KR102128852B1 - Device and method for visualizing key words of features extracted by applying principal component analysis to word vectors from text data - Google Patents

Abstract

The present invention relates to a key word visualizing device of a feature. The key word visualizing device of a feature comprises: a feature information input unit for inputting importance values for target variables of a plurality of features extracted from text data using principal component analysis; a reference importance input unit for setting a reference value related to the importance value based on user input information; a word-specific importance extracting unit for extracting the degree of influence of each feature for each word vector for the plurality of features; a word-by-word importance calculation unit converting the importance of the value extracted by the word-specific importance extracting unit into a percentage; and a visualization execution unit visualizing the importance of each word and information on the extracted features in the form of a Venn diagram graph.

Description

DEVICE AND METHOD FOR VISUALIZING KEY WORDS OF FEATURES EXTRACTED BY APPLYING PRINCIPAL COMPONENT ANALYSIS TO WORD VECTORS FROM TEXT DATA}

The present invention relates to an apparatus and method for visualizing a key word of a feature extracted by applying principal component analysis to a word vector generated from text data.

In the existing text data visualization process, a morphological analysis is applied to the data, and a word cloud method indicating importance based on the frequency of words is applied to a collection of words extracted.

However, this method does not visualize by extracting a word that has a large influence on the feature extracted from the text data by applying the Principal Component Analysis (PCA) after making the word a vector.

In the process of text data analysis, it is necessary to grasp information about which words influenced a feature a lot in order to obtain a description of the feature in the process of vectorizing words and extracting features.

For this, it is necessary to visualize the key word of the feature extracted by applying principal component analysis to the word vector generated from the text data.

The background technology of the present application is disclosed in Korean Patent Publication No. 10-2015-0048751.

The present application is intended to solve the above-described problems of the prior art, and by applying principal component analysis to word vectors generated from text data, it is possible to visualize the key words of the extracted features to grasp information about which words have influenced the features a lot. It aims to provide a function.

The present application is to solve the above-described problems of the prior art, and aims to provide a visualization function of a word cloud method using a Venn diagram shape for three features extracted by applying principal component analysis to word vectors generated from text data. Should be

The present application is for solving the above-mentioned problems of the prior art, and extracts a feature by applying principal component analysis to a word vector generated from text data, and performs performance when analyzing data in which structured data and text data are mixed using machine learning. An object of the present invention is to provide an apparatus and method for extracting features by applying principal component analysis to word vectors generated from text data capable of increasing.

The present application is intended to solve the problems of the prior art described above, and extracts a feature by applying principal component analysis to a word vector generated by applying a TF-IDF method to text data when analyzing a mixture of structured data and text data. It is an object of the present invention to provide an apparatus and method for extracting features by applying principal component analysis to word vectors generated from text data.

However, the technical problems to be achieved by the embodiments of the present application are not limited to the technical problems as described above, and other technical problems may exist.

As a technical means for achieving the above technical problem, the apparatus for visualizing a key word of a feature according to an embodiment of the present application is a feature for inputting importance values for target variables of a plurality of features extracted using principal component analysis from text data An information input unit, a reference importance input unit that sets a reference value associated with the importance value based on user input information, and a word-specific importance extraction unit that extracts an influence on each feature for each word vector for the plurality of features, the words It may include a importance calculation unit for each word that converts the importance as a percentage of the value extracted by the importance extraction unit for each star, and a visualization execution unit for visualizing the importance of each word and information about the extracted feature in the form of a Venn diagram graph.

In addition, when the target variable is set, the feature information input unit may calculate an influence degree of the extracted feature and the target variable, and input an importance value for the target variable.

In addition, the visualization execution unit may visualize in the cloud form by determining an output location, an output size, and an output color according to the importance of each word and the frequency of the extracted features.

In addition, the feature key word visualization apparatus uses a plurality of dictionaries to extract a key word from the received text data, and extracts the key word into a text analysis algorithm, vectorizes it, and then uses the principal component analysis. A feature extraction execution unit for extracting features of the text data may be further included.

In addition, the feature information input unit may input a variable importance value for the target variable based on the feature extracted by the feature extraction execution unit when the variable name of the text data and the target variable are present.

According to an embodiment of the present disclosure, the method of visualizing a key word of a feature includes receiving an importance value for a target variable of a plurality of features extracted by using principal component analysis from text data, and based on the user input information. Setting an associated reference value, extracting an influence on each feature for each word vector for the plurality of features, and extracting an influence on each feature for each word vector for the plurality of features It may include the step of converting the importance to a percentage and visualizing the importance of each word and information about the extracted feature in the form of a Venn diagram graph.

The above-described problem solving means are merely exemplary and should not be construed as limiting the present application. In addition to the exemplary embodiments described above, additional embodiments may exist in the drawings and detailed description of the invention.

According to the above-described problem solving means of the present application, it is possible to visualize and display the main word of the extracted feature by applying principal component analysis to the word vector generated from the text data.

According to the above-described problem solving means of the present application, it is possible to provide explanatory power for features in the process of vectorizing words and extracting features during text data analysis.

According to the above-described problem solving means of the present application, it is possible to extract a feature by applying principal component analysis to a word vector generated from text data.

According to the above-described problem solving means of the present application, it is possible to increase performance when analyzing data in which structured data and text data are mixed using machine learning.

However, the effects obtainable herein are not limited to the effects described above, and other effects may exist.

1 is a schematic configuration diagram of a feature word visualization device according to an embodiment of the present application.
2 is a diagram for explaining a visualization tool of a device for visualizing a key word of a feature according to an embodiment of the present application.
FIG. 3 is a diagram visualizing a list of words used in features extracted from the first text data in the apparatus for visualizing key words of the feature according to an embodiment of the present application.
4A to 4C are diagrams of visualizing a word list used for features extracted from second text data in a device for visualizing a key word of a feature according to an embodiment of the present application.
5A to 5C are diagrams of visualizing a word list used for features extracted from third text data on a device for visualizing a key word of a feature according to an embodiment of the present application.
6A to 6C are diagrams of visualizing a word list used for features extracted from fourth text data on a device for visualizing a key word of a feature according to an embodiment of the present application.
7A to 7C are diagrams visualizing a word list used for features extracted from the fifth text data in the apparatus for visualizing key words of features according to an embodiment of the present application.
8 is a schematic block diagram of a text data feature extraction apparatus according to an embodiment of the present application.
9 is a schematic diagram schematically illustrating a feature extraction process of a text data device according to an embodiment of the present application.
10 is an operation flowchart of a method for visualizing a key word of a feature according to an embodiment of the present application.

Hereinafter, embodiments of the present application will be described in detail with reference to the accompanying drawings so that those skilled in the art to which the present application pertains may easily practice. However, the present application may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in order to clearly describe the present application in the drawings, parts irrelevant to the description are omitted, and like reference numerals are assigned to similar parts throughout the specification.

Throughout this specification, when a part is "connected" to another part, it is not only "directly connected", but also "electrically connected" or "indirectly connected" with another element in between. "It includes the case where it is.

Throughout this specification, when a member is positioned on another member “on”, “on the top”, “top”, “bottom”, “bottom”, “bottom”, it means that a member is on another member. This includes cases where there is another member between the two members as well as when in contact.

Throughout the present specification, when a part “includes” a certain component, it means that the component may further include other components, not to exclude other components, unless otherwise stated.

1 is a schematic configuration diagram of a feature word visualization device according to an embodiment of the present application.

Referring to FIG. 1, the key word visualization device 150 of the feature includes a feature information input unit 151, a reference importance input unit 152, a word importance extraction unit 153, a word importance calculation unit 154, and visualization execution It may include a portion 160. In addition, the key word visualization device 150 of the feature may be provided with features extracted from the text feature extraction device 120. The text feature extraction device 120 may include a word extraction processor 130 and a feature extraction execution unit 140. However, the configuration of the text feature extraction device 120 is not limited thereto. For example, the text feature extraction device 102 may include a text data input unit (not shown) for receiving text data. In addition, the text feature extraction device 120 may include a database (not shown) for storing the extracted features.

According to an embodiment of the present application, after inputting text data, and extracting a major noun, vectorization is performed using a text analysis algorithm (eg, TF-IDF technique) and then principal component analysis (eg, PCA) In the process of extracting three features by applying the technique, the user can check only the extracted features. However, this feature is extracted by the principal component analysis (eg, PCA) technique, so it is composed only of numbers, so the user cannot know which word influenced each feature a lot.

In addition, since the feature word visualizer 150 can know the influence of the feature on each word vector obtained by using PCA.components_, which is a function of deciposition of Scikit-Learn, the words with high influence are given to the user. Visualization is provided so that the user can intuitively know which words are mainly used. The visualization method was designed in such a way that it provides a value converted in percentage and only words with an influence level above a certain value.

In addition, when there are variable names and target variables of text data, the key word visualization device 150 of the feature may input a variable importance value for the target variable of the feature extracted by principal component analysis. Also, the apparatus 150 for visualizing a key word of a feature may input a reference value input by a user in order to extract a word having an importance level higher than a specific value. In addition, the main word visualization device 150 of the feature may extract the influence of each feature on each word vector for three features extracted by applying principal component analysis to the word vector. Also, the apparatus 150 for visualizing a key word of a feature may convert the importance level into a percentage of the value extracted by the importance level extraction unit for each word. In addition, the feature word visualization device 150 may visualize the importance of each word and information about the extracted feature using a word cloud method using a venn diagram shape.

According to one embodiment of the present application, the text feature extraction device 120 may receive text data 110 from an external server (not shown). Also, the text feature extraction device 120 may receive text data 110 from a user terminal (not shown). The text data 110 is data for extracting features, and may be a document including Korean, English, special characters, and numbers. In addition, text data 110 includes legal documents, contracts, specifications, ITB, offshore and land plant data, ERP (enterprise resource management), PMIS (business management information system), commercial data, public data, big data integrated information, public It can include data, Open API, etc. The user inputs the text (document) desired for feature extraction into the text feature extraction device 120, and the text feature extraction device 120 extracts the key words using the dictionary after morpheme analysis on the received text data and extracts the main words After vectorizing by applying a text analysis algorithm to a word, a pre-set number of features (for example, three) may be extracted by applying a principal component analysis to the vector.

According to one embodiment of the present application, the word extraction processor 130 may extract a key word from the received text data using a plurality of dictionaries. The word extraction processor 130 may extract key words using a plurality of dictionaries such as a user dictionary and its own dictionary. The plurality of dictionaries may include a user dictionary, own dictionary, non-noun dictionary, human dictionary, and self-explanatory dictionary. The word extraction processor 130 may build a user dictionary based on user input information. In addition, the word extraction processor 130 may construct a noun dictionary including a plurality of general nouns. Also, the word extraction processor 130 may obtain a noun dictionary including a plurality of general nouns from an external server. In addition, the word extraction processor 130 may separate morphemes from text data and extract morphemes that can be words.

For example, the word extraction processor 130 may extract noun candidates from text data. When the text data is Korean, the word extraction processor 130 may extract a noun candidate by applying a noun extraction function (eg, the LRNounExtractor_v2 function of soynlp). Also, the word extraction processor 130 may extract a noun using a dictionary. A noun is a class of words having a meaning in a sense, and may indicate the name of an object.

Also, the word extraction processor 130 may extract nouns using a plurality of dictionaries. For example, among words extracted using the noun extraction function (for example, LynounExtractor_v2 function of soynlp), in addition to nouns, there are single-letter words, words with special characters or foreign languages, and words with nouns. In addition, there may be words that the user wants to extract among words that the function does not extract, and among the extracted words, there may be a key word that the user thinks more particularly. The word extraction processor 130 improves the quality of words using a dictionary prepared for vectorization by leaving only the main words.

According to one embodiment of the present application, the feature extraction execution unit 140 may apply the extracted key words to a text analysis algorithm to vectorize and then extract features of text data using principal component analysis. For example, Principal Component Analysis (PCA) is a technique that rotates a data set so that variables are not statistically correlated. Depending on how important it is to describe the data after rotation, you often choose only a few of the new traits. This new property is called the main component (each main component is influenced by all variables in the existing data), and if only a part is selected, it is used for the purpose of dimension reduction because the dimension is reduced. When extracting features of text data using principal component analysis, when vectorized words are used as features, the number of features becomes too large, and thus there is a problem in that performance may be poor in prediction. Therefore, the feature extraction execution unit 140 may extract features using principal component analysis. The number of features to be extracted is easy to visualize, and a preset number (for example, 3) that can express 80% to 90% of the variance of each vector can be set as a default value.

Also, the feature extraction execution unit 140 may vectorize a major noun using a vector extraction package (for example, feature_extraction.text, which is one of the subpackages of feature_extraction of Scikit-Learn). A vector extraction package (e.g., feature_extraction.text, one of the subpackages of Scikit-Learn's feature_extraction) provides a document preprocessing class, which is a text analysis function (e.g. a function that vectorizes words based on frequency) For example, the key nouns are vectorized using the TfidfTransformer function.

According to one embodiment of the present application, the feature information input unit 151 may input an importance value for a target variable of a plurality of features extracted using the principal component analysis from the text data 110. For example, the feature information input unit 151 may receive information of a plurality of features extracted using the principal component analysis from the text data 110 from the text feature extraction device 120. For example, when the feature variable input unit 151 has a text variable name of'text' and the words extracted from this variable are'word 1','word 2', ..., each variable used in the PCA is extracted from the feature. Table 1 shows how much influence it has. The influence degree shown in Table 1 may still be shown as a value indicating a positive correlation and a negative correlation.

word text_PCA_1 text_PCA_2 text_PCA_3 Word 1 0.003124 -0.093125 0.007445 Word 2 0.024356 0.098631 0.024323 Word 3 -0.345432 -0.145223 -0.343252 Word 4 -0.002244 -0.032174 -0.001175 Word 5 0.000229 0.011312 -0.09022 ... ... ... ...

In addition, when the target variable is set, the feature information input unit 151 may calculate the degree of influence between the extracted feature and the target variable, and input an importance value for the target variable. The feature information input unit 151 may calculate the degree of influence between the extracted feature and the set target variable by using principal component analysis on the vectorized word in the feature extraction unit 142. Also, the feature information input unit 151 may input an importance value for the target variable. In addition, the feature information input unit 151 may input a variable importance value for the target variable based on the feature extracted by the feature extraction execution unit 140 when there are variable names and target variables of the text data 110. For example, when there is a target variable, the feature information input unit 151 may input a value (predictive variable impact degree) indicating how much the feature extracted from the text has a target variable. For example, when the target variable name is'target' and the text variable name is'text', the influence on the target variable can be calculated and input for each of the three features'text_PCA_1','text_PCA_2', and'text_PCA_3'. . The influence calculation method used herein may include a tree-based calculation method or a p-value test method, but is not limited thereto.

According to one embodiment of the present application, the reference importance input unit 152 may set a reference value associated with the importance value based on user input information. In addition, the reference importance input unit 152 may input a reference value input by the user to extract words having a importance level higher than a specific value. The reference importance input unit 152 may receive a reference value associated with the importance value from a user terminal (not shown). The reference importance input unit 152 may provide a reference value setting menu to a user terminal (not shown). The user may input reference value information associated with the importance value using a user terminal (not shown). For example, the reference importance input unit 152 may set the reference value (default) to 0.01 when the user input information is 0.01. Also, when the user input information is not received, the reference important input unit 152 may set a reference value associated with a preset importance value. The reference value associated with the preset importance value may be 0.001%, 0.005%, 0.01%, 0.05%, 0.1%, 0.5%.

According to an embodiment of the present application, the importance extracting unit 153 for each word may extract an influence degree of each feature for each word vector for a plurality of features. The importance level extractor 153 for each word may extract the influence of each feature for each word vector for the three features extracted by applying principal component analysis to the word vector.

In addition, when a feature is extracted by using the principal component analysis, the importance extracting unit 153 for each word may extract an influence degree of each variable used in the principal component analysis on the extracted feature. For example, the importance extracting unit 153 for each word may extract the influence degree using PCA.components_ which is a function of deciposition of Scikit-Learn. When a feature is extracted using each PCA, this function can express how much each variable used in the PCA affects the extracted feature.

For example, when two variables M and N are extracted by applying PCA to variables A, B, C, D, and E, the effects shown in [Table 2] can be extracted.

A B C D E M -0.3 -0.4 0.5 0.7 0.1 N 0.1 -0.7 -0.4 -0.5 0.3

The influences of variables A, BC, D, and E on variables M are -0.3, -0.4, 0.5, 0.7, and 0.1, respectively, and the influences of variables N are 0.1, -0.7, -0.4, -0.5, and 0.3, respectively. . That is, the variable with the highest positive correlation in M is D, the variable with the highest negative correlation in B, and the variable with the highest positive correlation in N is E, the highest in negative correlation. The variable is B. The importance extracting unit 153 for each word may be generated such that 1 is added when each influence degree is squared. For example, the importance extracting unit 153 for each word is applied to each influence degree of the variable shown in Table 2. Multiplied by the square, we can derive as follows.

According to an embodiment of the present disclosure, the importance level calculation unit 154 for each word may convert the importance level as a percentage of the value extracted by the importance level extraction unit 153 for each word. The feature extraction unit 142 may represent a percentage of how each existing variable affects the feature extracted by the PCA based on the value extracted by the word-specific importance extraction unit 153. In addition, since each influence degree is calculated so that 1 is added when squared and added, the importance calculation unit 154 for each word can calculate the percentage by multiplying each influence degree and multiplying by 100.

If the influence degree is squared in Table 2, it can be expressed as Table 3, and if it is expressed as a percentage, it can be expressed as Table 4. The importance level extracting unit 153 for each word can determine which variables are important (which influenced a lot) for each feature.

A B C D E M 0.09 0.16 0.25 0.49 0.01 N 0.01 0.49 -0.16 0.25 0.09

A B C D E M 9% 16% 25% 49% One% N One% 49% 16% 25% 9%

According to one embodiment of the present application, the visualization execution unit 160 may visualize the importance of each word and information on the extracted features in the form of a Venn diagram graph. In addition, the importance of each word and information about the extracted features may be visualized using a word cloud method using a Venn diagram shape. In addition, the visualization execution unit 160 may determine the importance of each word and the frequency of the extracted features. Depending on the output location, output size and output color can be determined and visualized in the form of a cloud. Cloud-type visualization can visualize key words to stand out and express words used in articles in different sizes and colors according to frequency. In addition, the visualization execution unit 160 may provide a graph visualized in a cloud form by a user terminal (not shown).

2 is a diagram for explaining a visualization tool of a device for visualizing a key word of a feature according to an embodiment of the present application. Referring to FIG. 2, the visualization execution unit 160 may visualize key words of a feature by applying a word cloud in the form of a Venn diagram.

For example, referring to FIG. 2(a), there are key words for each extracted plurality of features (for example, for each of three features). For example, when a group of words is W1, W2, and W3, W1, W2, and W3 are composed only of different words, and W1, W2, and W3 may have words shared with each other. These shared words influenced each feature, so they can be considered important words. Therefore, the visualization execution unit 160 may visualize in the form of a venn diagram to indicate this. For example, the visualization execution unit 160 may determine and output the first area (eg, W1 area) of the first feature variable as the first color. In addition, the output position and output size of the word may be output differently based on the importance of the variable in the first area of the first feature variable, the importance of the key word, and the influence degree. Also, the visualization execution unit 160 may determine and output the second region (eg, W2 region) of the second feature variable as the second color. In addition, the output position and output size of the word may be output differently based on the importance of the variable in the second area of the second feature variable, the importance of the key word, and the influence degree. Also, the visualization execution unit 160 may determine and output the intersection region of the first feature variable and the second feature variable (for example, W1&W2) as the third color. In addition, the output position and output size of the word may be output differently based on the importance of the variable, the importance of the key words, and the influence of the variable in the intersection region of the first feature variable and the second feature variable (for example, W1&W2). .

In addition, referring to (b) of FIG. 2, a word cloud (word cloud) means to visually calculate the frequency of words used in a document. By using a word cloud, words that appear frequently appear large, so you can grasp the core content of the document at a glance. In addition, a word cloud (word cloud) is also called a tag cloud. Tags are tags attached to clothes and objects to describe materials and handling methods. Keywords that you attach to describe content on a web page or social network service (SNS) are called tags. The tag cloud displays the importance of tags on a website in text size or color. In addition, the work cloud is divided into a text cloud and a data cloud according to the nature of the content to be expressed. If a document cloud is a visual representation of a word contained in a document, the data cloud is a representation of numeric information in size and color instead of words. For example, depending on the size of the population, the size or color of a country is expressed differently, or the size and color of a company's name are determined by reflecting the fluctuations in stock prices and the trading volume in the stock market. In addition, key words are displayed for each area of the Venn diagram using a word cloud technique. In addition, the visualization execution unit 160 assigns importance to a value corresponding to the frequency, so that the higher the importance, the larger the word size is displayed. When assigning importance, substitute the proper conversion so that the percentage becomes an integer. In addition, since the visualization execution unit 160 has a different color of the background for each area of the Venn diagram, if the color of the word is different, the user may be uncomfortable to view it, so the color of the word may be uniformly provided in black.

FIG. 3 is a diagram visualizing a list of words used in features extracted from the first text data in the apparatus for visualizing key words of the feature according to an embodiment of the present application.

Referring to FIG. 3 as an example, the key word visualization device 150 of the feature may visualize the word list used in the feature by generating a word cloud in the form of a Venn diagram. The key word visualization device 150 of the feature may show words having a feature impact level higher than a reference value in a word cloud in the form of a Venn diagram.

The key word visualization device 150 of the feature uses a word cloud in the Venn diagram for words having an impact on the feature that is greater than or equal to a reference value for the three features'text_PCA_1','text_PCA_2', and'text_PCA_3' extracted when the text variable name is'text' It can be expressed using Here, the apparatus 150 for visualizing a key word of a feature may display words on the overlapping part (intersection) of the Venn diagram that simultaneously influenced many features simultaneously. In other words, the most central words are words that have influenced all three features. In addition, the key word visualization device 150 of the feature may show a variable information display box. In the variable information display box, the target variable name is'target' and the text variable name is'text'. The extracted three features'text_PCA_1','text_PCA_2', and'text_PCA_3' are each influenced by the target variable and the key words for each feature It may include the number of.

8 is a schematic block diagram of a text data feature extraction apparatus according to an embodiment of the present application. Referring to FIG. 8, the word extraction processor 130 may include a user dictionary input unit 131, its own dictionary input unit 132, a word candidate extraction unit 133, and a main word extraction unit 134.

According to an embodiment of the present application, the user dictionary input unit 131 may obtain input information related to a key word from a user. The user dictionary input unit 131 may build a user dictionary based on user input information input by a user through a user terminal (not shown). In addition, the user dictionary input unit 131 may build a user dictionary to extract the words when the user wants to extract words from text data. In other words, the user dictionary input unit 131 may build a user dictionary based on user input information input at least one of Hangul and English special characters through a user terminal (not shown). The words included in the user dictionary may be given the highest priority when extracting words from the main word extraction unit. In addition, the user dictionary may overcome limitations that may not be able to extract all words matching data characteristics with its own dictionary included in its dictionary input unit 132, and to put words that the user considers to be particularly important in the highest priority. It is a dictionary input by the user.

According to an embodiment of the present application, the user dictionary input unit 131 may provide a word input menu to a user terminal (not shown). For example, a user terminal (not shown) downloads and installs an application program provided by the user dictionary input unit 131, and a word input menu may be provided through the installed application.

The user may input words to be preferentially extracted through the word input menu. The user dictionary input unit 131 may build a user dictionary by collecting words input from a user terminal (not shown).

The user dictionary input unit 131 may transmit and receive data, contents, and various communication signals to and from a user terminal (not shown) through a network, and may include all types of servers, terminals, or devices having functions of data storage and processing. .

The user terminal (not shown) is a device that is linked to the user dictionary input unit 131 through a network, for example, a smart phone, a smart pad, a tablet PC, a wearable device, and a PCS (Personal Communication System). ), Global System for Mobile Communication (GSM), Personal Digital Cellular (PDC), Personal Handyphone System (PHS), Personal Digital Assistant (PDA), International Mobile Telecommunication (IMT)-2000, Code Division Multiple Access (CDMA)-2000 , W-CDMA (W-Code Division Multiple Access), all kinds of wireless communication devices such as Wibro (Wireless Broadband Internet) terminals and fixed terminals such as desktop computers and smart TVs.

Examples of networks for sharing information between the user dictionary input unit 131 and a user terminal (not shown) are 3rd Generation Partnership Project (3GPP) networks, Long Term Evolution (LTE) networks, 5G networks, and World Interoperability for Microwave Access (WIMAX). ) Network, Wired and Wireless Internet, Local Area Network (LAN), Wireless Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), Bluetooth Network, Wifi Network, NFC( Near Field Communication (Near Field Communication) network, satellite broadcasting network, analog broadcasting network, DMB (Digital Multimedia Broadcasting) network, and the like may be included, but are not limited thereto.

According to an embodiment of the present application, the dictionary input unit 132 may obtain a noun dictionary including a plurality of common nouns. As an example, the noun dictionary may be a dictionary including a noun owned by itself. For example, the number of words included in the noun dictionary is 371,151. Although the number of technical terms (for example, special words used for each professional or major) such as legal terms is small, the number of ordinary words may be included. In addition, the dictionary input unit 132 itself may acquire a database including a plurality of common nouns from an external server. The dictionary input unit 132 may acquire a specific noun dictionary from an external server (not shown). In other words, the dictionary input unit 132 may obtain a specific noun dictionary from an external server (not shown) in consideration of the characteristics (characteristics) of the received text data 110. For example, when the received text data 110 is judgment data, the own dictionary input unit 132 may obtain a noun dictionary including legal terms from an external server (not shown).

In addition, the dictionary input unit 132 itself may include a non-noun dictionary. Non-noun dictionaries are'~struck, ~struck, ~struck, ~so, ~struck, ~seen, ~none, ~shall, ~same, Sikkim, ~same, Yes, None, ~to, ~to, It consists of words with parts of speech that are not nouns, such as'if, where, not,' and vocabularies that cannot be nouns such as'??, ?O'. Non-noun dictionary can contain words that correspond to parts of speech and typos, not to nouns. In addition, a non-noun dictionary can be used to remove words that are definitely not nouns for words that are not 100% noun words.

In addition, the dictionary input unit 132 may include a person name dictionary and a place name dictionary. The human name dictionary may be a dictionary consisting of human names. Also, the place name dictionary may be a dictionary including a region name, a subway station, a city name, and the like. People's names, local names, subway stations, and famous places are words that may or may not be needed depending on the data. If the user deems it unnecessary, it can be removed.

According to an embodiment of the present application, the word candidate extracting unit 133 may separate a morpheme from text data and extract a morpheme that can be a word. As an example, the word candidate extracting unit 133 may perform morpheme analysis using text data and extract morphemes that can be words such as sentences and phrases. The word candidate extracting unit 133 may perform morphological analysis of Komoran (Korean Morphological Analyzer) to extract nouns from text data. Unlike conventional morpheme analysis, Komoran morpheme analysis can analyze multiple words as one part of speech, so that text data with spaces can be more accurately tokenized.

Korean is L + [R] structure. The distribution of R that appears next to L is a good hint to judge whether L is a noun or not. However, it is not possible to find nouns based on rules while having Rs as investigations. For example,'-silver' is a representative survey, but'Sonnaeun' is not'Sonna + silver'. In entertainment news,'A pink member Son Na-eun' frequently appears, but words such as'Sona Na + Eun','Sona Na + Lee', and'Sona Na + To' do not appear frequently. In this way, the word extraction processor 130 may use the information of the bipartite graph of L-R to determine whether the word is a noun or not. The word extraction processor 130 collects the R sets appearing after the noun using the Sejong corpus, and the words in the R set are learned with a noun possible score. The range of noun possible scores is [-1, 1]. For example, the word "out" is -0.530702, the word "being" is 1.000000, the word "with" is 0.327824, the word "writing" is 0.079298, and the word "was" is -1.000000 , And the word "was" was 0.437399, and if "was + was said", "had" had a noun score of -1.0. If'Fun + Have' 3 times and'Fun + Have' 2 times, the fun noun score is (3 x 1.0 + 2 x 0.33) / 5 = 0.732. If the threshold of noun extraction is 0.5,'fun' can be extracted as a noun. In other words, the word candidate extracting unit 113 may extract a specific noun based on a threshold of preset noun extraction.

This method can learn the structure of words in a given document set and extract nouns from the given document set. That is, it is a statistical-based unsupervised learning method that does not require learning data. Therefore, there is an advantage of extracting nouns that have not been registered in advance. The word candidate extracting unit 133 of the present application may extract a noun candidate using a noun extraction function (for example, the LRNounExtractor_v2 function of soynlp).

Also, the word candidate extracting unit 133 may extract a morpheme that can be a word from two or more Hangul words. Among the words extracted using a specific function (for example, the LRNounExtractor_v2 function), there are words in Korean characters in addition to nouns, words containing special characters or foreign languages, and words that are not nouns. In addition, there may be words that the user wants to extract among words that the function does not extract, and among the extracted words, there may be a key word that the user thinks more particularly. The word candidate extracting unit 133 may improve the quality of words using a dictionary prepared for vectorization by leaving only the main words. The word candidate extracting unit 133 may separate morphemes from the text data 110 and extract nouns. In addition, the word candidate extracting unit 133 may extract a morpheme composed of only two or more Hangul characters from the text data 110.

For example, the word candidate extracting unit 133 may perform a task of leaving only two or more Hangul words. Since there are many meanings of Hangul characters, and there are many cases of investigation, the word candidate extracting unit 133 may exclude Hangul characters. Further, the word candidate extracting unit 133 may exclude (delete) the word when the special data or the foreign language is mixed in the text data in order to use only the word input by the user.

According to an embodiment of the present application, the key word extracting unit 134 may extract the key word using a plurality of dictionaries from the morpheme list extracted by the word candidate extracting unit 133. In other words, the main word extraction unit 134 uses a user dictionary, its own noun dictionary, a non-noun dictionary, a human name dictionary, and a geographical name dictionary from the morpheme list extracted by the word candidate extraction unit 133 To extract words.

Also, the main word extractor 134 may extract words that are nouns from text data using a noun dictionary. The key word extracting unit 134 may extract words that are nouns from words extracted using dictionaries of other konlpy such as kkma, mecab, and built-in noun dictionaries. The key word extracting unit 134 may extract words that are nouns from text data using a noun dictionary regardless of whether or not a user dictionary is present. These extracted words can be used as 100% nouns. The noun dictionary holds 31,151 words. The number of terminology (e.g., special words used by professions and majors such as legal terms) is small, but the number of everyday words is secured.

Also, the main word extracting unit 134 may remove words that are not nouns using a non-noun dictionary. The key word extracting unit 134 may delete (remove) words that are not 100% noun words using a non-noun dictionary. At this time, the non-noun dictionary is'~struck, ~sanda, ~spooky, ~, so, ~s, ~ss, ~nones, ~tos, ~tos, shikim, ~tos, yess, nones, ~tos, ~ It consists of words with parts of speech that are not nouns, such as'if, if, where, not,' and vocabularies that cannot be nouns such as'??, ?O'. After filtering through a non-noun dictionary, the remaining words can be regarded as new words, synthetic nouns, jargon, etc. obtained through unsupervised learning and can be used.

For example, people's names and local names, subway stations, and famous places are words that may or may not be needed depending on the data. The key word extracting unit 134 may delete (remove) the word based on the user's input information. The key word extracting unit 134 may remove (delete) words by applying a human dictionary and a geographical name dictionary when additionally wanting to remove a person's name and a place name for a removed word collection using a non-noun dictionary.

Also, the main word extracting unit 134 may remove words using a user dictionary. The user dictionary may be a dictionary in which words the user determines are important in extracting the corresponding text feature.

In addition, the main word extracting unit 134 may give a first priority to words included in a user dictionary among a plurality of dictionaries. Also, the main word extracting unit 134 may give a second priority to words included in the noun dictionary among the plurality of dictionaries. The key word extracting unit 134 may extract the key word from the list of words given priority. For example, if the word included in the user dictionary among the plurality of dictionaries exists in the morpheme list extracted by the word candidate extracting unit 133, the main word extracting unit 134 may give the word the first priority. . In addition, when the word included in the noun dictionary among the plurality of dictionaries exists in the morpheme list extracted by the word candidate extracting unit 133, the main word extracting unit 134 may give the word a second priority. The key word extracting unit 134 may extract the final key word based on the priority given to the morpheme list extracted by the word candidate extracting unit 133. For example, the key word extracting unit 134 may extract words within a preset rank as key words. Meanwhile, when a word included in a user dictionary among a plurality of dictionaries exists in the morpheme list extracted by the word candidate extracting unit 133, the main word extracting unit 134 may give a first weight to the corresponding word. In addition, when the word included in the noun dictionary among the plurality of dictionaries exists in the morpheme list extracted by the word candidate extracting unit 133, the main word extracting unit 134 may assign a second weight to the word. Here, the first weight value may be higher than the second weight value. The key word extracting unit 134 may re-generate the weighted word list and extract the key word.

According to one embodiment of the present application, the word vector generator 141 may vectorize the key words by applying the key words extracted by the key word extractor 134 to a text analysis algorithm. The word vector generator 141 may vectorize each word by applying a text analysis algorithm (eg, TF-IDF technique) to the words extracted by the main word extractor 134. As an example, the word vector generator 141 may vectorize key words using a vector extraction package (eg, feature_extraction.text, which is one of subpackages of feature_extraction of Scikit-Learn). A vector extraction package (e.g., feature_extraction.text, one of the subpackages of Scikit-Learn's feature_extraction) provides a document preprocessing class, which is a text analysis function (e.g. a function that vectorizes words based on frequency) For example, the key nouns can be vectorized using the TfidfTransformer function.

As an example, the word vector generator 141 may calculate the frequency of words included in the document. The frequency of words in text data (document) may be expressed as Equation (1). Here, d may be text data (document) and t may be a word. That is, Equation 1 may mean the frequency of the word t in the document d. The value of TF can be normalized depending on the situation.

[Equation 1]

In addition, the word vector generator 141 may use a Boolean frequency to indicate 1 when the text data 110 contains a word and 0 when the word is not included. As an example, the so called frequency may be expressed as Equation 2.

[Equation 2]

For example, the word vector generation unit 141 prevents the value of the TF (Term Frequency) from becoming excessively large by giving up the case where the word or the word data is not included in the text data 110 using the called frequency. can do. On the other hand, since the word vector generator 141 assigns the same weight to the text data 110 as 1 or 100 times, the TF is not important, that is, only the presence or absence of the word is important. When the frequency called when is applied, the word can indicate whether the word appears or not.

In addition, the word vector generator 141 may express a logarithmically scaled frequency as shown in Equation (3). The logarithmic scale frequency may mean a value converted to a lock scale to reduce the size. At this time, if the difference in the frequency of words appearing in the text data 110 is small, the change in the TF value is large, but when the frequency of words increases innumerably, there is almost no difference in the TF value.

[Equation 3]

In addition, the word vector generator 141 may adjust the relative frequency value of the word according to the length of the text data 110. In other words, augmented frequency means a method of adjusting the relative frequency value of a word according to the length of the document, and dividing the word frequency by the maximum value among the word frequencies of words in the document. The frequency of increase can be expressed by Equation 4. Equation 4 may be an equation adjusted to reduce the deviation.

[Equation 4]

For example, the word vector generator 141 may fix the denominator to 100 when the word A appears 100 times in the first document, 5 times in the second document, and 8 times in the third document.

IDF (inverse document frequency) may mean a number inversely proportional to the number of documents containing a specific word. IDF (inverse document frequency) can be used to reduce the importance of words that frequently occur in the entire document. At this time, if the inverse document frequency is increased, the IDF (inverse document frequency) value increases exponentially as the total number of documents increases. IDF (inverse document frequency) can be expressed as Equation (5).

[Equation 5]

TF-IDF encoding is a method of increasing the importance (TF) frequently occurring in one document and decreasing the importance of frequently occurring words in the entire document. TF-IDF may be expressed as Equation (6).

[Equation 6]

For example, Equation 6 is

값(중요도)이 커지게 됨을 의미할 수 있다.

It can mean that the value (importance) becomes large.

For example, the word vector generator 141 may receive text data 110 and assign numbers to words included in the text data 110.

Corpus = ['This is the first document.','This is the second second document.','And the third one.','Is this the first document?']

Number by word {'and': 0,'document': 1,'first': 2,'is': 3,'last': 4,'one': 5,'second': 6,'the' : 7,'third': 8,'this': 9}

The word vector generator 141 may generate a matrix by applying only TF values. Here, the row may indicate a number assigned to each word, and the column may indicate a sentence.

When taking only TF values: array([[0, 1, 1, 1, 0, 0, 0, 1, 0, 1],

[0, 1, 0, 1, 0, 0, 2, 1, 0, 1],

[1, 0, 0, 0, 0, 1, 0, 1, 1, 0],

[0, 1, 1, 1, 0, 0, 0, 1, 0, 1])

In addition, the word vector generator 141 may generate a matrix by applying the T TF-IDF value. Here, the row may indicate a number assigned to each word, and the column may indicate a sentence.

Array ([[0., 0.3894, 0.5577, 0.462, 0., 0., 0., 0.3294, 0., 0.4629],

[0. , 0.2415, 0., 0.2870, 0., 0., 0.8573, 0.2042, 0., 0.2870],

[0.5566, 0., 0., 0., 0., 0.5566, 0., 0.2652, 0.5566, 0.],

[0. , 0.3894, 0.5577, 0.4629, 0., 0., 0., 0.3294, 0., 0.4629]

The example described above is only an example, and is not limited thereto. It may be present in more various embodiments.

According to one embodiment of the present application, the feature extraction unit 142 may extract a feature using a principal component analysis on a vectorized word in the word vector generation unit 141. The feature extraction unit 142 may extract three new features by dimensionally reducing the word vector in three dimensions. Principal component analysis (PCA) is a technique that rotates a data set so that variables are not statistically correlated. Depending on how important it is to describe your data after rotation, you often choose only a few of the new traits. This new property is called the main component (each main component is influenced by all variables in the existing data), and if only a part is selected, it is used for the purpose of dimension reduction because the dimension is reduced.

In addition, the feature extracting unit 142 has a problem in that performance may not be good in predicting when the words are vectorized and all of them are used as features. Therefore, features are extracted using principal component analysis. The number of extracted features is easy to visualize, and is generally set to a preset number (e.g., 3) that can express the variance of each vector by 80% to 90%.

Also, the feature extraction unit 142 may generate the name of the feature to be extracted. For example, when the text variable name included in the existing text data 110 is'text', the names of the three extracted features may be sequentially generated as'text_PCA_1','text_PCA_2', and'text_PCA_3'.

According to one embodiment of the present application, when the feature extraction unit 142 extracts a feature using principal component analysis, each variable used in principal component analysis may extract an influence degree on the extracted feature. For example, the feature extraction unit 142 may extract the influence degree using PCA.components_ which is a function of deciposition of Scikit-Learn. When a feature is extracted using each PCA, this function can express how much each variable used in the PCA affects the extracted feature.

For example, the text feature extraction device 120 may receive text data 110. The text data 110 may include text shown in Table 5. For example, the text data 110 may include court judgment data.

precedent Text <Applicant, Judge Requester> <An opponent, Judge Requester> <Judgement> Daegu High Court 1981.3.31. Decision 80 Le 93 Decision <Order> Dismissed the appeal permission application. <Reason> The appeal permission application provided in Article 12 of the Special Act on Promotion of Litigation is applied to the civil action case. The application is clear in light of the provisions of Article 2 of the Transfer Act (applicants are also filing appeals with party member 81M43). Decide <Defendant> <Sendee> Defendant <Judgement> Daegu District Court 1990.2.2. Judge 89 No 1675 Judgment <Order> Dismisses the appeal. <Reason> Judging by the defendant's reason for appeal, Article 66 of the National Civil Service Act, Clause 1 of the Constitutional Law, Article 11, Paragraph 1, Article 21, Paragraph 1, There is no reason to dismiss the defendant's appeal, as it cannot be regarded as a law in violation of Article 31 (4), Article 33 or Article 37 (2) (see Judges, April 1990, Decision 90°332). In accordance with the opinions of the concerned judges, the judgment is made as in the order. <Current debtor> Gosung Dyne Laser Co., Ltd. <Management Inn> <Order> abolished the rehabilitation procedure for this case. <Reason> Submitted by the manager on May 31, 2010, and July 26, 2010 The rehabilitation plan approved for amendment was rejected due to the failure to obtain the consent of the provisions of Article 237 of the Act on Rehabilitation and Bankruptcy of the Debtor at a meeting related to the resolution of the rehabilitation plan held on July 26, 2010. Decisions are made in accordance with Article 286 (1) 2 of the Act on Rehabilitation and Bankruptcy. <Respondent> <Appeal Person> Prosecutor <Prosecutor> Bae Seong-Hun <Respondent Law> Judicial Trainee Min-Geun Cho (Kuksun) <Judgment Judgment> Seoul Southern District Court 2011. 4. 15. Decision 2010 High Court 4002 Judgment <Order> The prosecutor's appeal is dismissed. <Reason> The prosecutor argues that as a ground for appealing this case, the accused that the accused prescription is suspended while the accused is in Japan, the accused prescription is suspended, as judged by the ruling. Can't see Therefore, there is no reason to appeal the prosecutor's office, so it is dismissed pursuant to Article 364 (4) of the Criminal Procedure Law <Shin-Chung-In> Director of Dongdaegu Tax Office <Sang Dae Bang> Kim Kyung Jo <Judgment Judgment> Daegu High Court 1981.2.10. Decision 80 Order 200 Declaration <Order> Dismissed the appeal permission application. <Reason> The application for permission to file an appeal under Article 12 of the Special Act on Promotion of Litigation applies to the civil action case, and it applies to administrative action cases such as this. It is clear in light of the provisions of Article 2 of the Transfer Act, so the application for this case, which filed an appeal for an administrative lawsuit, is inappropriate and is decided to dismiss it as an order in accordance with the opinions of the concerned judges. <Resident Complaint> Lee Sang-hak <Won Sim> Seoul High School October 6, 1960 sentenced to 60 Min Gong 178, 179 <Reason> Under the premise, Article 34 of the Auction Act states that in the case of real estate auction under the Act, It is illegal that the auction court ordered the bidding in person without a stakeholder's right to authorize the meridian in spite of the fact that the auction court could not bid by the ex officio. Centrifugal decisions that cannot be overlooked and that have been overlooked are also illegal <Applicant> <Respondent> Minister of Education and Human Resources Development (Representative Jong-In Kim, Deputy Attorney) <Order> dismisses all cases. <Purpose of Application> Respondent filed on Feb. 5, 2004 Applicants Jung Hee, Byung-Ki Son, In-Ho Bae, The disposition for the removal of the approval of the executive office for Eun-Hyun Yoon, Sung-Kyun Jung, and Jin-Bok Lee, and the external statement made on the same day, Lee Jeong-Jung, In-Ki Kim, Byeong-Chan Kim, Jong-Chan Kim, Young-Shin Kang, Young-Shin Kang, and Jeong-Gi-Ol School were appointed as temporary directors of the Yushin Academy. The validity of the case is suspended until the judgment of the case is declared. <Shin Cheong-In> Shin Bun-Sik <Sang Dae-Bang> Nam-gu Office Director <Director Judgment> under direct management of Busan, Daegu High Court 1982.8.17. Decision 81 ward 120 Decision <Order> Dismisses the appeal permission application. <Reason> The appeal permission application provided for in Article 12 of the Special Act on Promotion of Litigation applies to civil lawsuits and is subject to administrative action cases such as this. It is clear in light of the provisions of Article 2 of the Transfer Act, so the application for this case, which filed an appeal for an administrative lawsuit, is inappropriate and is decided to dismiss it as an order in accordance with the opinions of the concerned judges. <Plaintiff, Defendant> Yoo Sang-rok and 1 other <Defendant, Defendant> Seoul Metropolitan Government Commissioner <Defendant Assistant, Defendant> Joo Gil-Hwan <Preliminary Judgment> 1st Seoul High School 1956. 8. 16. According to the U.S. record, the original judgment confirmed the fact that the plaintiffs first filed a lease contract for the defendant against the land where the plaintiffs were in an unmanaged state, and the plaintiffs judged that the plaintiffs were the owners of the land, but only with the facts Since it is said that the right to rely on the property belonging to it cannot occur, there is a misconduct in the original ruling that misinterprets the law. <The Defendant> <The Defendant> The Defendant <Judgement> Seoul Court of Justice 1991.7.24. Decision 91 No 2429 Decision <Order> Dismisses appeal. <Reason> Sees the reason for appeal. According to the evidence of the first judgment of the original referee, the defendant paints the advertisement sign owned by another person with white paint, and then the advertisement text. There is no violation of the law of misunderstanding or the law of debt in a centrifugal judgment because it constitutes a property damage crime if it is possible to admit that it has been erased. Adjudicate

The word extraction processor 130 may extract key words using a dictionary after morpheme analysis on the judgment data'precedentText' shown in Table 5. The feature extraction execution unit 140 may vectorize the extracted key words in a TF-IDF manner, and then extract three features by applying a principal component analysis to the vector. Here, the number of judgments used is 20000, the words used in the user dictionary are 9400 statute terms, the number of words in the noun dictionary is 31,7151, the number of words in the non-noun dictionary is 9682, and the human name ( The number of words in the 人名) dictionary may be 3467,5174, and the number of words in the geographical name dictionary may be 50675.

In addition, the number of words extracted by the word candidate extracting unit 133 is 104887, and some of them are as follows.

('"Feraud" part, "For suspension of architect's business", ""Design Academy Light Sales"", "Request for trial to confirm the scope of right", "Judgment for refusal to object to refusal of judgment", "Reason for judgment." Korea Urban Development Co., Ltd., '1966.11.23','Article 95, Article 100','Application for permission for river branch office','Without rampant','Branch material','Interim registration','Appropriate use', 'Deposit issuance','Survey evaluation','Loan available', …,'Haetae business','Happy driver','Danghae business','Correction possible','Early food','Hearing place', '5 days ago In','Estimated Damage','Contract Number','Obtained Lease','Earth 29', '197 pyeong','Permitted','Submission of Objection','Single Debt')

In addition, the number of words extracted by the main word extraction unit 134 is 101625, and some of them are as follows.

('Subtraction','Virtual Quarantine Certificate','Price Decision','Price Decision','Price Standard','Price Trend','Price Change','Price Change Index','Price Comparison','Price Eval' ','Price level','Price start','Price stabilization','Price information', …,'Hope price','Hope pig','Hope pig savings','Hope department store','Hope application', 'Application for Hope','Rare Hame','Hiropon Ingredients','Hiropon Manufacturing','Hiropon Manufacturing Ingredients','Hiropon Injection')

4A to 4C are diagrams of visualizing a word list used for features extracted from second text data in a device for visualizing a key word of a feature according to an embodiment of the present application.

According to one embodiment of the present application, the second text data may be judgment data'precedentText'. The key word visualization device 150 of the feature performs principal component analysis on the vectorized word obtained through morphological analysis, word extraction, and vectorization of the word as shown in Table 5, and which word affects the extracted feature You can visualize how crazy you are in the word cloud form of the Venn diagram. The morpheme analysis, word extraction, and vectorized words obtained through vectorization of words may be performed by the text feature extraction device 120 that performs principal component analysis. Text The text feature extraction device 120 extracts the key words after morpheme analysis on the judgment data'precedentText', vectorizes the words in a TF-IDF method, and then applies the principal component analysis to the vectors to extract three features Can. At this time, the number of rulings used may be 20000, and the number of words used when extracting features may be 13185. In addition, the key word visualization device 150 of the feature may visualize and provide a vowel collection of 0.01%, 0.05%, 0.1%, 0.5%, 1% or more, respectively, and the importance shown in FIGS. 4A to 4C It is possible to intuitively grasp which word has a high influence according to the reference value of.

For example, FIG. 4A(a) may visualize a word having a feature influence degree of 0.01% or more in the second text data in a word cloud method in the form of a Venn diagram. Also, FIG. 4A (b) may be a visualization of a word having a feature influence of 0.05% or more in the second text data in the form of a Venn diagram word cloud. FIG. 4B(c) may visualize a word having a feature impact ratio of 0.1% or more in the second text data in a word cloud method in the form of a Venn diagram. Also, FIG. 4B (d) may be a visualization of a word having a feature impact of 0.5% or more in the second text data in a word cloud form in the form of a Venn diagram. In addition, FIG. 4C(e) may visualize a word having a feature influence of 1% or more in the second text data in a word cloud method in the form of a Venn diagram.

5A to 5C are diagrams of visualizing a word list used for features extracted from third text data on a device for visualizing a key word of a feature according to an embodiment of the present application. According to one embodiment of the present application, the third text data may include Naver movie review data. Text The text feature extraction device 120 may perform principal component analysis on the vectorized words obtained through morpheme analysis, word extraction, and vectorization of words using the third text data. The third text data may be data generated by collecting 15931 reviews registered in the Naver movie review. Text The text feature extraction device 120 may generate a user dictionary by directly entering keywords indicating positive and negative in a review based on user input information. The user dictionary may include a total of 393 words. This word contains not only nouns, but also parts of speech, such as adjectives and verbs, and may include new words (eg: honey jam, yes jam, etc.). Text The text feature extraction device 120 uses all of its own noun dictionaries, non-noun dictionaries, human dictionaries, and nominated dictionaries when extracting nouns. After reflecting the user dictionary, the total number of extracted words is 2308. The key word visualization device 150 of the feature may generate a word count distribution for each feature as shown in Table 6. In other words, Table 6 may be generated based on the importance section of the word count distribution for each feature in the third text data. The high-priority word for each feature may include viewing (47%), spectators (47%), fun (1%), and movies (1%) in the first feature (review_PCA_1). Also, the key words of the second feature (review_PCA_2) may include a movie (74%), funny (7%), acting (3%), too (2%), and viewing (1%). In addition, the key words of the third feature (review_PCA_3) are funny (39%), funny (36%), funny (19%), movie (3%), and not funny (1%). You can see that the words related to'movie, fun, and watching' were of high importance.

Importance section review_PCA_1 review_PCA_2 review_PCA_3 0 or more and less than 0.001 2120 2138 2228 0.001 or more but less than 0.005 139 88 43 0.005 to less than 0.01 18 27 12 0.01 to 0.05 20 26 15 0.05 or more and less than 0.1 4 11 2 0.1 to less than 0.3 One 9 One 0.3 to 0.5 One 2 2 0.5 or more but less than 1 One 2 0 1 to 10 2 4 2 10 or more and less than 20 0 0 One 20 to 30 0 0 0 30 to 40 0 0 2 40 to 50 2 0 0 50 or more 0 One 0

As an example, FIG. 5A (a) may be a visualization of a word having a feature influence of 0.001% or more in a third text data in a word cloud form of a Venn diagram. In addition, FIG. 5B(b) may be a visualization of a word having a feature influence of 0.005% or more in the third text data in a word cloud method in the form of a Venn diagram. FIG. 5B(c) may visualize a word having a feature impact degree of 0.01% or more in the third text data in a word cloud form in the form of a Venn diagram. Further, FIG. 5B(d) may visualize a word having a feature impact factor of 0.05% or more in the third text data by using a word cloud method in the form of a Venn diagram. In addition, FIG. 5C(e) may visualize a word having a feature influence of 0.1% or more in the third text data in a word-cloud form of a Venn diagram. In addition, (f) of FIG. 5C may visualize a word having a feature impact degree of 0.5% or more in a third text data in a word cloud form of a Venn diagram shape. FIGS. 6A to 6C are diagrams according to an embodiment of the present application. This is a diagram visualizing the word list used for the feature extracted from the fourth text data in the feature's main word visualization device.

According to one embodiment of the present application, the fourth text data may be Naver news title data. The key word visualization device 150 of the feature performs principal component analysis on the vectorized word obtained through morpheme analysis, word extraction, and vectorization of the fourth text data, and which word has a significant influence on the feature You can visualize whether it is a Venn diagram-shaped word cloud. The morpheme analysis, word extraction, and vectorized words obtained through vectorization of words may be performed by the text-text feature extraction device 120 that performs principal component analysis. Text The text feature extracting device 120 extracts key words after morphological analysis on the fourth text data, vectorizes the words in a TF-IDF method, and applies three components to the vector to extract three features have. In this case, the text text feature extraction device 120 may extract a noun using only a noun dictionary, a noun dictionary, a life dictionary, and a noun dictionary without using a user dictionary. The fourth text data used herein may include data generated by collecting 96,000 Naver News titles. Text The text feature extraction device 120 may extract a total of 5324 words after reflecting the user dictionary.

The key word visualization device 150 of the feature may generate a word number distribution for each feature as shown in Table 7. In other words, Table 7 may be generated based on the importance section of the word count distribution for each feature in the fourth text data. The first feature (text_PCA_1) is weather (26%), dust (18%), fine (17%), fine dust (17%), tomorrow (3%), nationwide (3%), heat (2%), Can include key words from Seoul (1%). In addition, the key words of the 2nd feature (text_PCA_2) are key words such as talks (37%), summit (18%), Trump (12%), North America (12%), inter-Korean (5%), president (2%), etc. Words can be included. In addition, the third feature (text_PCA_3) is weather (19%), fine (13%), dust (13%), fine dust (13%), Seoul (6%), Korea (5%), nationwide (4%) ), heat waves (4%), tomorrow (3%), heat (3%), cold waves (1%) can include the key words. It can be seen that text_PCA_1 and text_PCA_3 have high importance for words related to weather, and text_PCA_2 has high importance for words related to the North American summit and North-South summit. In addition, the key word visualization device 150 of the feature may visualize and provide a vowel collection of 0.01%, 0.05%, 0.1%, 0.5%, 1% or more, respectively, and the importance shown in FIGS. 4A to 4C It is possible to intuitively grasp which word has a high influence according to the reference value of.

Importance section text_PCA_1 text_PCA_2 text_PCA_3 0 or more and less than 0.001 4848 4699 4849 0.001 or more but less than 0.005 286 383 256 0.005 to less than 0.01 52 88 56 0.01 to 0.05 80 102 84 0.05 or more and less than 0.1 12 20 25 0.1 to less than 0.3 25 18 32 0.3 to 0.5 10 8 5 0.5 or more but less than 1 3 0 6 1 to 10 4 2 7 10 or more and less than 20 3 3 4 20 to 30 One 0 0 30 to 40 0 One 0 40 to 50 0 0 0 50 or more 0 0 0

As an example, FIG. 6A (a) may be a visualization of a word having a feature influence of 0.001% or more in the fourth text data in a word cloud form of a Venn diagram. In addition, FIG. 6B (b) may be a visualization of a word having a feature impact of 0.005% or more in the fourth text data in a word cloud form of a Venn diagram. FIG. 6B(c) may be a visualization of a word having a feature impact of 0.01% or more in the fourth text data, using a word cloud method in the form of a Venn diagram. In addition, FIG. 6B(d) may visualize a word having a feature impact factor of 0.05% or more in the fourth text data in a word cloud method in the form of a Venn diagram. In addition, FIG. 5C(e) may visualize a word having a feature influence of 0.1% or more in the fourth text data in a word cloud method in the form of a Venn diagram. In addition, (f) of FIG. 5C may visualize a word having a feature impact degree of 0.5% or more in the fourth text data in a word cloud form in the form of a Venn diagram. FIGS. 7A to 7C are diagrams according to an embodiment of the present application. This is a diagram visualizing a list of words used for features extracted from the fifth text data in the feature's main word visualization device. According to an embodiment of the present application, the fifth text data may include judgment data. The fifth text data may be data including 20,000 rulings. The key word visualization device 150 of the feature performs principal component analysis on the vectorized words obtained through morpheme analysis, word extraction, and vectorization of words for the fifth text data containing 20,000 rulings, and which words are extracted to the extracted features You can visualize how much influence you have on the features in a word-cloud like Venn diagram. The morpheme analysis, word extraction, and vectorized words obtained through vectorization of words may be performed by the text-text feature extraction device 120 that performs principal component analysis. Text The text feature extraction device 120 extracts a key word after morphological analysis on the fifth text data containing 20,000 rulings, vectorizes the word in TF-IDF method, and applies the principal component analysis to the vector to feature Three can be extracted. Text The text feature extraction device 120 may select a user dictionary as a dictionary consisting of a collection of legal terms using the '2015 Korean-English Standard Terminology Glossary' produced by the Korea Legal Research Institute and perform a noun extraction process. There are 9,400 legal terms included in the user dictionary, 1965 legal names, and 3809 organization names and occupations, among which 9400 legal terms can be input into the user dictionary to perform feature extraction. When extracting nouns, we used all of our own noun dictionaries, non-noun dictionaries, human dictionary, and geographical names dictionary. After reflecting the user dictionary, the total number of extracted words is 13185. In addition, the key word visualization device 150 of the feature may visualize and provide a vowel collection of 0.01%, 0.05%, 0.1%, 0.5%, 1% or more, respectively, and the importance shown in FIGS. 7A to 7C It is possible to intuitively grasp which word has a high influence according to the reference value of. The key word visualization device 150 of the feature may generate a word count distribution for each feature as shown in Table 8. In other words, Table 8 may be generated based on the importance section for the number of words for each feature in the fifth text data. The first feature (precedentText_PCA_1) is determined (34%), appeal (33%), auction (4%), appealer (3%), reason for appeal (2%), burden (2%), cost (2%) , The facts (1%), meridians (1%) can include the key words. In addition, the second feature (precedentText_PCA_2) is deliberate (11%), litigation (9%), postmark (8%), burden (5%), cost (5%), fact (3%), litigation agency (3 %), litigation agent (3%), agent (3%), agent (3%), detention (3%), number of days of detention (3%), income (3%), prosecutor (2%), pleading (2) %), Closing of Pleadings (2%), Consumption (2%), Imprisonment (2%), Civil Litigation (2%), Crime (2%), Prosecution (2%), Resolution (1%), Appealants ( 1%) and litigation costs (1%). In addition, the 3rd feature (precedentText_PCA_3) is postmark (15%), appeal (12%), decision (9%), appeal (4%), closing pleading (4%), consumption (4%), resolution (3) %), Protection (3%), Pleading (3%), Auction (2%), Description (2%), Request for Protection (2%), Text (1%), Detention (1%), Days of Detention (1) %), registered (1%), input (1%), ownership (1%), litigation cost (1%), prosecution (1%), intention (1%), crime (1%) It can contain. Through this, it can be seen that the words related to a specific judgment are not seen much, and the importance of frequently used words is generally high.

Importance section precedentText_PCA_1 precedentText_PCA_2 precedentText_PCA_3 0 or more and less than 0.001 12350 12155 12486 0.001 or more but less than 0.005 392 521 362 0.005 to less than 0.01 120 157 90 0.01 to 0.05 201 219 138 0.05 or more and less than 0.1 53 39 30 0.1 to less than 0.3 27 51 39 0.3 to 0.5 15 16 13 0.5 or more but less than 1 12 12 9 1 to 10 13 12 17 10 or more and less than 20 One 3 0 20 to 30 One 0 0 30 to 40 0 0 0 40 to 50 0 0 One 50 or more 0 0 0

As an example, FIG. 7A (a) may be a visualization of a word having a feature influence of 0.001% or more in the fifth text data in a word cloud form of a Venn diagram. In addition, FIG. 7B (b) may be a visualization of a word having a feature influence of 0.005% or more in the fifth text data in a word cloud form in the form of a Venn diagram. 7B (c) may be a visualization of a word having a feature impact of 0.01% or more in the fifth text data, using a word cloud method in the form of a Venn diagram. Further, FIG. 7B(d) may be a visualization of a word having a feature influence of 0.05% or more in the fifth text data in a word cloud method in the form of a Venn diagram. In addition, FIG. 5C(e) may visualize a word having a feature influence of 0.1% or more in the fifth text data in a word cloud method in the form of a Venn diagram. In addition, (f) of FIG. 5C may visualize a word having a feature impact degree of 0.5% or more in the fifth text data in a word cloud method in the form of a Venn diagram. FIG. 9 is a text data device according to an embodiment of the present application This is a schematic diagram schematically illustrating the feature extraction process.

For example, referring to FIG. 9, the text feature extraction device 120 may receive text data 110. The text feature extraction device 120 may extract nouns from the input text data 110. Also, the text feature extraction apparatus 120 may generate a Hangul noun candidate of 2 or more characters from the extracted noun. The text feature extraction device 120 may extract a noun in the dictionary, that is, a noun included in its own dictionary, from the generated Hangul noun candidate. Also, the text feature extraction apparatus 120 may filter nouns that are not in the dictionary from noun candidates, that is, nouns that are not included in the dictionary. In this case, the filtering may be performed using a non-noun dictionary, a person dictionary, and a place name dictionary, and the text feature extraction apparatus 120 filters to remove the noun included in the noun dictionary, the person dictionary, and the place name dictionary in the noun candidate You can do The text feature extraction device 120 may generate a list of major nouns based on the nouns included in the user dictionary, its own noun dictionary, and filtered noun candidates. In this case, the text feature extraction apparatus 120 may generate a list of main nouns so that nouns (words) included in the user dictionary are placed in the highest priority. Also, the text feature extraction device 120 may generate a list of major nouns such that nouns (words) included in the noun dictionary are placed after nouns (words) included in the user dictionary. The text feature extraction apparatus 120 may vectorize a text noun list by applying a key noun list to a text analysis algorithm (eg, TF-IDF). The text feature extraction device 120 may extract a preset number (eg, 3) of features from the vectorized noun.

Hereinafter, based on the details described above, the operation flow of the present application will be briefly described.

10 is an operation flowchart of a method for visualizing a key word of a feature according to an embodiment of the present application.

The key word visualization method of the feature illustrated in FIG. 10 may be performed by the key word visualization device 150 of the feature described above. Therefore, even if omitted, the description of the feature's main word visualization device 150 may be equally applied to the description of the feature's main word visualization method.

In step S101, the key word visualization device 150 of the feature may receive an importance value for a target variable of a plurality of features extracted using principal component analysis from text data.

In step S102, the key word visualization device 150 of the feature may set a reference value associated with the importance value based on user input information.

In step S103, the apparatus 150 for visualizing a key word of a feature may extract an influence degree of each feature for each word vector for a plurality of features.

In step S104, the apparatus 150 for visualizing a key word of a feature may convert the importance as a percentage of a value obtained by extracting an influence on each feature for each word vector for a plurality of features.

In step S105, the key word visualization device 150 of the feature may visualize the importance of each word and information about the extracted feature in the form of a Venn diagram graph.

In addition, although not shown in the drawing, before the step S101, the key word visualization device 150 of the feature may receive a plurality of feature information extracted by the text feature extraction device 120.

According to an embodiment of the present application, the text feature extraction device 120 may receive text data 110. For example, the text feature extraction device 120 may receive text data 110 from an external server (not shown). Also, the text feature extraction device 120 may receive text data 110 from a user terminal (not shown). In addition, the text feature extraction device 120 may analyze the morpheme in the text data 110. The text feature extraction device 120 may perform morpheme analysis by applying text data 110 composed of sentences to a morpheme analysis algorithm. Morphological analysis is the first step in the analysis of natural language, which means converting an input string into a morphological column. Morpheme is the smallest unit of meaning and refers to the smallest element of meaning that can no longer be analyzed. In addition, a morpheme refers to a word or part of a word that represents a grammatical or relational meaning. The Korean morpheme analysis may be subjected to preprocessing, grammar morpheme separation, body analysis, verbal analysis, single morpheme analysis, compound word estimation, omission of investigation, verbal processing, and post-processing. In this case, morpheme analysis may be performed on text data based on a grammatical morpheme dictionary, a vocabulary morpheme dictionary, a terminology dictionary, a user-defined dictionary, and a pre-analysis dictionary. As an example, the text feature extraction device 120 may perform morpheme analysis using konlpy. In addition, the text feature extraction device 120 may extract a key word from the text data whose morpheme is analyzed. Also, the text feature extraction device 120 may vectorize the extracted key words by applying it to the text analysis algorithm. In addition, the text feature extraction apparatus 120 may extract a feature by applying a vectorized key word to principal component analysis. Also, the text feature extraction device 120 may provide the extracted feature to the key word visualization device 150 of the feature.

In the above description, steps S101 to S105 may be further divided into additional steps or combined into fewer steps, depending on the implementation herein. In addition, some steps may be omitted if necessary, and the order between the steps may be changed.

The method for visualizing a key word of a feature according to an embodiment of the present application may be implemented in a form of program instructions that can be executed through various computer means and may be recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floptical disks. Includes hardware devices specifically configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language code that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the present invention, and vice versa.

In addition, the key word visualization method of the above-described feature may be implemented in the form of a computer program or application executed by a computer stored in a recording medium.

The above description of the present application is for illustrative purposes, and those skilled in the art to which the present application pertains will understand that it is possible to easily modify to other specific forms without changing the technical spirit or essential features of the present application. Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as distributed may be implemented in a combined form.

The scope of the present application is indicated by the claims below, rather than the detailed description, and it should be interpreted that all changes or modifications derived from the meaning and scope of the claims and equivalent concepts thereof are included in the scope of the present application.

110: text data
120: text feature extraction device
130: word extraction processor
140: feature extraction execution unit
150: feature key word visualization device
160: visualization execution unit

Claims (6)

In the feature word visualization device,
A feature information input unit for inputting importance values for target variables of a plurality of features extracted from text data using principal component analysis;
A reference importance input unit for setting a reference value associated with the importance value based on user input information;
An importance extracting unit for each word that extracts an influence on each feature for each word vector for the plurality of features;
A importance calculation unit for each word that converts the importance as a percentage of the value extracted by the importance extraction unit for each word; And
Visualization execution unit to visualize the importance of each word and information about the extracted features in the form of a Venn diagram graph,
Including,
The feature information input unit,
When the target variable is set, a key word visualization device of the feature is to calculate the degree of influence between the extracted feature and the target variable, and input an importance value for the target variable. delete According to claim 1,
The visualization execution unit,
The key word visualization device of the feature is to determine the output location, output size, and output color according to the importance of each word and the frequency of the extracted features to visualize in the cloud form. According to claim 1,
A word extraction processor that extracts key words from the received text data using a plurality of dictionaries; And
A feature extraction execution unit that vectorizes the extracted key words by applying them to a text analysis algorithm and then extracts the features of the text data using principal component analysis.
Feature word visualization device further comprising a. According to claim 4,
The feature information input unit,
When there is a variable name of the text data and the target variable, a key word visualization device of the feature is to input a variable importance value for the target variable based on the feature extracted by the feature extraction execution unit. A method for visualizing a key word of a feature in which each step is performed by a key word visualization device of a computer-implemented feature,
Receiving an importance value for a target variable of a plurality of features extracted by using principal component analysis from text data;
Setting a reference value associated with the importance value based on user input information;
Extracting a degree of influence on each feature for each word vector for the plurality of features;
Converting the importance as a percentage of a value obtained by extracting the influence on each feature for each word vector for the plurality of features; And
Visualizing the importance of each word and information about the extracted features in the form of a Venn diagram graph,
Including,
The step of receiving the importance value,
When the target variable is set, calculating the degree of influence between the extracted feature and the target variable, and inputting an importance value for the target variable, the key word visualization method of the feature.

Images