KR102009029B1 - A contents filtering system for comparative analysis of feature information - Google Patents

Abstract

A content filtering system for the comparative analysis of feature information according to the present invention includes: a content collection module collecting and storing content consisting of any one of a video, an image, and a sound; a feature information extraction module extracting hash value feature information of the content and storing the same as feature information; a filtering module filtering illegal content by comparing and analyzing the feature information in the content; and a result providing module transmitting the filtering result of the filtering module to a content provider. According to the present invention, an illegal copy of the content shared through an online storage service and the like can be filtered by extracting and comparing the feature information of the content, and thus the illegal copy of the content can be effectively blocked and managed.

Description

Content filtering system through comparative analysis of feature information {A CONTENTS FILTERING SYSTEM FOR COMPARATIVE ANALYSIS OF FEATURE INFORMATION}

The present invention relates to a content filtering system by comparing and analyzing feature information. More specifically, the present invention relates to an illegal copy of a content shared through a web hard, etc., by extracting and comparing feature information of the content and filtering the illegal copy of the content. The present invention relates to a content filtering system through feature information comparison and analysis that can be blocked and affiliated with each other.

First, Web hard refers to an Internet content management service that allows a user to store, view and edit a document or content by securing a certain amount of storage space, and to share content with a large number of people. In such a web hard, a variety of contents such as movies, music, dramas and entertainment are widely distributed, and users can easily exchange these contents through the web hard. At this time, the distribution of such content is common to give and receive without the consent of the copyright holder, and in general, since the unspecified majority is shared, the distribution of such content not only causes serious damage to the creator and the related person of the content but also violates the copyright law. do. Therefore, various methods for preventing unauthorized distribution of counterfeit contents have been researched and applied.

As a prior art, Korean Patent No. 10-1035867 discloses a method and system for filtering content in a user terminal using a watermark.

The prior art is a content filtering method in a user terminal using a watermark, the method comprising: extracting a watermark of the content through a program equipped with a watermark extraction module; Transmitting the extracted watermark to a watermark management server; Receiving watermark information and control information based on the extracted watermark from the watermark management server; And filtering the content based on the control information, wherein the watermark information includes at least one of user information, copyright information, copy information of the content, distribution information, information on illegality, and information on billing. Including the above, the control information includes a content filtering method including information on whether to filter the content.

The prior art is to solve the problem of illegal spread of the content when the service to the watermark-embedded content in the content service environment using the user terminal, re-transmitted directly from the user terminal to another user terminal or downloaded from another user terminal However, the real-time copyright management system can be provided by extracting the watermark of the corresponding content in real time and providing a function of filtering and filtering the illegal content, but the watermark may be displayed on the screen unnecessarily. The disadvantage is that there is a lack of precision filtering.

Therefore, there is a need to develop a new and advanced content filtering system capable of extracting characteristic information (DNA) of content and filtering illegally copied contents such as black and white, up, down, left, right, and size manipulation.

SUMMARY OF THE INVENTION The present invention has been made to overcome the problems of the above technology, and the illegal copy of the content shared through the web hard and the like to extract and compare the feature information of the content to filter and effectively manage the illegal copy of the content The main objective is to provide a system that can

Another object of the present invention is to enable illegal content extraction using hash values.

It is a further object of the present invention to enable illegal content extraction through keyword comparison between the protected content and the analyzed content.

It is a further object of the present invention to extract a target keyword from text related to the analysis target content as well as the keyword included in the analysis target content when extracting illegal content using the keyword, and to use it for similarity analysis.

In order to achieve the above object, the content filtering system through the comparative analysis of the feature information according to the present invention, a content collection module for collecting and storing the content consisting of any one video, image, sound; A feature information extraction module for extracting the hash value feature information of the content and storing it as feature information; A filtering module for filtering illegal content by comparing and analyzing the feature information among the contents; And a result providing module for transmitting the filtering result of the filtering module to a content provider.

The content collection module may include a protection target collection unit configured to receive and store protected content from the content provider, and an analysis target collection unit configured to collect and store analysis target content uploaded to a plurality of content service sites. The feature information extraction module extracts and stores the hash value feature information of the protected content, and the filtering module compares and analyzes the protected content and the analyzed content based on the feature information.

Furthermore, the protected object collecting unit includes a feature keyword collecting part that receives and stores a feature keyword of the protected object content from the content provider, and the analyzing object collecting unit collects and stores related keywords related to the analyzed object content. And a related keyword collection part, wherein the system further comprises: a target keyword extracting module configured to extract and store a target keyword including a characteristic word from a plurality of the related keywords, wherein the filtering module comprises: the protection module; And a keyword comparison analysis unit which compares and analyzes feature keywords of target content and target keywords of the analysis target content.

According to the content filtering system through the comparative analysis of the feature information according to the present invention,

1) By extracting and comparing the feature information of the content and filtering, it is possible to efficiently block and affiliate management of illegal copies of the content,

2) It is possible to extract illegal content using hash value.

3) keyword comparisons not only enable illegal content extraction,

4) It is effective to provide more accurate analysis tool by extracting target keyword from text based on more precise and objective algorithm.

1 is a block diagram showing a detailed configuration of a system of the present invention.
2 is a process diagram showing the sequence of the filtering system of the present invention.
3 is a conceptual diagram illustrating a method of extracting feature information of content of the present invention.
4 is a conceptual diagram illustrating an embodiment of the feature information (DNA) of the content of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. The accompanying drawings are not drawn to scale, and like reference numerals in each of the drawings refer to like elements.

First, each component constituting the content filtering system 10 through the comparative analysis of the feature information of the present invention will be described.

The content filtering system 10 through the comparative analysis of the feature information of the present invention may basically include a main server.

The main server may also be referred to as a collection of server PCs for the implementation of the system 10 of the present invention. As will be described later, the system includes various databases, so that the system server for implementing the system of the present invention may be used as well. The main server of the present invention. The main server implements a content monitoring system to identify affiliated and non-affiliated contents among the contents uploaded to the service site, and further performs filtering on illegal contents, that is, contents that are illegally sold without the permission of the content provider. do.

Here, the affiliated content is content that is normally sold through a contract with a content provider, that is, a person having a copyright or a distribution right for the content, that is, content that is shared through a formal permission.

Non-partnered content is a wide range of illegal content, which is content that is sold without a contract with a content provider. In other words, it can be said to be content that is shared and sold without formal permission for sale and sharing.

Therefore, the main server of the present invention can perform the role of extracting non-affiliated content by discriminating affiliated content and non-affiliated content among contents uploaded to a service site through the implementation of the system 10, and furthermore, among non-participated contents. It is to allow illegal content extraction in a detailed sense. At this time, the extraction target means the content that the protection request has been received from the content provider, that is, the content that has been protected, and this content is prohibited from being legally shared without the affiliation. As such, illegal content in a narrow sense means content that is highly similar to protected content among non-affiliated contents. Alternatively, the system of the present invention can concentrate on the role of dividing the content into two categories, normal content and illegal content without distinguishing affiliated / non-partnered content or non-partnered / illegal content. It will be explained in detail.

That is, according to the system of the present invention, the distribution of illegal content and normal sale of affiliated content are monitored in real time on a service site, i.e., the web hard, and the request for blocking and normal content (or protected content) and illegal content, etc. Distribution statistics and the like can be provided digitally.

The content filtering system 10 through the comparative analysis of the feature information of the present invention will be described in more detail as follows.

1 is a block diagram showing a detailed configuration of the system of the present invention, Figure 2 is a process diagram showing the sequence of the filtering system of the present invention.

Referring to Figures 1 and 2, the content filtering system 10 through the feature information comparison analysis of the present invention is basically a content collection module 100, feature information extraction module 200, filtering module 300, It includes a result providing module 400.

3 is a conceptual diagram illustrating a method of extracting feature information of content of the present invention, and FIG. 4 is a conceptual diagram showing an embodiment of feature information (DNA) of content of the present invention.

The content collection module 100 collects and stores content consisting of any one of a video, an image, and a sound, and may play a role of collecting and storing content by a main server of the system of the present invention.

The content collection module may be composed of a protection target collection unit 110, the analysis target collection unit 120.

The protected object collecting unit 110 receives and stores protected content from a content provider. The protected content means a content that a protection request is received from a content provider, that is, a content provided through a normal distribution process.

The analysis target collection unit 120 collects and stores the analysis target content uploaded to a plurality of content service sites, wherein the analysis target content refers to the content requested to determine whether it is illegal content in the system of the present invention. In other words, this refers to content that includes both protected and illegal content that has not yet been analyzed.

 That is, the analysis target collecting unit 120 filters the analysis target content through a configuration and a function to be described later, that is, a series of functions for comparing and analyzing the protected content and the analysis target content based on specific information to determine whether or not it is illegal copy content. It can be determined.

The feature information extraction module 200 extracts the hash value feature information of the content and stores it as feature information. The feature information extraction module 200 may extract and store feature information of the protected content stored in the database by the above-described content collection module. Here, the hash value characteristic information means a hash value of the corresponding content, and the hash value is a numeric value such as a password abbreviated as a file characteristic in order to prove the identity of the copied digital evidence. It can be called the characteristic of the content because it is referred to as the 'fingerprint of evidence'. In this case, the hash value feature information is described as an example of a video. Each of the feature points (DNA) of the video frame (maximum of about 300 to 800 frames can be extracted based on an image of an hour) It can be hash value feature information.

The filtering module 300 compares and analyzes the feature information between contents to filter illegal content, wherein the feature information of the protected content (original DNA file) extracted from the feature information extraction module and the illegal content (manipulated DNA file) Compare feature information. Here, the feature information may be a DNA of the content, that is, a hash value (hash value feature information) as described above.

To describe the feature information of the image, for example, when the illegally copied content is copied in black and white after being protected in black and white, it is judged to be illegally copied by comparing the feature information of the extracted image. Can be filtered. That is, as shown in FIG. 4, when the feature information is extracted and stored, the presence or absence of duplication may be determined and filtered in black and white processing, up, down, left and right inversion processing size manipulation processing, and the like.

The result providing module 400 transmits the filtering result of the filtering module to the content provider. The result providing module 400 stores the filtering result of the content determined to be illegally copied from the filtering module, that is, the result database, and provides the result to the user. In other words, the result of the filtering module may be illegal copy content or normal distribution content. In this case, the result of the derived filtering may be provided in a cloud type by accessing a server type or the Internet according to a user's request.

In this case, the protection target collection unit 110 of the above-described content collection module 100 may include a feature keyword collection part 111, and the analysis target collection unit 120 may include an associated keyword collection part 121. Can be. Accordingly, the system may further include a target keyword extraction module 500, and the filtering module 300 may include a keyword comparison analyzer 310.

The feature keyword collection part 111 receives and stores a feature keyword of the protected content from the content provider, and sets and stores a keyword that can characterize the protected content in advance. For example, if the feature keyword for the movie "Frozen, 2013" is stored, the director's name "Chris Buck", the producer "Disney", the character name " Elsa "," Anna "," Olaf ", etc. may be stored as feature keywords.

In addition, the feature keyword collection part 111 may include a search keyword collection unit 112, a frequency determining unit 113, and a feature keyword parallel setting unit 114.

The search keyword collecting unit 112 collects search keywords derived as a search result when searching for a feature keyword in a plurality of portal sites. The search keyword collecting unit 112 collects result data or related search terms derived when the corresponding content is searched on the portal site. will be.

The frequency determining unit 113 determines the frequency of the search keyword, and may determine that a high frequency among the search keywords derived through the search keyword collection unit described above is related to the corresponding content.

The feature keyword parallel setting means 114 sets at least one of the search keywords as the feature keyword based on the frequency. In other words, it is determined that the frequency of the search keywords is high from the above-described frequency determining means to be the feature of the corresponding content, and is added together as the feature keyword.

The related keyword collection part 121 collects and stores related keywords related to the analysis target content. In this case, the method of collecting related keywords of the analysis target content is applied to a method of collecting related search terms that can be provided when the analysis target content is searched on the portal site, analyzing the words of the content introduction or description field, and inputting them directly. Can be.

Accordingly, the target keyword extraction module 500, which may be additionally included in the system (! 0) of the present invention, extracts and stores a target keyword including a feature word from a plurality of related keywords. Among the related keywords collected according to the related keyword collection part 121 described above, the keyword determined as not being a feature of the content to be analyzed is filtered, and only the keywords that can be featured are extracted and stored.

In addition, the keyword comparison analysis unit 310 of the filtering module 300 compares and analyzes the feature keyword of the protected content and the target keyword of the analysis target content, and compares and compares the degree of matching between the feature keyword and the target keyword. Is to compare the content. In general, when illegal content is uploaded to a site, the title of a copy object is not used as it is, but it is expressed indirectly. Through comparison analysis of these keywords, it is possible to confirm whether the same content is used. As an example, it could be uploaded under the title "Times Best Film of the Year, Chris Buck's Best Disney Film" as a replica of the movie "Frozen, 2013." In this case, it is possible to confirm that the content is identical to the work of "Frozen, 2013" through the target keywords of "Cris Buck" and "Disney" among the keywords of the duplicate content, that is, the analysis target content.

In addition, the basic function of the target keyword extraction module 500 may be referred to as keyword analysis for a plurality of related keywords, and further, target keyword extraction from keywords. In general, in the case of keyword analysis, AHP hierarchical analysis has been mainly used, but more preferably, the target keyword can be generated by performing a more detailed analysis through potential hierarchical analysis. It will be more effective to generate the target keyword feature information.

To this end, the target keyword extraction module 500 includes a word detection unit 510, an analysis keyword generation unit 520, a stair classification unit 530, a keyword group generation unit 540, and a target keyword designation unit 550. can do. Each configuration will be described in more detail as follows.

The word catching unit 510 is to identify the words included in the related keyword. In this case, the criteria for distinguishing words is based on spacing. The number of words may vary depending on the spacing, but when the associated keyword does not perform spacing, words are divided according to roots and proper nouns. In the case of 'Amazing Spider-Man', it is identified as two words, and when the relevant text of 'Amazing Spider-Man' is entered, it is divided into two words, 'Amazing' (root) and 'Spider-Man' (proper noun). will be. In other words, if no spacing is performed, it is basically divided according to roots. If two or more roots constitute one proper noun, the proper noun itself is determined as one word.

The analysis keyword generation unit 520 generates an analysis keyword by filtering the plurality of words, and generates an analysis keyword by filtering the plurality of words extracted through the word detection unit 510. Punctuation marks and symbols, such as periods, quotes, and question marks, do not significantly affect text analysis. Also, the filter process is preferably based on nouns on the extracted words. In this case, pronouns other than personal pronouns are already nouns, even if they are kept in this form, and thus no additional conversion is required. In this case, words that cannot be noun, for example, prepositions, etc., cannot be noun, and thus are deleted.

The stair classification unit 530 classifies the plurality of analysis keywords into a plurality of hierarchies by performing latent hierarchical analysis (LCA) on the plurality of analysis keywords generated by the analysis keyword generation unit 520. Do this. At this time, latent hierarchical analysis is a part of constructing the structural equation model, and it is a part showing causality and correlation among latent variables. Since the latent variable is an analysis keyword, it is possible to classify the keyword according to the causality and correlation between the analysis keywords.

Therefore, since the analysis keywords stratified through the latent hierarchical analysis are estimated groups based on similarity, the target keyword characteristic information generated by the latent hierarchical analysis method can also be sorted by type.

In this case, hierarchical analysis of AHP method, which has been widely used in text analysis, is based on clustering. Cluster analysis is a simple method of attempting classification based on data values and based on coefficients estimated by specific statistical methods such as mixed models. Categorizing them as (eg, categorizing them based on estimated rates of change in the potential growth model) is not possible. In the case of latent hierarchical analysis, it can be combined with various analyzes by including various statistical indices, longitudinal analysis, influence variables and outcome variables to determine the number of groups, which is very powerful and flexible. have.

The stair classification unit 530 is said to perform a function of classifying a plurality of keywords into a plurality of stairs. For this purpose, the stair classification unit 530 is basically a factor extraction part 531 and a model application part 532. ) Is included by default.

The factor extraction part 531 performs a function of extracting a category factor from a plurality of the analysis keywords. In this case, a category factor is a keyword that can show characteristics of the corresponding content to be provided. For example, a keyword related to the genre of the content to be provided, a keyword related to a title or a file name, etc. may be extracted as a category factor.

The model application part 532 performs hierarchical classification based on the mutual similarity of the category factor extracted through the factor extraction part 531 through the latent variable model. In the latent variable model, the determination of the number of hierarchies and the determination of parameters are essential, and thus, an additional configuration for determining the number of hierarchies and parameters is required. Therefore, in order to smoothly apply the model application part 532, it is necessary to determine the parameters and the layer coefficient through the detailed configuration of the stair classification unit 530. For this purpose, the stair classification unit 530 has a parameter range determination part 533. ), The step determination part 534 may be further included.

The parameter range determination part 533 performs a function of determining a parameter range according to the number of keywords extracted from the analysis keyword. The value of any one of these parameter ranges is the number of stairs to which the actual analysis keyword is classified. At this time, the reason for determining only the parameter range and not determining the layer coefficient in advance is to enable more exploratory and technical analysis of the analysis keyword, and to inductively determine the corresponding data based on the analysis keyword, that is, the data. In addition, because the model is sufficiently verified when determining the layer coefficients, accurate analysis is possible without setting a separate prior number of layers. At this time, the parameter range is not limited, but the parameter range is determined as a range including the number of extracted keywords. That is, the size of the parameter range is not limited, but, for example, if 100 keywords are extracted, the parameter range must be a range including 100. That is, the parameter range should be determined as a range including 100, such as 1 or more and 100 or less, 1 or more and 1000 or less, 50 or more and 150 or less.

The step determination part 530 calculates the akaike information index, the Bayesian information index, and the modified Bayesian information index for each of the integers included in the parameter range, and compares the values to determine the value of any one of the parameter ranges. Perform the function specified by. As described above, the range of the number of layers is not calculated in advance, and then the most suitable value is designated as the number of layers by applying the actual model. More preferably, each of the integers included in the parameter range is designated as a preliminary parameter, and for each of the preliminary parameters, an akaike information index, a Bayesian information index, a corrected Bayesian information index are calculated, and the calculated values are analyzed and substituted. An integer representing the value closest to the maximum likelihood estimate compared to the value of one integer is designated as the number of layers.

To this end, it is basically necessary to calculate the maximum likelihood estimate, which can be calculated through the following equations (1) to (2).

Equation 1,

는 분석 키워드

에 대해 설정된 지수,

는 지수

및 예비모수

에 대한 우도함수,

는 상기 모수범위에 포함된 수 중 어느 하나로써 예비모수,

은 분석 키워드의 총 개수를 의미한다.)(here,

Analyze keywords

The exponent set for,

Is the exponent

And preliminary parameters

Likelihood function for,

Is any one of the numbers included in the parameter range as a preliminary parameter,

Means the total number of analysis keywords.)

Equation 2,

는 최대우도추정치,

는 분석 키워드

에 대해 설정된 지수,

은 분석 키워드의 총 개수를 의미한다.)(here,

Is the maximum likelihood estimate,

Analyze keywords

The exponent set for,

Means the total number of analysis keywords.)

에 대해 설정된 지수라 함은 관련텍스트에 분석 키워드

가 나타난 빈도수를 의미하며, 분석 키워드의 총 개수는 연관 키워드에 포함된 단어에서 추출된 분석 키워드의 전체 개수를 의미한다.If you explain in more detail,

The exponent set for

Denotes a frequency of appearing, and the total number of analysis keywords refers to the total number of analysis keywords extracted from words included in the related keyword.

At this point, it is difficult to calculate the likelihood function in general calculation. Therefore, when calculating the likelihood function, it is better to use a statistical program. For this purpose, a statistical program such as MPlus may be preferably used. In the case of likelihood function, the maximum value is a general goal. In this case, the smaller the value of the preliminary parameter and the larger the value of the likelihood function may be a preferable maximum value to be extracted.

Further, the step determination part 534 designates each of the integers included in the parameter range as a preliminary parameter, calculates the Akaike Information Index, the Bayesian Information Index, and the Corrected Beige Information Index, and analyzes the calculated values and substitutes them. The integer that represents the value closest to the maximum likelihood estimation value is designated as the number of hierarchies, and each calculation formula is as follows.

Calculation of Akaike Information Index

Equation 3,

Bayesian Information Index

Equation 4,

Corrected Bayesian Information Index

Equation 5,

는 아카이케정보지수,

는 베이지안정보지수,

는 수정베이지안정보지수,

는 분석 키워드

에 대해 설정된 지수,

는 지수

및 예비모수

에 대한 우도함수,

는 상기 모수범위에 포함된 수 중 어느 하나로써 예비모수,

은 분석 키워드의 총 개수를 의미한다.)(here,

Is the Akaike Information Index,

Is the Bayesian Information Index,

Is the Corrected Bayesian Information Index

Analyze keywords

The exponent set for,

Is the exponent

And preliminary parameters

Likelihood function for,

Is any one of the numbers included in the parameter range as a preliminary parameter,

Means the total number of analysis keywords.)

The reason for calculating the three information indices and comparing the values is the first to determine which model is most suitable. Furthermore, each information index imposes different penalties according to the number of parameters and the number of samples. Therefore, by calculating all three information indexes that impose different penalties, and comparing them, the model of the information index that shows the likelihood value that shows the closest value to the maximum likelihood estimation value is compared with the value of the substituted integer. It is determined by the information index model, and the value of the integer is designated as the number of layers.

In addition, obtaining each information index can be difficult with general calculations. Therefore, it is good to use a statistical program when calculating each information index, for this purpose, a statistical program such as MPlus may be preferably used.

Therefore, when the number of stairs is designated through the above-described configuration, the stair classification through the model application part 532 is possible. In this case, the stair classification is possible through Equation 6 below.

Equation 6,

는 종속변수의 벡터값에 따른 잠재변수모형의 값,

는 분석 키워드

에 대한 종속변수의 벡터,

는 추출된 카테고리팩터의 계급값,

는 각각의 층계,

는 층계수판단파트에 의해 판단된 층계수를 의미한다.)(here,

Is the value of the latent variable model according to the vector value of the dependent variable,

Analyze keywords

Vector of dependent variables for,

Is the rank value of the extracted category factor,

Each staircase,

Means the floor coefficient judged by the floor determination part.)

In this case, the rank value means a frequency in which the analysis keyword designated as the category factor appears in the relevant text, and in the case of the layer coefficient, it is calculated through Equations 1 to 5 described above. In addition, even when applying the latent variable model, if you use a statistical program such as MPlus will be available to those skilled in the art.

Therefore, the comparison process through various models and the determination of the layer coefficients through them can be used for stratification with excellent accuracy, and an unbiased value can be obtained. Furthermore, as mentioned earlier, the stratified keywords are estimated groups based on similarity, and thus, analysis keywords typed according to the latent hierarchy analysis method can be obtained.

Furthermore, the keyword group generator 540 performs a function of grouping the classified analysis keywords to generate a keyword group. This is a function of generating keyword groups for each type by grouping keywords classified according to the hierarchy into one keyword group.

The target keyword designation unit 550 designates at least one of the analysis keywords included in each of the keyword groups as a target keyword, and all of the analysis keywords included in the keyword group may be designated as the target keyword, or each Only one of the keyword groups of may be designated as the target keyword. In the monitoring and similarity analysis, the number of target keyword designations may be changed by the system administrator according to whether the weight is more weighted or the accuracy is weighted.

As described above, although the content filtering system through the comparative analysis of the feature information according to the present invention has been represented in the above description and the drawings, this is merely an example, and thus the spirit of the present invention is not limited to the above description and the drawings. Various changes and modifications are possible without departing from the spirit of the invention.

10: system 100: content collection module
110: protection target collector 111: feature keyword collection part
112: search keyword collection means 113: frequency determination means
114: feature keyword parallel setting means 120: analysis target collection unit
121: related keyword collection part 200: feature information extraction module
300: filtering module 310: keyword comparison analysis unit
400: result providing module 500: target keyword extraction module
510: word detection unit 520: analysis keyword generation unit
530: stair classification unit 531: factor extraction part
532: Model application part 533: Parameter range determination part
534: Step determination part 540: Keyword group generation unit
550: target keyword specifying unit

Claims (10)

Content filtering system through comparative analysis of feature information,
A protected collection unit for receiving and storing protected content and feature keywords of the protected content from a content provider, and collecting and storing analysis content uploaded to a plurality of content service sites and related keywords associated with the analysis content; A content collection module having an analysis target collection unit;
A feature information extraction module for extracting hash value feature information of the content to be protected and storing it as feature information;
A target keyword extraction module for extracting and storing a target keyword including a word characterizing from a plurality of related keywords;
A filtering module which compares and analyzes the protected content and the analysis content based on the feature information, and compares and analyzes feature keywords of the protected content and target keywords of the analysis content;
And a result providing module for transmitting the filtering result of the filtering module to a content provider.
The target keyword extraction module,
A word grasping unit which grasps the words included in the related keywords, an analysis keyword generating unit which generates an analysis keyword by filtering the plurality of words, and a plurality of the analysis keywords through latent hierarchical analysis (LCA); A target classifier which designates at least one of a stair classification unit classified into a stair, a keyword group generation unit that groups the classified analysis keywords to generate a keyword group, and the analysis keywords included in each of the keyword groups as target keywords. Content filtering system, characterized in that consisting of a keyword designation unit. The method of claim 1,
The feature keyword collection part,
Search keyword collection means for collecting a search keyword derived as a search result when the feature keyword is searched in a plurality of portal sites;
Frequency determining means for determining the frequency of the search keyword;
And a feature keyword parallel setting means for setting at least one of the search keywords as a feature keyword on the basis of the frequency. The method of claim 1,
The stair classification unit,
A factor extraction part for extracting a category factor from a plurality of analysis keywords;
And a model application part for classifying the plurality of category factors into a plurality of layers through a latent variable model based on mutual similarity. The method of claim 3, wherein
The stair classification unit,
A parameter range determination part for determining a parameter range based on the number of extracted analysis keywords, and
Each of the integers included in the parameter range is designated as a preliminary parameter, and for each of the preliminary parameters, an akaike information index, a Bayesian information index, and a modified Bayesian information index are calculated, and the calculated values are analyzed and compared to the values of the substituted integers. And a step coefficient determining part that designates an integer representing a value closest to the maximum likelihood estimate as a layer coefficient. The method of claim 4, wherein
The maximum likelihood estimate is
The content filtering system, characterized in that calculated through the following equation (1) to (2).
Equation 1,

(here,

Analyze keywords

The exponent set for,

Is the exponent

And preliminary parameters

Likelihood function for,

Is any one of the numbers included in the parameter range as a preliminary parameter,

Means the total number of analysis keywords.)
Equation 2,

(here,

Is the maximum likelihood estimate,

Analyze keywords

The exponent set for,

Means the total number of analysis keywords.) delete delete delete delete delete

Images