form of text and/or graph so that it can be performed.

Various embodiments described herein may be implemented in a recording medium that can be read by a computer or a similar device using, for example, software, hardware, or a combination thereof.

According to hardware implementation, the embodiments described herein include application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs). , Processors, controllers, micro-controllers, microprocessors, and electric units for performing functions may be used. In some cases, such embodiments may be implemented by the text consistency analysis unit 200.

According to a software implementation, embodiments such as procedures or functions may be implemented together with separate software modules that perform at least one function or operation. The software code can be implemented by a software application written in an appropriate programming language. In addition, the software code may be stored in the storage unit 400 and executed by the text consistency analysis unit 200.

On the other hand, although not shown in the drawing, power required for each of the above-described units, that is, the preprocessing unit 100, the text consistency analysis unit 200, the display unit 300, the storage unit 400, and the communication unit 500, is supplied. It is preferable that a power supply unit (not shown) is further included. The power supply unit is preferably implemented to receive external AC power (eg, AC 220V) and convert it into various DC power sources, but the present invention is not limited thereto, and may be implemented as a conventional portable battery.

The apparatus for analyzing consistency between sentences in the text document according to the embodiment of the present invention configured as described above is preferably implemented by, for example, a personal PC or a notebook PC, but is not limited thereto, and through wireless Internet or portable Internet. It may be implemented as at least one mobile terminal device of a smart phone, a smart pad, or a smart note that communicates. In addition, it may be implemented with a Palm PC, a mobile play-station, a Digital Multimedia Broadcasting (DMB) phone with a communication function, a tablet PC, an iPad, and the like.

If, in a text document according to an embodiment of the present invention, when the device for analyzing the consistency between each sentence is made by a smartphone, the smartphone is a variety of applications desired by the user, unlike a general mobile phone (a feature phone). Application) As a phone based on an open operating system that can be downloaded, freely used, and deleted, all mobile phones with mobile office functions as well as commonly used functions such as voice/video calls and internet data communication Alternatively, it is preferable to understand it as a communication device including all Internet phones or tablet PCs that do not have a voice call function but can access the Internet.

Such a smartphone can be implemented as a smartphone equipped with a variety of open operating systems, and examples of the open operating systems include Symbian from NOKIA, Blackberry from RIMS, iPhone from Apple, and Microsoft. (MS)'s Windows Mobile, Google's Android, and Samsung Electronics' Sea.

As described above, since a smartphone uses an open operating system, unlike a mobile phone having a closed operating system, a user can arbitrarily install and manage various application programs.

That is, the aforementioned smart phone basically includes a control unit, a memory unit, a screen output unit, a key input unit, a sound output unit, a sound input unit, a camera unit, a wireless network communication module, a short-range wireless communication module, and a battery for power supply. do.

The control unit is a generic term for functional configurations for controlling the operation of the smartphone, includes at least one processor and an execution memory, and is connected to each functional configuration unit provided in the smartphone through a bus.

The control unit controls the operation of the smartphone by loading at least one program code provided in the smartphone through the processor to the execution memory for calculation, and transmitting the result to at least one functional configuration unit through the bus. .

The memory unit is a generic term for a nonvolatile memory provided in a smartphone, and stores and maintains at least one program code executed through the control unit and at least one data set used by the program code. The memory unit basically stores a system program code and a system data set corresponding to an operating system of a smartphone, a communication program code and a communication data set for processing wireless communication connection of the smartphone, and at least one application program code and application data set, and , The program code and data set for implementing the present invention are also stored in the memory unit.

The screen output unit is composed of a screen output device (e.g., a liquid crystal display (LCD) device) and an output module driving the same, and is connected to the control unit via a bus to receive an operation result corresponding to the screen output among various operation results of the control unit. Output to the screen output device.

The key input unit is composed of a key input device having at least one key button (or a touch screen device interlocking with the screen output unit) and an input module driving the same, and is connected to the control unit via a bus to perform various operations of the control unit. A command for commanding is inputted or data required for calculation of the control unit is inputted.

The sound output unit is composed of a speaker that outputs a sound signal and a sound module that drives the speaker, and is connected to the control unit by a bus to output an operation result corresponding to the sound output among various calculation results of the control unit through the speaker. . The sound module decodes sound data to be output through an existing speaker and converts it into a sound signal.

The sound input unit includes a microphone receiving a sound signal and a sound module driving the microphone, and transmits sound data input through the microphone to the control unit. The sound module encodes and encodes a sound signal input through the microphone.

The camera unit is composed of an optical unit, a charge coupled device (CCD), and a camera module driving the same, and acquires bitmap data input to the CCD through the optical unit. The bitmap data may include both image data of a still image and video data.

The wireless network communication module is a generic term for a communication configuration connecting wireless communication, and includes at least one antenna for transmitting and receiving radio frequency signals of a specific frequency band, an RF module, a baseband module, and a signal processing module, and the control unit and It is connected by a bus and transmits the calculation result corresponding to wireless communication among various calculation results of the control unit through wireless communication, or receives data through wireless communication and transmits it to the control unit, and at the same time, accessing and registering the wireless communication Maintain the procedures of communication, handoff.

In addition, the wireless network communication module includes a mobile communication configuration that performs at least one access to a mobile communication network, location registration, call processing, call connection, data communication, and handoff according to the CDMA/WCDMA standard. Meanwhile, according to the intention of a person skilled in the art, the wireless network communication module may further include a portable Internet communication configuration for performing at least one access to the portable Internet, location registration, data communication, and handoff according to the IEEE 802.16 standard, and the wireless network It is to be clear that the present invention is not limited by the wireless communication configuration provided by the communication module.

The short-range wireless communication module is composed of a short-range wireless communication module that connects a communication session using a radio frequency signal as a communication medium within a certain distance, and preferably, RFID communication, Bluetooth communication, Wi-Fi communication, and public communication according to ISO 180000 series. It may include at least one of wireless communication. In addition, the short-range wireless communication module may be integrated with the wireless network communication module.

The smart phone configured as described above means a terminal capable of wireless communication, and any device may be applied as long as it is a terminal capable of transmitting and receiving data through a network including the Internet in addition to a smart phone. That is, the smart phone may include at least one notebook PC, a tablet PC having a function of transmitting a short message and a function of connecting to a network, and a portable terminal capable of carrying and moving.

Hereinafter, a method for analyzing consistency between sentences in a text document according to an embodiment of the present invention will be described in detail.

2 is an overall flowchart illustrating a method for analyzing consistency between sentences in a text document according to an embodiment of the present invention, and FIG. 3 is a cross-reference relationship between sentences among a first consistency analysis method applied to an embodiment of the present invention. A diagram expressed as a matrix, and FIG. 4 is a conceptual diagram showing a sentence embedding model based on a convolutional neural network among the third coherence analysis methods applied to an embodiment of the present invention, and FIG. 5 is a third coherence analysis method applied to an embodiment of the present invention. It is a conceptual diagram showing the structure of a convolutional neural network-based model for measuring document consistency, and FIG. 6 is a diagram illustrating an example in which input data of a document consistency measurement model is shown in a table form among a third consistency analysis method applied to an embodiment of the present invention.

1 to 6, the method for analyzing consistency between sentences in a text document according to an embodiment of the present invention is, first, unstructured data, that is, a text document in a document-based natural language through the preprocessor 100. Is received as an input and decomposed into paragraphs and sentences, and then the decomposed paragraphs and sentences are morphologically analyzed and output as a text document (S100).

Thereafter, with respect to the text document morphologically analyzed and output in step S100 through the text consistency analysis unit 200, the consistency is quantitatively calculated by analyzing the keyword of the sentence excluding the first sentence of the paragraph and the keyword of the previous sentence. The first consistency analysis method, the second consistency analysis method, which extracts the features of each sentence and expresses it as a vector, then calculates the similarity of each sentence and uses it to quantitatively calculate the consistency. Using at least one of the third consistency analysis methods, which randomly generates and learns it using a deep learning-based convolutional neural network, and quantitatively calculates consistency through the combined results, each unit sentence is the context of the entire text document. Quantitatively analyze whether the phase consistency is maintained (S200).

At this time, in step S200, the first consistency analysis method compares the morphemes of each sentence to determine how semantically related adjacent sentences are. Therefore, the consistency can be quantitatively calculated using Equation 4 below.

(Equation 4)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence. As an element of the cross-referencing relationship of a sentence, it is determined in units of morphemes due to the characteristics of Korean (aggregating words), and is limited to parts of speech that have meaning in the sentence, such as nouns, adjectives, verbs, and roots.

That is, if there is at least one cross-reference relationship between two adjacent sentences (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise, ' It is defined as 0'.

The document consistency measurement process is described in detail through the example below.

(example)

Sentence 1: When water is heated, it boils and eventually evaporates .

Sentence 2: When the evaporated gas is cooled, it turns into water again.

Sentence 3: The transformation of a substance into a completely different substance is called chemical change.

Sentence 4: The same goes for melting ice into water .

Here, since two or more identical morphemes exist in sentence 1 and sentence 2, the value of R(1, 2) is '1'. Since the same morpheme does not exist in sentence 2 and sentence 3, the value of R(2, 3) is '0'. In the same way, the value of R(3, 4) becomes '0'. When the cross-reference relationship between sentences is expressed as a matrix, it is shown in FIG. 3.

In the end, the cross-reference relationship values of adjacent sentences are calculated as an average.

That is, (R(1, 2)+R(2, 3)+R(3, 4))/3 = (1+0+0)/3 = 0.333...

In addition, in the step S200, the second consistency analysis method includes word embedding in which each sentence is expressed as a feature vector through dimension reduction and abstraction by learning context information of words and sentences of text, and Sentence embedding (Sentence Embedding) is used to express the sentences constituting the text document as each sentence vector, and after calculating the similarity between each sentence vector through each expressed sentence vector, the similarity between each calculated sentence vector is calculated. Using this, the semantic consistency of the text document can be calculated quantitatively. In this case, the sentence embedding is preferably defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.

The second consistency analysis method will be described in more detail by way of example.

That is, sentence feature expression using Sentence Embedding is to express the sentence as a feature vector through dimension reduction and abstraction by learning the context information of words and sentences in text. By expressing the document as a sentence vector (Sentence Vector), it is possible to calculate the semantic similarity between sentences.

At this time, it is preferable to use'Sent2Vec' as the model for sentence embedding. 'Sent2Vec' is a model that extends the CBOW (Continuous Bag of Words Model) of the'FastText' model and the'Word2Vec' model in order to better learn the meaning of sentences, not words.

In the'Sent2Vec' model, sentence embedding is defined as an average of word vectors through embedding of words constituting the sentence, that is, word embedding. In addition, this model embeds not only the'uni-gram' word vector, but also the'bi-gram' of the word in each sentence into one new word, and learns together, all'uni-gram' word vectors and'bi-gram' 'Use the average of the vector of word pairs.

For example, when a sentence vector v _s of a sentence S and a target word vector to be predicted in CBOW for a word w in a vocabulary is v _w , a sentence vector is generated by the following equation (5). R(S) is a list of'n-grams (uni-gram + bi-gram)' created in sentence S.

(Equation 5)

And, the sentences constituting a document are expressed as vectors, and the semantic consistency of the document is measured by using the similarity between the sentence vectors. Cosine similarity is used as a measure of similarity between document vectors.

라고 할 때, 문장 간의 유사도(Sim(s_i,s_j))는 하기의 식 6의 아랫 부분 식과 같이 계산한다. 그리고, 한 문서의 일관성은 문서를 구성하고 있는 모든 문장들 간의 유사도의 평균을 이용하여 하기의 식 6의 윗 부분 식과 같이 계산한다.Document consisting of N sentences D={s ₁ ,s ₂ ,… The sentence vector corresponding to the sentence s _i for ,s _N }

In the case of, the similarity between sentences (Sim(s _i , s _j )) is calculated as the lower part of Equation 6 below. And, the consistency of a document is calculated as the upper part of Equation 6 below by using the average of the similarity between all sentences constituting the document.

(Equation 6)

And, in the step S200, the third consistency analysis method, in order to measure the consistency of the text document, is synthesized by using each sentence constituting the text document, the preceding sentence of each sentence, and the back sentence of each sentence. Perform multiplication neural network training.

In order to be used as an input of the convolutional neural network, one sentence is converted into a vector of words and expressed. That is, from the result of morpheme analysis of an input sentence, one morpheme is assumed as a word, and it is expressed by converting it into a word vector using word embedding. And, it is used by replacing the vectors of words constituting each sentence into a concatenated sentence matrix (Sentence Matrix).

로 이루어져 있다. 이때,

는 문장을 구성하는 단어 개수이다. 단어 임베딩을 이용하여 d차원의 단어 벡터를 사용한다고 하였을 때, 문장 매트릭스

은 입력 문장 s를 구성하고 있는 단어들의 벡터로 구성되어 있다. 즉, S의 i번째 열은 s의 i번째 단어의 벡터이다.For example, the sentence matrix is made of the input sentence s, and the words constituting the sentence

Consists of. At this time,

Is the number of words that make up the sentence. Sentence matrix when d-dimensional word vectors are used using word embedding

Is composed of a vector of words that make up the input sentence s. That is, the ith column of S is a vector of the ith word of s.

는 너비 m과 문장 매트릭스

와 같은 차원 크기 d를 가지는 가중치 행렬이다. 도 3과 같이 컨볼루션 필터는 문장 매트릭스를 stride 1씩 움직이면서 벡터

를 출력으로 생성한다. c의 구성요소는 하기의 식 7과 같이 계산된다.In addition, the convolution layer extracts important features from the sentence matrix using a convolution filter. Convolution filter

Is the width m and the sentence matrix

It is a weight matrix with a dimension d of equal to. As shown in Figure 3, the convolution filter moves the sentence matrix by stride 1,

Is generated as output. The component of c is calculated as in Equation 7 below.

(Equation 7)

는 문장 매트릭스의 열을 따라 움직이는 크기가 m인 행렬의 한 부분이고,

는 행렬 간의 성분곱(element-wise multiplication)을 뜻하는 연산자이다. 컨볼루션(Convolution) 연산 뒤에는 비선형 활성함수 'ReLU'를 적용한다.here,

Is a part of a matrix of size m moving along the columns of the sentence matrix,

Is an operator that means element-wise multiplication between matrices. After the convolution operation, the nonlinear activation function'ReLU' is applied.

In the pooling layer, the dimension of the feature maps resulting from the convolution layer is reduced and combined using max pooling. The maximum pooling operates in one column of the feature map matrix C, and returns the maximum value of the output of the convolutional layer as shown in Equation 8 below.

(Equation 8)

5 is a convolutional neural network-based model structure for measuring document consistency. In other words, each sentence constituting the document and the first and second sentences, and a total of three sentences, are used as inputs. A coherent document consisting of N (total sentences in the document) D=(s ₁ ,s ₂ ,... For ,s _N }, three sentences are defined as one set q to form the input data of the model.

At this time, error learning data is generated by replacing the middle sentence with another random inconsistent sentence among the three sentences in a set. 6 is an example of input data generated for a document consistency measurement model.

For example,'y _q =1' is set for a set of three sentences that are one training data for sentences constituting a consistent document. Then, the model is trained by setting'y _q = 0'for the error learning data generated by replacing the middle sentence of the three sentences with other random inconsistent sentences.

And, the three vectors x ₁ , x ₂ , x ₃ output from each of the three sentence embedding models enter the input of the join layer and are concatenated to become one vector for three sentences ( See equation 9).

(Equation 9)

In addition, in the hidden layer, a vector h is calculated as shown in Equation 10 below using a weight matrix W _h , a bias b _h and a nonlinear function f. In this case, h may be regarded as a final vector for three sentences that have undergone a series of operations such as convolution, pooling, and dimensional transformation from the input layer.

(Equation 10)

Then, h outputted from the output of the hidden layer is input to the Softmax classification layer, and the consistency probability for the three sentences is finally calculated as shown in Equation 11 below. W _s is the weight matrix and b _s is the bias.

(Equation 11)

As described above, the convolutional neural spot-based document consistency measurement model is trained by Equation 12 below using, for example,'negative conditional log-likelihood'.

(Equation 12)

는 전체 모델의 파라미터들이다. 단어 임베딩 행렬 W, 각 문장 임베딩 모델의 컨볼루션 레이어에서 사용하는 컨볼루션 필터 가중치와 바이어스(

), 그리고 히든 레이어의 가중치 행렬과 바이어스(W_h;b_h), 소프트맥스 레이어의 가중치 행렬과 바이어스(W_s;b_s)로 구성되어 있다. g⁽ⁱ⁾는 N개의 문장으로 구성되어 있는 문서 D={s₁,s₂,…,s_N}에서 모델 입력을 위해 구성된 i번째 3개의 문장 세트이다. 모델 학습은 상기의 식 12를 최소화하면서 모델 파라미터

들을 최적화한다.here,

Are the parameters of the entire model. The word embedding matrix W, the convolution filter weight and bias used in the convolution layer of each sentence embedding model (

), and the weight matrix and bias of the hidden layer (W _h ;b _h ), and the weight matrix and bias of the softmax layer (W _s ;b _s ). g ⁽ⁱ⁾ is a document consisting of N sentences D=(s ₁ ,s ₂ ,... ,s _N } is the i-th set of three sentences constructed for model input. Model training minimizes Equation 12 above and

Optimize them.

And, through the combined result of the convolutional neural network, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... , s _N} consistency score (S _D) for the can be calculated quantitatively by Equation 13 below.

(Equation 13)

인 경우이고, p는 상기 합성곱 신경망의 합습된 결과로서 3개의 문장 세트에 대한 일관성 확률이다.Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.

At this time, the consistency of the document is determined by the consistency of all sets of documents. Since inconsistent sentences have too much negative effect on the overall consistency of the document, the consistency of a document is calculated as the product of all set qs with'y _q =1'.

In addition, although not shown in the drawing, after the step S200, displaying the consistency score between each sentence quantitatively analyzed in the step S200 on a display screen of a separate display unit 300 so that the user can visually see it. It may contain more.

Furthermore, after the step S200, the step of converting the consistency information data between each sentence quantitatively analyzed in the step S200 into a database (DB) for each text document and/or for each sentence and storing it in a separate storage unit 400 It may include more.

In addition, after the step S200, the step of transmitting the consistency information data between each sentence quantitatively analyzed in the step S200 to an external user terminal (not shown) through wired and/or wirelessly through a separate communication unit 500 It may include more.

Meanwhile, the method for analyzing consistency between sentences in a text document according to an embodiment of the present invention may be implemented as a computer-readable code on a computer-readable recording medium. The computer-readable recording medium includes all types of recording devices that store data that can be read by a computer system.

For example, computer-readable recording media include ROM, RAM, CD-ROM, magnetic tape, hard disk, floppy disk, removable storage device, and non-volatile memory. And optical data storage devices.

In addition, the computer-readable recording medium can be distributed to a computer system connected through a computer communication network, and stored and executed as code that can be read in a distributed manner.

In the above-described text document according to the present invention, a device for analyzing consistency between sentences and a preferred embodiment of the method have been described, but the present invention is not limited thereto, and the scope of the claims, the detailed description of the invention, and the accompanying drawings It is possible to carry out various modifications within, and this also belongs to the present invention.

100: pretreatment unit,
200: text consistency analysis unit,
300: display unit,
400: storage unit,
500: communication department

Claims (23)

A pre-processing unit that receives a text document in a natural language in a document unit as an input, decomposes it into paragraphs and sentences, morphemely analyzes the decomposed paragraphs and sentences, and outputs a text document; And
For the text document output through the morpheme analysis from the preprocessor, a first consistency analysis method that quantitatively calculates consistency by analyzing the keywords of the sentences excluding the first sentence of the paragraph and the keywords of the preceding sentence, the feature of each sentence ) Is extracted and expressed as a vector, and through this, the second consistency analysis method that calculates the similarity of each sentence and uses it to quantitatively calculate the consistency, and randomly generates inconsistent sentences using a machine learning method, and this is a convolutional neural network based on deep learning. Text consistency that quantitatively analyzes whether each unit sentence maintains the contextual consistency of the entire text document by using at least one consistency analysis method among the third consistency analysis methods that quantitatively calculates consistency through the result of learning using and combined results. Including an analysis unit,
In the first consistency analysis method, each sentence in a text document, characterized in that the morphemes of each sentence are compared to determine how semantically adjacent sentences are related, and the consistency is quantitatively calculated using Equation 1 below. Device for analyzing the consistency of the liver.
(Equation 1)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence, and at least one cross-reference relationship to two adjacent sentences. If there is (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise it is defined as '0'.
delete The method of claim 1,
The second consistency analysis method includes word embedding and sentence embedding in which each sentence is expressed as a feature vector through dimension reduction and abstraction by learning context information of words and sentences in text. The sentences constituting the text document are expressed as sentence vectors, the similarity between each sentence vector is calculated through each sentence vector, and the meaning of the text document is used using the similarity between each sentence vector. Apparatus for analyzing consistency between sentences in a text document, characterized in that quantitatively calculating the appropriate consistency.
The method of claim 3,
The sentence embedding is defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.
The method of claim 3,
Each sentence in a text document, characterized in that the semantic consistency of the text document through the second consistency analysis method is calculated by Equation 2 below by using the average of the similarities between all sentences constituting the text document. Device for analyzing the consistency of the liver.
(Equation 2)

Here, N is the total number of sentences in the document, and Sim(s _i , s _j ) is a text document consisting of N sentences using Cosine Similarity D=(s ₁ , s ₂ ,... The sentence vector corresponding to the sentence s _i for ,s _N }

When said, the degree of similarity between sentences was calculated.
The method of claim 1,
In the third consistency analysis method, in order to measure the consistency of the text document, the convolutional neural network learning is performed using each sentence constituting the text document, the preceding sentence of each sentence, and the back sentence of each sentence, In order to use the convolutional neural network as an input, after morphological analysis of the input sentence, assuming one morpheme as a word, converting it into a word vector using word embedding, and constructing each sentence. A device for analyzing consistency between sentences in a text document, characterized in that the vectors of existing words are replaced with a connected sentence matrix (Sentence Matrix).
The method of claim 6,
As the input of the convolutional neural network, a total of three sentences consisting of each sentence constituting the text document, the front sentence of each sentence, and the back sentence of each sentence are used, but N (the number of total sentences in the document) sentences Consistent document consisting of D={s ₁ ,s ₂ ,… For ,s _N }, three sentences are defined as one set (q) to form the input data of the model, and the middle sentence among the three sentences in one set (q) is replaced with another random inconsistent sentence. Error learning data is generated, and'y _q =1' is set for a set of three sentences, which is one learning data, for sentences constituting a consistent document, and the middle sentence among the three sentences is Consistency analysis between sentences in a text document, characterized in that the model is trained by setting'y _q = 0'for the error learning data generated by replacing with other inconsistent sentences, and the consistency is calculated quantitatively through the combined results. Device.
The method of claim 7,
Consistent document D={s ₁ , s ₂ ,... consisting of N sentences through the combined result of the convolutional neural network. Consistency (S _D ) for,s _N } is a device for analyzing consistency between sentences in a text document, characterized in that it is quantitatively calculated by Equation 3 below.
(Equation 3)

Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.
The method of claim 1,
Consistency between sentences in a text document, characterized in that further comprising a display unit that displays the consistency score between each sentence quantitatively analyzed in the text document under the control of the text consistency analysis unit on the display screen so that the user can visually see the consistency score Analysis device.
The method of claim 1,
In a text document, characterized in that it further comprises a storage unit that converts and stores consistency information data between each sentence quantitatively analyzed in the text document under the control of the text consistency analysis unit into a database (DB) for each text document or each sentence. Consistency analysis device between each sentence.
The method of claim 1,
Consistency between each sentence in a text document, characterized in that further comprising a communication unit for transmitting the consistency information data between each sentence quantitatively analyzed in a corresponding text document to an external user terminal by wire or wireless under the control of the text consistency analyzer Analysis device.
As a method of analyzing the consistency between each sentence in a text document using a device including a preprocessor and a text consistency analysis unit,
(a) receiving a text document in a natural language in a document unit as an input through the preprocessor, decomposing it into paragraphs and sentences, morpheme analysis of the decomposed paragraphs and sentences, and outputting the text document; And
(b) A system that quantitatively calculates consistency by analyzing keywords of sentences other than the first sentence of the paragraph and keywords of the preceding sentence with respect to the text document output by morpheme analysis in step (a) through the text consistency analysis unit. 1 Consistency analysis method, the second consistency analysis method, which calculates the similarity of each sentence by extracting the features of each sentence and expresses it as a vector, and uses this to quantitatively calculate the consistency, and randomly randomizes inconsistent sentences using the machine learning method. By using at least one of the third consistency analysis methods that generate and learn them using a deep learning-based convolutional neural network, and quantitatively calculate the consistency through the combined results, each unit sentence is in the context of the entire text document. It includes the step of quantitatively analyzing whether it is consistent,
In the step (b), the first consistency analysis method is characterized in that the morphemes of each sentence are compared to determine how semantically related adjacent sentences are, and consistency is quantitatively calculated using Equation 4 below. How to analyze the consistency between each sentence in a text document.
(Equation 4)

Here, N is the total number of sentences in the document, and R(i, i+1) indicates the state of the cross-reference relationship between the i-th sentence and the i+1th sentence, and at least one cross-reference relationship to two adjacent sentences. If there is (or if morphemes with meaningful parts of speech exist identically), the value of R(i, i+1) is '1', otherwise it is defined as '0'.
delete The method of claim 12,
In the step (b), the second coherence analysis method includes word embedding in which each sentence is expressed as a feature vector through dimension reduction and abstraction by learning context information of words and sentences of text, and Sentence embedding (Sentence Embedding) is used to express the sentences constituting the text document as each sentence vector, and after calculating the similarity between each sentence vector through each expressed sentence vector, the similarity between each calculated sentence vector is calculated. A method for analyzing consistency between sentences in a text document, characterized in that the semantic consistency of a corresponding text document is calculated quantitatively by using.
The method of claim 14,
The sentence embedding is defined as an average of word vectors through word embedding, which is the embedding of words constituting the sentence.
The method of claim 14,
Each sentence in a text document, characterized in that the semantic consistency of the text document through the second consistency analysis method is calculated by Equation 5 below using the average of the similarity between all sentences constituting the text document. How to analyze the consistency of the liver.
(Equation 5)

Here, N is the total number of sentences in the document, and Sim(s _i , s _j ) is a text document consisting of N sentences using Cosine Similarity D=(s ₁ , s ₂ ,... The sentence vector corresponding to the sentence s _i for ,s _N }

When said, the degree of similarity between sentences was calculated.
The method of claim 12,
In the step (b), the third consistency analysis method is performed by using each sentence constituting the text document, the preceding sentence of each sentence, and the back sentence of each sentence to measure the consistency of the corresponding text document. Multiply neural network learning is performed, but from the result of morphological analysis of the input sentence for use as an input of the convolutional neural network, one morpheme is assumed as a word and expressed by converting it into a word vector using word embedding. Thereafter, a method for analyzing consistency between sentences in a text document, characterized in that the vectors of words constituting each sentence are replaced with a sentence matrix in which they are connected.
The method of claim 17,
As the input of the convolutional neural network, a total of three sentences consisting of each sentence constituting the text document, the front sentence of each sentence, and the back sentence of each sentence are used, but N (the number of total sentences in the document) sentences Consistent document consisting of D={s ₁ ,s ₂ ,… For ,s _N }, three sentences are defined as one set (q) to form the input data of the model, and the middle sentence among the three sentences in one set (q) is replaced with another random inconsistent sentence. Error learning data is generated, and'y _q =1' is set for a set of three sentences, which is one learning data, for sentences constituting a consistent document, and the middle sentence among the three sentences is Consistency analysis between sentences in a text document, characterized in that the model is trained by setting'y _q = 0'for the error learning data generated by replacing with other inconsistent sentences, and the consistency is calculated quantitatively through the combined results. Way.
The method of claim 18,
Consistent document D={s ₁ , s ₂ ,... consisting of N sentences through the combined result of the convolutional neural network. , s _N} consistency (S _D) is consistent way analysis between each sentence in the text document, characterized in that quantitatively calculated by the following formula 6 for the.
(Equation 6)

Here, a consistent document consisting of N sentences D={s ₁ ,s ₂ ,... The set (q) defined for ,s _N } is

Is the case, and p is the probability of consistency for three sets of sentences as a result of the convolutional neural network combined.
The method of claim 12,
After the step (b), displaying the consistency score between each sentence quantitatively analyzed in the step (b) on a display screen of a separate display unit so that the user can visually see the text How to analyze the consistency between each sentence in the document.
The method of claim 12,
After the step (b), the step of converting the consistency information data between each sentence quantitatively analyzed in the step (b) into a database (DB) for each text document or each sentence, and storing the data in a separate storage unit. A method for analyzing consistency between sentences in a text document, characterized in that.
The method of claim 12,
After the step (b), transmitting the consistency information data between each sentence quantitatively analyzed in the step (b) by wire or wirelessly through a separate communication unit to an external user terminal, characterized in that it further comprises A method of analyzing the consistency between each sentence in a text document.
A computer-readable recording medium storing a program capable of executing the method of claim 12, 14 to 22 by a computer.

Images