KR102607826B1 - Deep neural network-based document analysis system and method, and computer program stored in recording media and media in which the program is stored - Google Patents

Abstract

It relates to a deep neural network-based document analysis system and method according to this embodiment, a recording medium storing a program for implementing the same, and a computer program stored in the medium, including extracting keywords from a sentence or document through morphological analysis and prototype processing. Wow, a step of determining the presence of a dictionary with multiple individual dictionaries and creating a dictionary if there is no dictionary, and quantifying each sentence using the extracted keywords and dictionaries, and calculating the number of bits corresponding to the number of individual dictionaries. A deep neural network-based document analysis system and method including converting the numerical value into a numerical value and analyzing the sentence or document using a deep neural network, and a recording medium and a medium storing a program for implementing the same. Provides computer programs.

Description

A deep neural network-based document analysis system and method, a recording medium storing a program to implement the same, and a computer program stored on the medium IS STORED}

The present invention relates to a document analysis system and method based on a deep neural network, a recording medium storing a program for implementing the same, and a computer program stored in the medium, and analyzes the risk of the document by learning and classifying semantic elements for each sentence of a text document. Provides a deep neural network-based document analysis system and method, a recording medium storing a program for implementing the same, and a computer program stored on the medium.

In general, documents such as contracts or proposals are defined in a somewhat certain format in terms of form and content in accordance with laws, regulations, etc. In addition, there is a disadvantage that it is very difficult for users without expertise in relation to the document to read the document, understand its meaning, and find risks or missing parts in the document. Of course, this can be resolved by consulting and reviewing the document through an expert, but since the expert reviews the document within a limited scope of information, the accuracy of the review is low, and even if he or she is an expert, he or she cannot review all items in a large document. It has the disadvantage of not being able to effectively analyze missing or incorrect items, even unfavorable items, and risks within the document due to the inability to cover them.

(Patent Document 1) Korean Patent Publication No. 10-2118603 (Patent Document 2) Korean Patent Publication No. 10-2020-0103152

The present invention was developed to solve the above-mentioned problems. In order to classify documents according to classification criteria set by the user, the semantic elements of each sentence are quantified, deep neural network-based learning is performed on these, and classification is based on the learning results. Provides a deep neural network-based document analysis system and method that can classify documents according to standards, a recording medium storing a program for implementing the same, and a computer program stored in the medium.

A preprocessing module that extracts keywords for each sentence of the target document according to the present invention, a dictionary module that groups similar categories of keywords by the preprocessing module and the learning control module and categorizes a number of words into a plurality of individual dictionaries; A quantification module that quantifies the keywords extracted through the preprocessing module based on the dictionary module, a deep neural network module that learns or predicts and analyzes the numerical value of the quantification module through a deep neural network, and quantification through a deep neural network module according to learning standards. It provides a deep neural network-based document analysis system that includes a learning control module that controls learning of sentences that have passed through the module and an analysis module that controls analysis of analyzed documents.

For the above-described operation, the preprocessing module includes a sentence separation module that separates sentences from a document, a word separation module that separates words from the separated sentences, and a prototype extraction module that extracts and selects the original form from the separated word.

The prototype extraction module includes a morpheme analysis unit that analyzes the morphemes of separated words, a stop word removal unit that removes stop words or unnecessary particles according to the morpheme analysis results, matches the original form of the word, and multi-semantic words and compound nouns. It includes a circular selection unit that selects a circular shape, such as.

The dictionary module includes a category classification unit that receives keywords in a document and classifies them, an individual dictionary creation unit that creates an individual dictionary corresponding to the classified category, a keyword expansion unit that expands keywords located in the generated individual dictionary, and a generation unit. It includes a separate dictionary storage unit that stores individual dictionaries.

The numerical module digitizes sentences using keywords extracted for each sentence based on the dictionary module, expresses them in bits corresponding to the number of individual dictionaries, checks whether each keyword is in each individual dictionary, and checks whether each keyword is in the nth individual dictionary. If there is a corresponding keyword word, the corresponding individual dictionary position value is expressed as 1.

The numerical module determines for each sentence unit whether or not the keyword in the sentence matches a word in the category word part of each individual dictionary, and if it matches, it is digitized as 1, and if it does not match, it is digitized as 0. do.

The quantification module sequentially compares keywords with individual dictionaries and quantifies them as 0 and 1. However, if multiple keywords match in one individual dictionary, the individual dictionary position value is quantified as 1, and if one keyword has multiple If it matches an individual dictionary of , the position value of each individual dictionary is quantified as 1.

In addition, a step of extracting keywords from a sentence or document according to the present invention through morphological analysis and prototype processing, determining whether a dictionary has a plurality of individual dictionaries, and generating a dictionary if there is no dictionary, and extracting keywords In-depth analysis includes the steps of quantifying each sentence using a dictionary and converting them into numbers with the number of bits corresponding to the number of individual dictionaries, and analyzing the sentences or documents using a deep neural network. Provides a neural network-based document analysis method.

The step of extracting the keyword includes separating sentences from the document and distinguishing words from the separated sentences, extracting the original form of the word based on the separated words, removing unnecessary particles, and, in the case of nouns, singular form. In the case of verbs, it is replaced with the current base form, and if multiple nouns or meanings are combined in one word through morphological analysis, the original form of each of them is extracted. If both the original form of the noun and the verb exist, both the original form is extracted. It includes a step of extracting all of the words and a step of deleting words included in the terminology dictionary through a separate unnecessary terminology dictionary.

The step of creating the dictionary includes classifying categories that can be similarly grouped through keyword analysis, that is, TF-IDF analysis or category analysis, and classifying the classified categories into individual dictionaries, with the category name in the dictionary classification part. This includes the steps of separating and storing keyword words corresponding to the category into category word parts and increasing the words in each individual dictionary by expanding synonyms and synonyms for the keywords classified in each individual dictionary.

To quantify each sentence above, each keyword in the sentence is compared to whether it matches the category word part of the entire individual dictionary, and the vector value of the individual dictionary that matches is expressed as 1, and if it does not match, it is expressed as 0, and the total individual dictionary is expressed as 0. It is characterized by digitizing sentences with the number of bits equal to the number of dictionaries.

If one keyword matches a word in multiple individual dictionaries, the output value corresponding to all matched individual dictionaries is set to 1, and if multiple keywords match within one individual dictionary, the output value of this single individual dictionary is set to 1. It is characterized by having a value of 1.

In addition, a recording medium in which a program is stored is provided to implement the deep neural network-based document analysis method described above.

Additionally, a computer program stored on a medium is provided to implement a deep neural network-based document analysis method.

In this way, the present invention quantifies semantic elements for each sentence and learns them based on a deep neural network in order to classify documents according to classification criteria set by the user, and can classify documents according to classification criteria based on the learning results. .

In addition, the analysis ability of deep neural network analysis can be improved and the operation speed can be increased by classifying the categories of words that appear in the document, dictionary them, and then using this dictionary to quantify the sentences in the document. .

Additionally, by limiting the scope of quantification to the number of individual dictionaries within the dictionary, simplification of the structure of the deep neural network can be expected.

1 is a conceptual diagram illustrating a deep neural network-based document analysis system according to an embodiment of the present invention.
Figure 2 is a block diagram for explaining a preprocessing module according to an embodiment.
Figure 3 is a block diagram for explaining a prototype extraction module according to an embodiment.
Figure 4 is a block diagram for explaining a dictionary module according to an embodiment.
Figure 5 is a conceptual diagram for explaining the operation of a digitization module according to an embodiment.
Figure 6 is a block diagram for explaining a deep neural network module according to an embodiment.
Figure 7 is a flowchart illustrating a deep neural network-based document analysis method according to an embodiment of the present invention.
Figure 8 is a flowchart illustrating a keyword extraction method according to an embodiment.
Figure 9 is a flowchart illustrating a method for quantifying sentences according to an embodiment.
Figure 10 is a flowchart for explaining a sentence learning method according to an embodiment.
Figure 11 is a flowchart for explaining a dictionary creation method according to an embodiment.

Hereinafter, embodiments of the present invention will be described in more detail with reference to the attached drawings. However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms. These embodiments only serve to ensure that the disclosure of the present invention is complete and to those skilled in the art to fully convey the scope of the invention. This is provided to inform you. In the drawings, like symbols refer to like elements.

It is intended to be clear that the division of components in this specification is merely a division according to the main function each component is responsible for. That is, two or more components, which will be described below, may be combined into one component, or one component may be divided into two or more components for more detailed functions. In addition to the main functions it is responsible for, each of the components described below may additionally perform some or all of the functions handled by other components, and some of the main functions handled by each component may be performed by other components. Of course, it can also be carried out exclusively by . Therefore, the presence or absence of each component described throughout this specification should be interpreted functionally. For this reason, the deep neural network-based document analysis system and method of the present invention and the recording medium on which the program for implementing the same is stored and the configuration of the components of the computer program stored on the medium will be different within the extent of achieving the purpose of the present invention. Please make it clear that you can.

In this specification, relational terms such as first and second, top and bottom are used to distinguish one entity or action from another without necessarily requiring or implying an actual relationship or order between such entities or actions. Can only be used. The terms “comprises,” “comprising,” or other variations thereof mean that a process, method, product, or device that includes a list of components does not include only those components, but that such process, method, product, or device includes a list of components. , or other components not explicitly listed or inherent in the device, are intended to cover non-exclusive inclusions. The reference to an element as “comprising an element” excludes the presence of an additional identical element within a process, method, product, or device including the element, without further limitation.

The features and advantages of the present invention and the method of implementing the same will be described in detail with the accompanying drawings regarding the deep neural network-based document analysis method, the recording medium storing the program for implementing the same, and the computer program stored in the medium.

However, the drawings and detailed descriptions briefly describe or omit the deep neural network-based document analysis method, the recording medium on which the program for implementing it is stored, and the configuration and method commonly known to those in the field of computer programs stored on the medium. The disclosure of the present invention is limited to the parts necessary for a clear understanding. Accordingly, the present invention is not limited to the embodiments disclosed below and may be implemented in various forms and methods, and the present invention is defined by the scope of the claims.

Figure 1 is a conceptual diagram for explaining a deep neural network-based document analysis system according to an embodiment of the present invention, Figure 2 is a block diagram for explaining a preprocessing module according to an embodiment, and Figure 3 is an embodiment. FIG. 4 is a block diagram for explaining a prototype extraction module according to an embodiment, FIG. 4 is a block diagram for explaining a dictionary module according to an embodiment, FIG. 5 is a conceptual diagram for explaining the operation of a digitization module according to an embodiment, and FIG. 6 is a block diagram for explaining a deep neural network module according to an embodiment.

1 to 6, the deep neural network-based document analysis system according to this embodiment includes a preprocessing module 100 that extracts keywords for each sentence of the target document, a preprocessing module 100, and a learning control module 500. ), a dictionary module 200 in which a number of words are categorized and classified into multiple individual dictionaries by grouping similar categories of keywords, and the keywords extracted through the preprocessing module 100 are quantified based on the dictionary module 200. A numericization module 300 that learns or predicts and analyzes the numerical value of the numeric module 300 through a deep neural network, and a digitalization module 300 through the deep neural network module 400 according to a learning standard. ) includes a learning control module 500 that controls learning of sentences that have been processed, and an analysis module 600 that controls analysis of the analysis document.

It is also possible to further include a DB module in which information and data processed or stored in the system and each module are stored.

The preprocessing module 100 extracts keywords from sentences or documents through morpheme analysis and prototype processing. It is possible to use a computer text document as the document.

The preprocessing module 100 separates sentences from the target document and distinguishes each word from the separated sentences. At this time, it is possible to separate sentences using sentence delimiters, and to separate words based on the spacing between letters, but it is not limited to this and it is possible to separate and distinguish them using various technologies. Here, it is effective to separate sentences from the document and then assign a unique identification number or symbol to each sentence. When keywords are extracted through word classification within a sentence, it is effective to display the unique identification number of the sentence in front of the keyword. For this purpose, it is possible to use various types of morpheme analysis techniques, such as classifying them into independent morphemes and dependent morphemes, or classifying them into substantive morphemes and formal morphemes. Through morphological analysis, it is possible to classify words including not only their actual meaning but also their grammatical meaning.

Then, based on the separated words, the original form of the word is extracted and stored as a keyword for each sentence. To extract the original form of a word, unnecessary particles are removed, and nouns are treated as singular. Remove pronouns, articles, and at the same time remove non-main words. Additionally, verbs are expressed in the present tense and basic tense. Additionally, if a word contains a noun or a verb, it is effective to extract both as original forms. When two nouns are combined in one word through morphological analysis, it is possible to extract each of them as their original form. Of course, it is not limited to this, and it is possible to check the prototype of a word through various prototype processing, and it is effective to use verbs and nouns as the prototype of the word. Of course, it is possible to use articles in their original form. However, it is effective to remove particles, interjections, and pronouns, and to remove unnecessary words using a separate dictionary, that is, a stop word dictionary.

For the above-described operation, the preprocessing module 100 includes a sentence separation module 110 that separates sentences from a document, a word separation module 120 that separates words from the separated sentences, and a word separation module 120 that separates words from the separated words. It may include a prototype extraction module 130 for extraction and selection. The original form extraction module 130 matches the original form of the word with a morpheme analysis unit 131 that analyzes the morphemes of the separated word, a dead word removal unit 132 that removes unused words or unnecessary particles according to the morpheme analysis results, and the original word of the word. It is possible to include a prototype selection unit 133 that selects the prototype of multi-meaning words and compound nouns. Of course, the morpheme analysis unit is included in the word classification module, so it is possible to simultaneously classify words and classify morphemes.

For example, considering the case of extracting keywords from the sentence “Admiral Yi Sun-sin achieved a great victory during the Japanese Invasion of Korea by using the turtle ship, etc.,” first classify by word. “Admiral Yi Sun-sin” can be classified as “in the Japanese Invasion of Korea”, “turtle ship”, “on his back”, “using”, “achieving a great victory”, and “achieving”. Afterwards, particles, exclamations, and unused words are removed from each word, and the original form of the word is extracted: “Admiral Yi Sun-sin,” “Imjin War,” “Turtle Ship,” “Use,” “Use,” “Great Victory,” and “Reap.” It can be classified as: And, as mentioned earlier, in the case of compound nouns with consecutive nouns, it is effective to separate them and separate "General Yi Sun-sin" into "Yi Sun-sin" and "General." Here, in the case of using, the original form can be both “use” as a noun and “use” as a verb. In this embodiment, it is effective to judge both of them as circular. In addition, in the case of verbs such as "do" and "be" as verbs through morphological analysis, it is effective to judge them as prototypical nouns and verbs in most cases. In other words, it is possible to extract all of the words 'study', 'study', 'discover' and 'discovered' into original forms. In addition, in the case of compound nouns, they are separated, but all possible cases of the original form are extracted as the original form. For example, in the case of “independent sentence”, “independent”, “sentence”, and “independent sentence” are extracted as the original form, and in the case of “English grammar”, “English sentence”, “grammar”, “law”, and “English grammar” are extracted as the original form. It is possible. The original form of the word extracted in this way is stored. At this time, it is effective to display the unique identification number of the sentence classified in the document at the front of the stored data. Through this, it is possible to check how many keywords one sentence consists of. As mentioned earlier, the same number of keywords as the number of words in one sentence may appear, but the present invention is not limited to this, and a larger number of keywords may be created.

After the keywords in the document and sentence are extracted through the preprocessing module 100, the dictionary module 200 creates an individual dictionary using the keywords, and the digitization module 300 converts each sentence into the dictionary module 200. Quantify it using .

In the following, the creation of individual dictionaries for numerical changes through the dictionary module 200 will be described. The dictionary module 200 is created by the learning control module 500, and the contents of the dictionary can be reinforced, changed, or added by the analysis module 600.

The dictionary module 200 groups keywords generated by the preprocessing module 100 into similar categories and classifies these grouped keywords into individual dictionaries.

The dictionary module 200 receives all generated keywords as input. Categories are extracted through analysis of input keywords. Category refers to keywords, that is, words that have similar meanings or belong to the same category and can be grouped. To distinguish these categories, in this embodiment, it is possible to use TF-IDF analysis or to create a classification system based on criteria such as meaning or similarity. In other words, it is possible to classify words by assigning weights to them according to the frequency of words in the document. Of course, it is also possible to distinguish by evaluating the importance of keywords rather than frequency. Additionally, it is possible to evaluate both frequency and importance of keywords. Through this, it is possible to indicate in numerical terms how frequently a word appears in the document and how important it is. Based on this, using various classification systems, it is possible to determine which categories of words are most common and their numerical value is high. It is possible to distinguish whether it is low or not.

In this way, keywords are divided into identical and/or similar categories, and then these categories are selected into an individual dictionary. At this time, the individual dictionary is divided into a dictionary division part and a category word part. The dictionary classification part may be a name, ID, or classification standard that distinguishes the category, and the category word part may be a space where words belonging to that category are located. Each of these distinct categories consists of a separate dictionary.

Of course, it is effective to store words in one category without duplication. Additionally, it is possible for one word, that is, a keyword, to belong to multiple categories. It is effective to determine the meaning of a word and place words with multiple meanings in individual dictionaries of the category to which they belong.

If the category classification operation of the dictionary module 200 is explained as a case of classifying a construction company's contract or proposal (RFP) document, the dictionary module 200 analyzes the keywords classified by the preprocessing module 100 and It may be possible to separate into categories, material categories, device or equipment categories, location categories, people categories and cost (money) categories, etc. Of course, there may be more such categories, and it is also possible that they do not exist at the time of the current analysis but are added later. That is, the category classification is not limited to this and can be added or changed through analysis of various will documents in the dictionary module 200.

After the categories are divided in this way, the category word portion of the dictionary is expanded based on the category, that is, the word in the individual dictionary, that is, the keyword. In other words, we proceed with the expansion of synonyms and synonyms. To expand category words, it is effective to use data from language dictionaries. At this time, it is possible to use the frequency of synonyms and assign weight by examining the frequency. Additionally, it is also possible to use dictionary location information, that is, index information technology. Through this, if there are words "price" and "cost" in the cost (money) category through keywords, if you proceed with the expansion, price, cost, expenditure, expense, consumption, value, market price, market price, sacrifice, It can be expanded to words such as loss. In other words, if there is a word building in the category, individual dictionary, it can be expanded to include words such as structure, edifice, and grocerles.

As described above, when the categories are separated using keywords, a number of individual dictionaries corresponding to the categories are formed, and the classified dictionaries are visualized according to the dictionary module 200 that expands the keywords, as shown in the table below. Expression may be possible.

dictionary division part Category Word Part Individual Dictionary 1 TIME Time, april, day, dec, December, everyday, hour, jan, January, jul, july, jun, june, march, may, minute, month
… Individual Dictionary 2 PERSON Adjudicator, advocate, affidavit, arbiter, arbitral, arbitration, employ, employee, employer, engineer, person, people… Individual Dictionary 3 MONEY Bankrupt, benefit, bidding, bill, bond, capital, cash, cent, coin, cost, debt, debtor, discount, dollar, earn, euro, exchange… Individual dictionary 4 PLACE Boundary, bridge, building, central, country, factory, farm, gate, geo, hall, highway, house, indoor, land, location… ... ... ...

As shown in the table, more individual dictionaries can be added, and more categories of dictionary divisions can also be added. In addition, it is possible to add more category word parts according to the analysis of the document and to add more as it expands.

The dictionary module 200 that performs this operation includes a category classification unit 210 that receives and classifies keywords in the document for the above-described operation, and an individual dictionary creation unit 220 that generates an individual dictionary corresponding to the classified category. and a keyword expansion unit 230 that expands keywords located in the generated individual dictionary. It is also possible to further include a separate dictionary storage unit 240 that stores the created individual dictionary.

The numericalization module 300 uses keywords extracted for each sentence based on the dictionary module 200 described above and converts them into numerical form. In the case of individual dictionaries, they have an independent state with no relationship to each other. Accordingly, it can be expressed in vector form with as many dimensions as the number of dictionaries.

According to the learning control module 500 and the analysis module 600, the numericalization module 300 matches the keywords extracted for each sentence by the preprocessing module 100 to each individual dictionary of the dictionary generated by the dictionary module. It is displayed numerically depending on the presence or absence. In other words, it is possible for one sentence to be expressed with the number of bits according to the number of individual dictionaries.

For each sentence unit, the digitization module 300 determines whether the keyword in the sentence matches the word in the category word portion in each individual dictionary. At this time, if they match, it is expressed as 1, and if they do not match, it is expressed as 0. Because of this, one sentence can be expressed with the number of bits corresponding to the number of individual dictionaries. For example, if there are 5 individual dictionaries, the total number of bits that can be expressed is “00000”, which can be 5 bits.

The digitization module 300 checks whether each keyword is in each individual dictionary, and if the corresponding keyword word is in the nth individual dictionary, it expresses the corresponding individual dictionary position value as 1. For example, if one sentence consists of 5 keywords and there are 5 individual dictionaries as before, the keywords and the category word portion of the individual dictionaries are compared sequentially or simultaneously, and the results are as shown in the table below. It is possible to express

division Individual Dictionary 1 Individual Dictionary 2 Individual Dictionary 3 Individual dictionary 4 Individual Dictionary 5 Keyword 1 0 One 0 0 0 Keyword 2 0 0 0 0 0 Keyword 3 0 One 0 One 0 Keyword 4 0 0 0 0 0 Keyword 5 0 0 0 0 0 Numerical result 0 One 0 One 0

Looking at Table 2, since there are five individual dictionaries, one sentence can be expressed with 5 bits. There are 5 keywords in one sentence, and when compared to individual dictionaries, the first bit is expressed as 0 because the keywords are not located in individual dictionary 1. And, keyword 1 and keyword 3 are located in individual dictionary 2. In this case, because there are two keywords, the second bit is expressed as 1. At this time, quantification by keyword is expressed as 1 even if at least one of the keywords is located. Of course, even if all keywords are in the individual dictionary, they are expressed as 1. Next, since keywords are not located in individual dictionaries 3 and 5, the 3rd and 5th bit values are also 0. Since keyword 3 is located in individual dictionary 4, the 4th bit is expressed as 1. At this time, in the case of keyword 3, it is simultaneously located in individual dictionary 2 and individual dictionary 4, but since the individual dictionaries are operated and reviewed separately, they can each be expressed as 1. Therefore, if the sentence with the five keywords in the example is digitized using the digitization module, the value of “01010” is obtained.

As another example, if the numerical value is "000010001000010100", there are 18 individual dictionaries, and the word corresponding to the character keyword for the sentence is located in the 5th individual dictionary and the 9th, 14th, and 16th individual dictionary. can be seen.

In the present invention, the number of dictionaries reviewed by the digitization module 300 is not limited and may vary depending on category classification. Preferably, it is possible to use 10 to 1000 individual dictionaries. However, if the number of individual dictionaries is too large, a load may occur on the entire system, and if the number of individual dictionaries is too small, document analysis may not be accurate. Therefore, it is preferably effective to have 50 to 200 individual dictionaries. Of course, it is effective to vary this depending on the field of the document being analyzed. Additionally, as mentioned earlier, the number of individual dictionaries may increase depending on the field of documents. Additionally, if necessary, it is possible to reduce the number by integrating individual dictionaries.

In the example of the previous preprocessing module 100, if keywords are extracted from the sentence “Admiral Yi Sun-sin achieved a great victory during the Japanese invasions of Korea by using the turtle ship, etc.”, the keywords are “Admiral Yi Sun-sin,” “Imjin invasions of Korea,” “turtle ships,” and “use.” , it is classified as “use”, “mahayana”, and “reap”. At this time, two keywords are created for the word use: the keyword use and the keyword use. This can be seen as keyword weighting. However, in the individual dictionary generated by the dictionary module 200, since use and use belong to one category, the individual dictionary where the keyword use is located and the individual dictionary where the keyword use is located are the same, so the individual dictionary where the keyword use is located is the same. The value may be displayed as 1 under constant conditions so that the weight assigned to it is offset. However, there is the advantage of being able to analyze documents more specifically by classifying keywords such as use and use.

The deep neural network module 400 is capable of performing learning for sentence and document analysis or analyzing sentences and documents using the numerical values of the digitization module 300. The deep neural network module 400 includes an input unit 410 that receives data, an output unit 420 that outputs results, and a hidden unit 430 that performs deep neural network analysis. In this embodiment, it is effective for the input unit 410 to be equal to the number of individual dictionaries, that is, the number of bits digitized by the digitization module 300. That is, the deep neural network module 400 includes an input unit 410 with a plurality of input nodes, an output unit 420 with one output node, and a hidden unit 430 with a plurality of hidden nodes located between them. do. It is effective for the input node of the input unit 410 to have an input node corresponding to the numerical value by the numerical conversion module 300.

It is effective that the hidden part 430 includes one or more layers connected as a whole, and the deep neural network module 400 of this embodiment is effectively divided by the relationship between each hidden layer.

When performing learning, assuming that the risk of documents and sentences is reviewed in this embodiment according to the provided information, it is effective for the output value of the output unit 420 to be expressed as 0 or 1. At this time, if there is a toxic clause, that is, an issue or risk, in the document/sentence, it is expressed as 1, and if there is no clause, it is expressed as 0.

When performing predictive analysis, if the output value is close to 1, it can be expressed as a toxic clause, that is, there is an issue or risk, and if it is close to 0, it can be predicted that there is no risk. In the case of predictive analysis, values between 0 and 1 are output. Practically 0.94, 0.85. Values such as 0.23 and 0.01 appear. Therefore, it is possible to judge that there is a risk if it is close to 1, and that there is no risk if it is close to 0.

When the learning control module 500 is running, the entire system learns whether a sentence is risky in a provided document. At this time, the learning is keyworded by the preprocessing module 100, and these keywords form an individual dictionary by the dictionary module, or are added to the category word portion in the already formed individual dictionary, and similar/synonyms thereof can also be added. .

Sentences classified as keywords by the numericization module 300 are converted into binary numbers with the same number of bit values as individual dictionaries, and are learned through the deep neural network module 400 to determine the corresponding sentences. Learn about the presence or absence of risk.

The analysis module 600 causes the deep neural network module 400 to perform an operation of predicting and analyzing the presence or absence of risk of the corresponding character based on the previously learned results.

In the following, a deep neural network-based document analysis method using the deep neural network-based document analysis system described above will be described.

In the method description described later, descriptions that overlap with the above-described system description will be omitted or simplified.

FIG. 7 is a flowchart for explaining a document analysis method based on a deep neural network according to an embodiment of the present invention, FIG. 8 is a flowchart for explaining a keyword extraction method according to an embodiment, and FIG. 9 is for an embodiment. FIG. 10 is a flowchart for explaining a sentence quantification method according to an embodiment, and FIG. 11 is a flowchart for explaining a dictionary creation method according to an embodiment.

Based on Figures 7 to 11, the deep neural network-based document analysis method is described as follows. First, as shown in Figure 7, the document is prepared to detect risks or toxic provisions in the document. The prepared document is preprocessed by the preprocessing module 100 to extract keywords corresponding to constituent words for each sentence in each document (S100).

To explain this in detail with reference to FIG. 8, sentences in the document are separated (S110). Then, words from each sentence are extracted through morphological analysis, etc. (S120). At this time, the original form of the word is extracted, unnecessary words or stop words are removed, and particles or articles are removed (S130). More specifically, in the case of verbs/adjectives, the original form is extracted according to the change in ending, and in the case of nouns, singular form processing is performed (S140). In addition, keywords corresponding to the sentence are extracted by removing words that are not important through semantic analysis. Here, if the original form of the word includes both a verb and a noun, it is effective to display both as keywords. Also, in the case of compound nouns, it is effective to separate them and extract all the extracted nouns as keywords. Additionally, it is desirable to delete stop words from a separate stop word dictionary (S150). The dictionary of stop words will be described later.

Next, the digitization module 300 uses a dictionary with a plurality of individual dictionaries and keywords for each sentence to numerically perform analysis through a deep neural network (S200).

Referring to Figure 9, each keyword in the sentence is compared to see whether it matches the category word portion of the entire individual dictionary (S210), and the vector value of the individual dictionary that matches is expressed as 1, and if it does not match, it is expressed as 0. Do it (S220). Through this, the sentence is expressed with the number of bits as the number of individual dictionaries (S230). When comparing and analyzing all keywords in an entire sentence, it is effective to express 0 for individual dictionary values that do not correspond to even one, and 1 if one or more keywords overlap. In this way, it is possible to digitize the bit value in binary form. In other words, it becomes possible to quantify it into n-dimensional coordinate values using keywords and individual dictionaries. In more detail, it checks whether each keyword word exists in each individual dictionary. As a result of checking, if it is in the mth dictionary, change the mth bit to 1. In addition, it is desirable to extract the final numerical data through sequential, continuous, or simultaneous comparison of all other keywords as above.

The presence or absence of risk is determined through deep neural network analysis of the quantified document through the deep neural network module 400 (S300). At this time, numerical data is input to the input unit, passes through a hidden unit with multiple layers, and is then output as a value between 0 and 1 to the output unit. And, the closer it is to 1, the more risky the sentence is judged to be (S400).

The layers of the input section are equal to the number of individual dictionaries applied. Additionally, it is effective to define the number of nodes and layers of the hidden part differently depending on the applied model.

In the above description, prediction and analysis of the target document were explained. For this purpose, it is necessary to conduct learning through a deep neural network in advance.

To proceed with such learning, as shown in FIG. 10, keywords are first calculated from various documents through a preprocessing module (S110). It is effective to extract keywords through preprocessing steps such as morphological analysis and prototype processing.

Then, each sentence is digitized using the keywords and dictionary extracted through the digitization module 300 (S1200). Numerical sentences are learned using a deep neural network. At this time, it is effective to learn 1 as a risky sentence and 0 as a risk-free sentence through learning using a deep neural network (S1300).

As described above, for learning and prediction/analysis using a deep neural network, quantification of sentences is necessary, and for this quantification, dictionary creation is necessary.

In this embodiment, to create a dictionary, keywords are calculated from various documents through a preprocessing module as shown in FIG. 11 (S2100). It is effective to extract keywords through preprocessing steps such as morphological analysis and prototype processing.

The generated keywords are analyzed and grouped into various categories. Then, the grouped categories are classified into one individual dictionary (S2200). Category classification is effectively performed through TF-IDF analysis and category analysis.

Next, each keyword is classified and placed into an individual dictionary (S2300), and the classified keywords are expanded into similar words and synonyms to complete the final dictionary (S2400).

Of course, when a new document is analyzed by the preprocessing module 100, new keywords can be added, new categories can be created, and new individual dictionaries can be created. When a new individual dictionary is created in the dictionary, the vector value changes when the sentence is digitized. In other words, this means that the number of bits is expanded. This means that the layer of the input terminal of the deep neural network increases, and changes by the dictionary module 200 induce changes in the deep neural network module 400. Accordingly, the deep neural network module 400 must perform new learning when the number of new individual dictionaries increases. It is also possible to construct dictionaries of different categories depending on the nature of the work to which it is applied, the type of document, and its intended use. In this embodiment, it is possible to first build a dictionary and then build the deep neural network module 400. Of course, it is also possible to make the deep neural network variable according to the dictionary.

The present invention is not limited to the configuration and method of the above-described embodiments, and various modifications and other embodiments are possible. That is, if the original form of a word includes a noun and a verb, the preprocessing module 100 can select either the noun or the verb rather than selecting both. Also, rather than separating it into a compound noun, it is possible to classify the compound noun itself as a prototype, that is, a keyword.

Additionally, the present invention is not limited to this, and it is possible to place all keywords of words with two or more original forms in the same dictionary in the dictionary module 200. And, it is possible to use nouns for keyword extraction. In other words, the efficiency of keyword analysis can be increased by using nouns rather than different parts of speech, and it is possible to prevent the duplication of words with similar meanings due to the use of different parts of speech, such as verbs.

In addition, when only nouns are extracted as keywords, dictionary construction can be simplified and words can be easily expanded, allowing for more effective deep neural network analysis.

Additionally, the present invention is not limited to this, and it is possible to provide a pre-processing module in a separate configuration. Since the dictionary module 200 serves as a basic standard for sentence analysis, results from analyzing various documents are collected. Therefore, as mentioned earlier, the preprocessing module can simultaneously analyze one sentence at once without separately distinguishing sentences from the document, separating them into words, and then distinguishing their original forms. In other words, it may be possible to have a separate pre-processing module for pre-processing. This improves the operational efficiency of the pre-processing module, efficiently derives keywords from the entire document rather than sentence analysis, and can be very effective in later classification and classification tasks, that is, category classification.

Although the technical idea of the present invention described above has been described in detail in preferred embodiments, it should be noted that the above-described embodiments are for illustrative purposes only and are not intended for limitation. In addition, an expert in the technical field of the present invention will understand that various embodiments of the present invention are possible within the scope of the technical idea of the present invention.

100: Preprocessing module 110: Sentence separation module
120: word classification module 130: prototype extraction module
131: Morpheme analysis unit 132: Stop word removal unit
133: circular selection unit 200: dictionary module
210: Category classification unit 220: Individual dictionary creation unit
230: Keyword expansion unit 240: Dictionary storage unit
300: Numericalization module 400: Deep neural network module
500: learning control module 600: analysis module

Claims (11)

A preprocessing module that extracts keywords for each sentence of the target document;
a dictionary module that groups similar categories of keywords by the pre-processing module and the learning control module and categorizes a number of words into a plurality of individual dictionaries;
a quantification module that quantifies sentences using keywords extracted through the pre-processing module based on a dictionary module;
A deep neural network module that learns or predicts and analyzes the numerical values of the numerical module through a deep neural network;
A learning control module that controls the learning of sentences that have gone through the numerical module through a deep neural network module according to learning criteria; and
Includes an analysis module that controls the analysis of the analysis document,
The numerical module digitizes sentences using keywords extracted for each sentence based on the dictionary module, expresses them in bits corresponding to the number of individual dictionaries, checks whether each keyword is in each individual dictionary, and checks whether each keyword is in the nth individual dictionary. If there is a corresponding keyword word, the corresponding individual dictionary position value is expressed as 1,
The numerical module determines for each sentence unit whether or not the keyword within the sentence matches a word in the category word part of each individual dictionary. If it matches, it is numerically 1, and if it is not a match, it is numerically 0.
The deep neural network module includes an input unit that receives data, an output unit that outputs results, and a hidden unit that performs deep neural network analysis,
The deep neural network module determines the risk of a sentence through deep neural network analysis of the numerical sentence by inputting numerical data into an input unit, passing through a hidden unit with multiple layers, and then outputting a value between 0 and 1 to the output unit. A deep neural network-based document analysis system characterized by:
According to paragraph 1,
For the above-described operation, the preprocessing module includes a sentence separation module that separates sentences from a document, a word separation module that separates words from separated sentences, and a prototype extraction module that extracts and selects the original form from the separated word. Neural network-based document analysis system.
According to paragraph 2,
The prototype extraction module includes a morpheme analysis unit that analyzes the morphemes of separated words, a stop word removal unit that removes stop words or unnecessary particles according to the morpheme analysis results, matches the original form of the word, and multi-semantic words and compound nouns. A document analysis system based on a deep neural network that includes a prototype selection unit that selects the prototype, etc.
According to paragraph 1,
The dictionary module includes a category classification unit that receives keywords in a document and classifies them, an individual dictionary creation unit that creates an individual dictionary corresponding to the classified category, a keyword expansion unit that expands keywords located in the generated individual dictionary, and a generation unit. A deep neural network-based document analysis system that includes a separate dictionary storage unit that stores individual dictionaries.
delete A document analysis method using a deep neural network-based document analysis system according to any one of claims 1 to 4, comprising:
A preprocessing module extracting keywords from a sentence or document through morphological analysis and prototype processing;
The dictionary module determines whether a dictionary has a plurality of individual dictionaries and creates a dictionary if there is no dictionary;
The digitization module digitizes each sentence using the extracted keywords and dictionaries, and converts them into numbers with the number of bits corresponding to the number of individual dictionaries; and
The deep neural network module includes analyzing the sentence or document using the numerical value using a deep neural network,
The digitization module digitizes each sentence by comparing each keyword in the sentence to whether it matches the category word part of the entire individual dictionary, expressing the vector value of the individual dictionary that matches as 1, and if it does not match, expresses it as 0. , A deep neural network-based document analysis method characterized by quantifying sentences with the number of bits as the number of individual dictionaries.
The method of claim 6, wherein the step of extracting the keywords includes:
A sentence separation module separates sentences from a document, and a word separation module separates words from the separated sentences;
The original form extraction module extracts the original form of the word based on the separated words, removes unnecessary particles, treats nouns as singular, replaces them with the current basic form for verbs, and adds them to a single word through morphological analysis. If plural nouns or meanings are combined, extracting the original form of each of them, and extracting both the original form of the noun and the verb if both exist; and
A deep neural network-based document analysis method in which the prototype extraction module deletes words included in the terminology dictionary through a separate unnecessary terminology dictionary.
The method of claim 6, wherein the step of generating the dictionary includes:
The category classification unit of the dictionary module includes the steps of classifying categories that can be similarly grouped through keyword analysis, that is, TF-IDF analysis or category analysis;
An individual dictionary creation unit classifies the classified categories into individual dictionaries, separately storing the category name as a dictionary division part and the keyword words corresponding to the category as a category word part; and
The keyword expansion unit is a deep neural network-based document analysis method that includes the step of increasing words in individual dictionaries by expanding synonyms and synonyms for keywords classified in each individual dictionary.
delete A recording medium storing a program to implement the deep neural network-based document analysis method described in paragraph 6.
A computer program stored on a medium to implement the deep neural network-based document analysis method described in clause 6.