JP5364010B2 - Sentence search program, server and method using non-search keyword dictionary for search keyword dictionary - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a text classification program, server and method for preventing text information that is not illegal or harmful from being classified into an illegal/harmful category when keywords registered in advance are used for the classification. <P>SOLUTION: Multiple legitimate learned text information which do not belong to a specific category, and multiple illegitimate learned text information which belong to the specific category are stored in a learned text storage means. Learned text information including a search keyword is searched. Modification keywords for the search keyword are extracted. A legitimacy rate of each modification keyword is calculated as the number of pieces of legitimate learned text information to the number of all the pieces of learned text information. The modification keywords whose legitimacy rates are not lower than a predetermined threshold are registered as non-search keywords to generate a non-search keyword dictionary. The text information including the non-search keyword as the modification keyword for the search keyword is prevented from being searched. <P>COPYRIGHT: (C)2011,JPO&amp;INPIT

Description

  The present invention relates to a text search program, server, and method using a search keyword dictionary.

  With the spread of the Internet, various texts are described on blogs, bulletin boards, or Web sites that publish reviews. A “blog” is a site that is generally run by an individual and can be updated in a diary to express their opinions on current news and specialized topics. A “bulletin board” is a site for sequentially exchanging discussions with other people on various themes. Furthermore, “review comments” refers to a site where comments about things such as people's rumors can be described. With the spread of these sites, general users can freely send information on the Internet.

  Along with this, it is a social problem that these websites contain text that slanders and slanders individuals and organizations, or illegal and harmful text information such as crime notices, crime assistants, or adult sites. It has become. Therefore, it is necessary to detect sentences and individuals that describe these illegal and harmful sentences while giving consideration to freedom of expression.

  Conventionally, there is a filtering technique for denying access from a terminal to a site describing illegal / harmful text (see, for example, Patent Document 1). According to this technique, a keyword or URL (Uniform Resource Locator) is registered in advance, and access from a terminal to a Web site including the keyword or URL is denied. However, it is necessary to constantly update keywords and URLs registered in advance, which requires time and cost.

  On the other hand, there is a technique for automatically extracting and filtering keywords and URLs (see, for example, Patent Document 2). According to this technology, keywords can be automatically extracted from illegal / harmful text information such as adult sites and crime notice sites. By registering these keywords in advance, illegality / harmfulness is determined depending on whether or not the text information to be analyzed contains those keywords.

JP 2007-128119 A JP 2009-37420 A

Nagaoka University of Technology, Natural Language Processing Laboratory, “CaboCha”, [online], [Search on December 28, 2009], Internet <URL: http://nlp.nagaokaut.ac.jp/CaboCha> Department of Intelligent Informatics, Graduate School of Informatics, Kyoto University, Kurohashi Laboratory, "Japanese Parsing System KNP", [online], [Search on December 28, 2009], Internet <URL: http: // nlp .kuee.kyoto-u.ac.jp / nl-resource / knp.html>

  However, when a general keyword is registered as belonging to an illegal / harmful category, there is a problem that a sentence that is not illegal / harmful is picked up as illegal / harmful text. For example, when the keyword search technology is simply used, “blasting” is registered as an illegal / harmful keyword in order to detect illegal / harmful keywords such as “blast the school”. In this case, even expressions such as “blast the mine” and “no blast” are detected as illegal and harmful text information. As a result, the classification accuracy into the illegal / harmful category is lowered, and the illegal / hazardous is determined by visual inspection.

  Therefore, the present invention, for example, when determining whether or not a pre-registered keyword belongs to an illegal / harmful category, as long as it is possible to classify text information that is not illegal / harmful into an illegal / harmful category. An object of the present invention is to provide a sentence classification program, a server, and a method capable of reducing and expanding the range of applicable sentences as much as possible.

According to the present invention, there is provided a text search program having search keyword dictionary means for storing search keywords, and causing a computer to function to search text information including the search keywords,
A learning sentence storage means for storing a plurality of legitimate learning sentence information not belonging to a specific category and a plurality of illegal learning sentence information belonging to a specific category;
Learning sentence keyword search means for searching learning sentence information including a search keyword using learning sentence storage means;
A learning dependency analysis means for extracting dependency keywords for the search keyword for learning sentence information including the search keyword;
Concept dictionary storage means in which word nodes are arranged in a tree shape;
Using a concept dictionary storage means, a dependency-related keyword extracting means for extracting a plurality of dependency-related keywords that are subordinate concepts as viewed from a node corresponding to the dependency keyword;
The number of valid learning text information and the number of illegal learning text information are counted for each of the dependency keywords and the dependency related keywords, and the number of valid learning text information with respect to the number of all learning text information is valid. A legitimate proportion calculating means for calculating the proportion;
Sentences that include a non-search keyword as a dependency keyword for a search keyword by causing the computer to function as a non-search keyword dictionary means for registering dependency keywords and dependency-related keywords with a legitimate ratio equal to or greater than a predetermined threshold as non-search keywords The computer functions so that information is not retrieved.

According to another embodiment of the text search program of the present invention,
The dependency-related keyword extracting means extracts a plurality of dependency-related keywords that are lower concepts as viewed from a higher concept node at least one level higher than a node corresponding to the dependency keyword, using the concept dictionary storage means. It is also preferable to make the computer function like this.

According to another embodiment of the text search program of the present invention,
The dependency-related keyword extraction means uses the concept dictionary storage means to determine the number of higher concept nodes to be viewed from the hierarchy n based on the height of the hierarchy n of the node corresponding to the dependency keyword. It is also preferable to make the computer function like this.

According to another embodiment of the text search program of the present invention,
The dependency-related keyword extracting means causes the computer to function based on the part-of-speech of the dependency keyword so as to determine the number of higher-level concept nodes to be viewed from the hierarchy n using the concept dictionary storage means. Is also preferable.

According to another embodiment of the text search program of the present invention,
The dependency related keyword extracting means preferably causes the computer to function based on the part of speech of the dependency related keyword so as to determine whether or not to extract the dependency related keyword.

According to another embodiment of the text search program of the present invention,
It further includes search related keyword extraction means for extracting a plurality of search related keywords that are lower concepts than the node corresponding to the search keyword using the concept dictionary storage means,
The learning sentence keyword searching means further searches the learning sentence information including the search related keyword using the learning sentence storing means,
The learning dependency analysis means preferably causes the computer to function so as to further extract the dependency keyword for the search related keyword with respect to the learning sentence information including the search related keyword.

According to another embodiment of the text search program of the present invention,
A target sentence input means for inputting target sentence information to be analyzed;
Target text keyword search means for determining whether the target text information includes a search keyword stored in the search keyword dictionary means;
A target dependency analysis means for extracting a dependency keyword for the search keyword for the target sentence information including the search keyword;
Non-search keyword determining means for determining whether or not the extracted dependency keyword matches the non-search dependency keyword registered in the non-search keyword dictionary means;
The target sentence information determined to be false by the target sentence keyword search means and the target sentence information determined to be true by the target sentence keyword search means and determined to be true by the non-search keyword determination means do not belong to a specific category. It is also preferable to cause the computer to function as text information classification means for classifying text information.

According to another embodiment of the text search program of the present invention,
The non-search keyword dictionary means registers the dependency ratio calculated by the validity ratio calculation means in association with each registered dependency keyword,
The non-search keyword determining means determines that the extracted dependency keyword matches the dependency keyword registered in the non-search keyword dictionary means, and the valid ratio in the dependency keyword is equal to or greater than a predetermined threshold. It is also preferable to make the computer function so as to determine that the dependency keyword is true.

According to another embodiment of the text search program of the present invention,
Search keywords are keywords set for the purpose of categorization,
It is also preferable to cause the computer to function so that the specific category in the learning sentence storage means is a group of sentence information based on the category classification.

According to another embodiment of the text search program of the present invention,
Search keywords are illegal and harmful keywords,
It is also preferable to make the computer function so that the specific category in the learning sentence storage means is a group of illegal and harmful sentence information.

According to another embodiment of the text search program of the present invention,
Search keywords are keywords specified by the user,
It is also preferable to make the computer function so that the specific category in the learning sentence storage means is a group of sentence information designated by the user.

According to another embodiment of the text search program of the present invention,
The target text information is also preferably text information described by an unspecified number of users in a blog (Weblog), bulletin board, and / or word-of-mouth comment published via the network.

According to the present invention, there is provided a text analysis server having search keyword dictionary means for storing search keywords and searching text information including the search keyword from text information acquired from another public server via a network. ,
A learning sentence storage means for storing a plurality of legitimate learning sentence information not belonging to a specific category and a plurality of illegal learning sentence information belonging to a specific category;
Learning sentence keyword search means for searching learning sentence information including a search keyword using learning sentence storage means;
A learning dependency analysis means for extracting dependency keywords for the search keyword for learning sentence information including the search keyword;
Concept dictionary storage means in which word nodes are arranged in a tree shape;
Using a concept dictionary storage means, a dependency-related keyword extracting means for extracting a plurality of dependency-related keywords that are subordinate concepts as viewed from a node corresponding to the dependency keyword;
The number of valid learning text information and the number of illegal learning text information are counted for each of the dependency keywords and the dependency related keywords, and the number of valid learning text information with respect to the number of all learning text information is valid. A legitimate proportion calculating means for calculating the proportion;
Sentence information having a non-search keyword dictionary means for registering a dependency keyword and a dependency-related keyword as a non-search keyword with a legitimate ratio equal to or greater than a predetermined threshold, and including the non-search keyword as a dependency keyword for the search keyword Is characterized by not being searched.

According to the present invention, there is provided a search keyword dictionary that stores search keywords, and a text search method for searching text information including the search keywords using a computer,
A plurality of legitimate learning sentence information that does not belong to a specific category and a plurality of illegal learning sentence information that belongs to a specific category are accumulated in the learning sentence accumulation unit,
A concept dictionary in which word nodes are arranged in a tree shape is stored in the concept dictionary storage unit,
A first step of searching for learning sentence information including a search keyword using the learning sentence storing means;
A second step of extracting dependency keywords for the search keyword for learning sentence information including the search keyword;
A third step of extracting a plurality of dependency-related keywords that are subordinate concepts when viewed from the node corresponding to the dependency keyword using the concept dictionary storage means;
The number of valid learning text information and the number of illegal learning text information are counted for each of the dependency keywords and the dependency related keywords, and the number of valid learning text information with respect to the number of all learning text information is valid. A fourth step of calculating a ratio;
And a fifth step of registering a dependency keyword whose dependency ratio is equal to or greater than a predetermined threshold and a dependency-related keyword as a non-search keyword, and sentence information including the non-search keyword as a dependency keyword for the search keyword is It is characterized by not being searched.

  According to the sentence classification program, server, and method of the present invention, for example, when determining whether or not a sentence that is illegal or harmful belongs to a keyword registered in advance, the sentence information that is not illegal or harmful is classified as illegal or harmful. Can be reduced as much as possible, and the range of applicable sentences can be expanded as much as possible.

It is a functional block diagram of the text classification program in this invention. It is explanatory drawing showing the production | generation of the non-search keyword in this invention. It is explanatory drawing showing the classification | category of the object text information in this invention. It is the 1st explanatory view showing extraction of a dependency related keyword using a concept dictionary in the present invention. It is the 2nd explanatory view showing extraction of a dependency related keyword using a concept dictionary in the present invention. It is a system configuration | structure figure of the text analysis server in this invention. It is a sequence diagram of a system in the present invention. It is a graph showing the precision (Precision) with respect to recall (Recall). It is a graph showing F value with respect to the number of keywords of a dictionary.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a functional block diagram of a sentence classification program according to the present invention.

  According to FIG. 1, the sentence classification program is divided into a search keyword dictionary unit 10, a non-search dictionary generation function unit 11, and a target sentence classification function unit 12. The non-search dictionary generation function unit 11 includes a learning sentence storage unit 110, a learning sentence keyword search unit 111, a learning dependency analysis unit 112, a concept dictionary storage unit 113, a dependency related keyword extraction unit 114, and a legal ratio. A calculation unit 115 and a non-search keyword dictionary unit 116 are included. The target text classification function unit 12 includes a target text input unit 120, a target text keyword search unit 121, a target dependency analysis unit 122, a non-search keyword determination unit 123, and a text information classification unit 124. These functional units can be realized by executing a program that causes a computer installed in the apparatus to function. The processing flow of these functional components is also understood as a sentence classification method in the apparatus.

  The sentence classification program causes the computer to function so as to classify the target sentence information according to whether or not it belongs to a specific category. Specifically, it is searched whether or not the search text registered in the search keyword dictionary unit 10 is included in the target sentence information. For example, in the case of classification based on whether or not the text information is illegal / harmful, the search keyword stored in the search keyword dictionary unit 10 is an illegal / harmful keyword. For example, in the case of a crime notice, a keyword such as “explode” is stored in the search keyword dictionary unit 10. Of course, the search keyword stored in the search keyword dictionary unit 10 may be a keyword designated by the user. According to the present invention, even text information including a search keyword registered in the search keyword dictionary unit 10 can prevent text information that does not belong to a target specific category from being searched.

[Non-search dictionary generation function]
FIG. 2 is an explanatory diagram showing generation of a non-search keyword in the present invention. In the following, the content of FIG. 2 will be described together with the description of FIG.

The learning sentence accumulating unit 110 accumulates a large number of legitimate learning sentence information that does not belong to the specific category and a large number of inappropriate learning sentence information that belongs to the specific category. The validity / injustice is distinguished by a flag added to the learning sentence information.
For example, when the specific category is a group of illegal / harmful text information, the learning text storage unit 110 stores the following text information group in a mixed manner.
(1) Group of text information that does not belong to illegal or harmful categories
= "Valid" flag is added to each text information
= Group of legitimate learning text information (2) Group of text information belonging to illegal / harmful categories
= "Unfair" flag is added to each text information
= Illegal Learning Text Information Group Of course, when the specific category is a text information group designated by the user, the learning text accumulating unit 110 accumulates the following text information groups together.
(1) A group of text information that does not belong to a specific category specified by the user
= "Valid" flag is added to each text information
= Group of legitimate learning sentence information (2) Group of sentence information belonging to a specific category designated by the user
= "Unfair" flag is added to each text information
= Group of inappropriate learning text information

  The learning text keyword search unit 111 searches the learning text information including the search keyword stored in the search keyword dictionary unit 10 for a large number of learning text information stored in the learning text storage unit 110. According to FIG. 2, all learning text information including the search keyword “blast” is searched regardless of the flag “valid” / “unjust”. A large number of retrieved pieces of learning text information are output to the learning dependency analysis unit 112. Here, the learning sentence information that does not include the search keyword “blast” is not a target of subsequent processing.

  The learning dependency analysis unit 112 extracts a dependency keyword for the search keyword from a large number of pieces of learning text information including the search keyword. The learning dependency analysis unit 112 also includes morphological analysis necessary as a pre-stage of dependency analysis. A “morpheme” refers to the smallest meaningful unit among the constituent elements of a sentence. The morphological analysis has a dictionary in which “part of speech” and “reading” are registered for each “word”. According to the present invention, only one sentence including a search keyword (which may be shorter than one sentence) is divided into morphemes, and grammatical rules of “part of speech” and “reading” are formed using a dictionary, , Analyze the dependency.

  As the dependency analyzer, for example, a conventional CaboCha (for example, see Non-Patent Document 1) or KNP (for example, see Non-Patent Document 2) may be used. “CaboCha” is a typical Japanese dependency analyzer, which uses a definitive analysis algorithm (Cascaded Chunking Model) that does not perform backtracking based on a support vector machine. “KNP” is a Japanese sentence syntactic / case analysis system for Japanese sentences. The morpheme sequence of the morpheme analysis system JUMAN is input, and the dependency relations and case relations between clauses and basic phrases are output. Technology.

According to FIG. 2, the following dependency keywords are extracted for the search keyword “blast”.
Learning text information a: “Unfair” flag added (belongs to illegal / harmful category)
"Blasting"<-"School"
Learning text information b: "Legacy" flag added (does not belong to illegal or harmful category)
"Blasting"<-"Mine"
Learning text information c: "Legacy" flag added (does not belong to illegal or harmful category)
"Blast"<-"No"
As a result, the following keyword set is output to the dependency related keyword extraction unit 114.
“Search keyword <-Dependency keyword: Valid / Unfair”
"Blasting <-School: Unfair"
"Blasting <-Mine: Justified"
"Blasting <-Useless: Justified"

  The concept dictionary storage unit 113 stores a concept dictionary. Here, the concept dictionary (thesaurus) refers to a dictionary in which words are classified according to semantic inclusion relations. The concept dictionary represents these relationships in a tree structure in which word nodes are arranged in a tree shape.

  The dependency related keyword extraction unit 114 inputs a keyword set from the learning dependency analysis unit 112. The dependency-related keyword extraction unit 114 extracts a plurality of dependency-related keywords that are subordinate concepts using a concept dictionary for the dependency keyword to which the “unfair” flag is added. For example, if there is a keyword set of a search keyword “blast” and a dependency keyword “school”, a dependency-related keyword for the dependency keyword “school” is extracted.

According to FIG. 2, the dependency-related keyword for the dependency keyword “school” is extracted from the concept dictionary as follows.
"School"->"ElementarySchool""TechnicalSchool"
The dependency-related keyword extraction unit 114 outputs the dependency keyword “school” and the dependency-related keywords “elementary school” and “special school” for the search keyword “explode” to the correct ratio calculation unit 115.

  Note that the dependency related keyword extraction unit 114 uses a concept dictionary to extract a plurality of dependency related keywords as subordinate concepts, even for dependency keywords to which a “valid” flag is added. Also good. In this case, a large number of keywords with a high legitimate ratio are registered in the non-search keyword dictionary unit 116. As a result, the non-search keyword dictionary unit 116 can accumulate a large number of non-search keywords, and can improve the determination accuracy of the non-search keywords.

  The correct ratio calculation unit 115 counts the number of valid learning text information and the number of illegal learning text information for each dependency keyword and dependency-related keyword, and calculates the correct learning text information for all the learning text information. Calculate the correct proportion of numbers. Based on a large amount of learned text information, a large number of valid / invalid dependency keywords for one search keyword are extracted. Then, a legitimate ratio is calculated for each dependency keyword and dependency-related keyword with respect to the search keyword.

According to FIG. 2, the correct ratio is calculated for each dependency keyword and dependency-related keyword as follows.
"Blasting"<-"School": Number of legitimate learning text information = 40
Number of inappropriate learning text information = 1960
Legitimate ratio = 40/2000 = 2%
"Blasting"<-"Primaryschool": Number of legal learning text information = 10
Number of inappropriate learning text information = 990
Legitimate ratio = 10/1000 = 1%
"Blasting"<-"Vocationalschool": Number of legitimate learning text information = 5
Number of inappropriate learning text information = 95
Legitimate ratio = 5/100 = 5%
"Blasting"<-"Coalmine": Number of legitimate learning text information = 90
Number of illegal learning text information = 10
Legitimate ratio = 90/100 = 90%
"Blast"<-"Mine": Number of legitimate learning text information = 99
Number of illegal learning text information = 1
Legitimate ratio = 99/100 = 99%
Then, the legal ratio calculation unit 115 outputs the legal ratio for each dependency keyword to the non-search keyword dictionary unit 116.
“Blast” <-“School”: 2%
<-"Elementary school": 1%
<-"Vocational school": 5%
<-"Coal mine": 90%
<-"Mine": 99%

  The non-search keyword dictionary unit 116 registers dependency keywords and dependency-related keywords whose legal ratio is equal to or greater than a predetermined threshold. According to FIG. 2, it is assumed that the predetermined threshold is set to 70% or more. In this case, the dependency keyword “coal mine” whose legitimate ratio is 70% or more with respect to the search keyword “blasting” and the “mine” that is a subordinate concept of “coal mine” are the dependency on the search keyword “blasting”. Is registered in the non-search keyword dictionary unit 116 as a non-search keyword.

The non-search keyword dictionary unit 116 may store not only the dependency keywords whose legal ratio is equal to or greater than a predetermined threshold value, but may also register the legal ratios in association with every dependency keyword. . Accordingly, the predetermined threshold as a level derived as a non-search keyword can be made variable by user designation.
Search keyword “Blast” <-Non-search / Dependency keyword “School”: 2%
<-Non-search and dependency related keyword "elementary school": 1%
<-Non-search / dependency related keyword “vocational school”: 5%
<-Non-search / dependency keyword "coal mine": 90%
<-Non-Search / Dependency Keywords "Mine": 99%
<-Non-Search / Dependency Keyword "Dame": 60%

[Target sentence classification function section]
FIG. 3 is an explanatory diagram showing classification of target sentence information in the present invention. In the following, the content of FIG. 3 will be described together with the description of FIG.

  The target sentence input unit 120 inputs target sentence information to be analyzed. The target sentence information may be input in response to a user operation, or may be received via a network. The target text information received via the network is, for example, text information described by an unspecified number of users in a publicly available blog, a bulletin board, and / or a review comment.

  The target text keyword search unit 121 determines whether the target text information includes the search keyword stored in the search keyword dictionary unit 10. According to FIG. 3, the search keyword “blast” is included in the sentence “.. Therefore, the target sentence information is output to the target dependency analysis unit 122 with the search keyword.

The target dependency analysis unit 122 extracts a dependency keyword for the search keyword from the target sentence information including the search keyword. According to FIG. 3, the dependency keyword “mine” is extracted for the search keyword “explosion”.
Target text information: “Blast” <-“Mine”
The extracted dependency keyword “mine” is output to the non-search keyword determination unit 123 together with the target sentence information.

  The non-search keyword determination unit 123 determines whether or not the extracted dependency keyword matches the non-search dependency keyword registered in the non-search keyword dictionary unit 116. According to FIG. 3, “Blasting” <-“Mine” is registered in the non-search keyword dictionary unit 116. Therefore, the non-search keyword determination unit 123 determines that the extracted dependency keyword “mine” matches the non-search dependency keyword “mine”. This determination result is output to the text information classification unit 124.

  The non-search keyword determination unit 123 matches the extracted dependency keyword with the dependency keyword registered in the non-search keyword dictionary unit 116, and the legitimate ratio in the dependency keyword is equal to or greater than a predetermined threshold. In this case, it may be determined that the dependency keyword is true. For example, for the matching dependency keyword “mine”, the legal percentage 99% registered in the non-search keyword dictionary unit 116 is equal to or greater than the predetermined threshold 70%, and thus is adopted as a non-search keyword. Conversely, non-search keywords with a predetermined threshold value less than 70% are not adopted.

  The text information classification unit 124 includes target text information determined to be false by the target text keyword search unit 121, and target determined to be true by the target text keyword search unit 121 and determined to be true by the non-search keyword determination unit 123. Text information is classified as text information that does not belong to a specific category.

  FIG. 4 is a first explanatory diagram showing extraction of dependency-related keywords using the concept dictionary in the present invention.

  Assume that the dependency keyword “school” is extracted for the search keyword “blast”. Here, it is assumed that the keyword set of “school, blast” has a low legitimate ratio. At this time, even if the “school” is a subordinate concept “elementary school” or “technical school”, the legitimate ratio for the search keyword “blasting” is as low as “school”.

  According to FIG. 4, the concept dictionary is represented by a tree structure. The tree-structured concept dictionary uses keywords as components, and each of these components is called a node. The nodes are arranged in a hierarchy. A word having a smaller value of n in the hierarchy n becomes a higher concept. For example, “school” at level 3 is a superordinate concept of “elementary school” and “specialty school” at level 4.

  On the other hand, the dependency-related keyword is a keyword included in the subordinate concept of the dependency keyword, for example, “elementary school” and “XX elementary school” included in the subordinate concept of “school”. On the other hand, it is a keyword included in the superordinate concept of the dependency keyword, for example, “public facilities” contained in the superordinate concept of “school”. Even if “Blasting public facilities” is just as low as “Blasting schools”.

  A threshold may be set in advance as the number of keywords extracted as dependency-related keywords. For example, keywords included in a lower concept are first extracted, and keywords included in a subtree of a higher concept may be extracted as long as the number of extracted keywords is equal to or less than a threshold value. For example, when the dependency keyword is “school” and the threshold value is set to 7, the number of keywords is less than or equal to the threshold value only for “elementary school”, “technical school”, and “XX elementary school” included in the subordinate concept. Therefore, “station” and “hospital” included in the subtree of the superordinate concept “public facilities” are also extracted.

  The dependency-related keyword is extracted by using a concept dictionary to extract a plurality of dependency-related keywords that are lower-level concepts when viewed from a higher-level concept node at least one level higher than a node corresponding to the dependency keyword. There may be. For example, when the dependency keyword is “school”, “station”, “subway station”, “private railway station”, and “hospital”, which are subordinate concepts, are extracted from the “public facilities” node one level above. For example, even if it is “Blasting a station”, the legitimate ratio is as low as “Blasting a school”.

  In addition, the dependency-related keyword is extracted based on the height of the hierarchy n of the node corresponding to the dependency keyword, and the number of higher-level concepts viewed from the hierarchy n is determined as a lower concept. A plurality of dependency-related keywords may be extracted using a concept dictionary. For example, when the dependency keyword is “school” of [node 3], a dependency-related keyword that is a lower concept of “public facilities” of [node 2], which is a higher concept node on one layer, is extracted. . On the other hand, when the dependency keyword is “elementary school” of [node 4], a dependency related keyword that is a subordinate concept of “public facilities” of [node 2], which is a higher concept node on the second hierarchy, is extracted. To do.

  Further, the dependency-related keyword extraction is based on the part-of-speech of the dependency keyword, and determines how many higher-order nodes are viewed from the hierarchy n, and a plurality of dependency-related keywords that are lower concepts are determined. You may extract using a concept dictionary. Moreover, based on the part of speech of a dependency keyword, you may determine whether it extracts as a dependency related keyword.

Depending on the part-of-speech of the dependency keyword, the keyword included in the subtree of the superordinate concept may not be extracted. For example, in the case of a noun, dependency-related keywords that are subordinate concepts when viewed from a superordinate concept node on one layer are likely to be semantically similar. On the other hand, in the case of verbs, if extracted in the same way, the dependency-related keywords tend to become larger in semantic shift. The following are examples of the noun “child” and the verb “kill”.
"Children"->"Daughter""Firstchild""Grandmotherchild""Hiddenchild""Childbrother""Seconddaughter""Twins"
"Kill"->"Assassinate""Insecticide""Damage""Crush""Thinning""Remove"

  FIG. 5 is a second explanatory diagram showing the extraction of dependency-related keywords using the concept dictionary in the present invention.

The score s may be used for extracting dependency-related keywords. According to FIG. 5, the score s (w ′) of the extracted dependency-related keyword is calculated by the following equation.
s (w ′) = s (w) × f (d (w ′))
f (x) = c ^x
s (w): dependency keyword score d (w ′): weighted distance between dependency keyword and dependency related keyword c: constant (for example, 0.8)

  As for the weighted distance d (w ′), the distance to the keyword that is a higher concept has a larger weight value than the distance to the keyword that is a lower concept. For example, the distance to the keyword that is a higher concept viewed from the dependency keyword is 0.6, whereas the distance to the keyword that is a lower concept is 0.2.

For example, the following example will be described.
w = “school”, s (w) = 2
w '= “elementary school”
d (w ′): Weighting distance of “elementary school” with respect to “school” At this time, the score of “elementary school” with respect to “school” is expressed as follows.
d (w ′) = 0.2 + 0.2 = 0.4
f (d (w ′)) = c ^0.4 = 0.8 ^0.4 = 0.92
s (w ′) = s (w) × f (d (w ′)) = 2 × 0.8 ^0.4 = 1.83

Similarly, the score of “station” for “school” is expressed as follows.
d (w ′) = 0.6 + 0.2 = 0.8
f (d (w ′)) = c ^0.8 = 0.8 ^0.8 = 0.84
s (w ′) = s (w) × f (d (w ′)) = 2 × 0.8 ^0.8 = 1.67

  Thus, when viewed from the “school” of the score 2, the “station” of the score 1.67 is farther than the “elementary school” of the score 1.83. By setting a threshold value for this weighted distance, the range of the weighted distance to be expanded from the dependency keyword can be determined. When the threshold is 0.5, for example, when viewed from “school” with a score of 2, “elementary school” with a score of 1.83 and “station” with a score of 1.67 are included in the range of the weighted distance. As a result, “School-> Blast” is expanded to “Elementary school-> Blast” and “Station-> Blast”.

  FIG. 6 is a system configuration diagram of the sentence analysis server according to the present invention.

  According to FIG. 6, the sentence analysis server 1 includes the search keyword dictionary unit 10, the non-search dictionary generation function unit 11, the target sentence classification function unit 12, and the communication interface unit 13 described above. The sentence analysis server 1 communicates with the Web server 2 via the Internet. In addition, the Web server 2 accumulates and discloses Web text information as target text information transmitted from the terminal 3. The target sentence classification function unit 12 of the sentence analysis server 1 receives target sentence information from a number of Web servers 2 via the communication interface unit 13. The target sentence classification function unit 12 classifies the target sentence information into a specific category by using the search keyword dictionary and the non-search keyword dictionary generated by the non-search dictionary generation function unit 11. The target sentence information received from the Web server 2 can be classified into, for example, a specific category that is illegal or harmful.

  FIG. 7 is a sequence diagram of the system according to the present invention.

(S701) The non-search dictionary generation function unit 11 accumulates, as learning text information, a large number of legitimate learning text information that does not belong to a specific category and a large number of inappropriate learning text information that belongs to a specific category. First, learning sentence information including a search keyword registered in a search keyword dictionary is searched from a large number of learning sentence information.
(S702) With respect to the learning sentence information including the search keyword, a dependency keyword for the search keyword is extracted.
(S703) Using a concept dictionary in which word nodes are arranged in a tree shape, a plurality of dependency-related keywords that are subordinate concepts are extracted from the node corresponding to the dependency keyword.
(S704) For each dependency keyword, the number of legitimate learning sentence information and the number of illegal learning sentence information are counted, and the legitimate ratio of the number of legitimate learning sentence information to the number of all learning sentence information is calculated.
(S705) Then, a dependency keyword whose legitimate ratio is equal to or greater than a predetermined threshold is registered as a non-search keyword.
(S706) The terminal 3 transmits the target text information to the Web server 2 in accordance with a user operation.
(S707) The sentence analysis server 1 receives target sentence information to be analyzed from the Web server 2. The target sentence information is output to the target sentence classification function unit 12.
(S708) It is determined whether the target sentence information includes the search keyword stored in the search keyword dictionary.
(S709) For the target sentence information including the search keyword, a dependency keyword for the search keyword is extracted.
(S710) It is determined whether or not the extracted dependency keyword matches the non-search dependency keyword registered in the non-search keyword dictionary.
(S711) Then, the target sentence information determined to be false by S708 and the target sentence information determined to be true by S708 and determined to be true by S710 are classified as sentence information not belonging to a specific category. Thus, when the dependency keyword for the search keyword is a non-search keyword registered in the non-search keyword dictionary, the sentence information is not searched as not belonging to the specific category.

  Finally, performance comparison evaluation results between the present invention and the prior art will be described. MeCab was used as a morphological analyzer, Cabocha was used as a dependency analyzer, and an EDR electronic dictionary was used as a concept dictionary. It consists of 400,000 learning documents (40,000 illegal and harmful articles, 360,000 harmless articles) that have been manually labeled, and 400,000 articles to be evaluated (40,000 illegal and harmful articles, 360,000 harmless articles). We used 800,000 commercial blog documents.

  The retrieval performance of an information retrieval system is evaluated by precision and recall. The relevance ratio is an index of accuracy indicating how many documents suitable for retrieval are included in the set obtained as a retrieval result. The recall is an index of completeness indicating how many documents (correct answer documents) that are suitable as search results among the documents to be searched can be searched. Here, there is a trade-off relationship that if the precision is increased, the recall is decreased, and if the precision is increased, the precision is decreased. Therefore, it is also preferable to evaluate the search performance using an F value (F-measure) that is a harmonic average of the precision and the recall. A higher F value means better search performance.

  FIG. 8 is a graph showing the precision (Precision) with respect to the recall (Recall). By using the concept dictionary, the recall rate improved by up to 4.2% and the precision rate improved by up to 2.0% compared to the prior art.

  FIG. 9 is a graph showing the F value with respect to the number of keywords in the dictionary. By using the concept dictionary, the F value improved by up to 3.9%. This is thought to be an effect of being able to correctly determine many new expressions by extending the concept dictionary based on a small number of dependency clauses obtained from the learning document. It is done.

  As described above in detail, according to the sentence classification program, server, and method of the present invention, it is not illegal / harmful when, for example, it is determined whether or not it belongs to an illegal / harmful category by a keyword registered in advance. The sentence information can be reduced from being classified into illegal and harmful categories as much as possible, and the range of applicable sentences can be expanded as much as possible. According to the present invention, in addition to the search keyword dictionary, a non-search keyword dictionary in which non-search keywords having a dependency relationship with the search keyword are registered. The non-search keyword dictionary can register not only dependency keywords but also related keywords as non-search keywords using a concept dictionary. As a result, even if the text information includes an illegal / harmful search keyword, it is determined that the text is not illegal / harmful text information if the dependency relation keyword matches the non-searched search keyword. According to the present invention, since the dependency relationship of search keywords is also analyzed, excessive determination is suppressed and classification accuracy is improved.

  In addition, even if the present invention is determined to be illegal / harmful text information, the present invention removes what should be determined not to be illegal / harmful text information, and illegal / harmful text information. Even if it is determined that it is not, it can be used to remove what should be determined to be illegal / harmful text information.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Text analysis server 10 Search keyword dictionary part 11 Non-search dictionary production | generation function part 110 Learning sentence storage part 111 Learning sentence keyword search part 112 Learning dependency analysis part 113 Concept dictionary memory | storage part 114 Dependence related keyword extraction part 115 Validity ratio calculation part 116 Non-search keyword dictionary part 12 Target sentence classification function part 120 Target sentence input part 121 Target sentence keyword search part 122 Target dependency analysis part 123 Non-search keyword judgment part 124 Text information classification part 13 Communication interface part 2 Web server 3 Terminal

Claims (14)

A text search program having search keyword dictionary means for storing search keywords, and causing a computer to function to search text information including the search keywords,
A learning sentence accumulating means for accumulating a plurality of legitimate learning sentence information not belonging to a specific category and a plurality of illegal learning sentence information belonging to the specific category,
Learning sentence keyword search means for searching learning sentence information including the search keyword using the learning sentence storage means;
A learning dependency analysis unit that extracts a dependency keyword for the search keyword for learning sentence information including the search keyword;
Concept dictionary storage means in which word nodes are arranged in a tree shape;
Using the concept dictionary storage unit, a dependency-related keyword extracting unit that extracts a plurality of dependency-related keywords that are subordinate concepts when viewed from a node corresponding to the dependency keyword;
For each keyword of the dependency keyword and the dependency related keyword, the number of the legitimate learning sentence information and the number of the illegal learning sentence information are counted, and the legitimate learning sentence with respect to the number of all the learning sentence information. A legal proportion calculating means for calculating a legal proportion of the number of information;
The computer functions as a non-search keyword dictionary means for registering the dependency keyword and the dependency-related keyword as a non-search keyword, and the non-search keyword is included as a dependency keyword for the search keyword. A text search program that prevents text information being searched from being searched.   The dependency-related keyword extracting unit uses the concept dictionary storage unit to view a plurality of dependency-related subordinate concepts as viewed from a higher concept node that is at least one level higher than a node corresponding to the dependency keyword. The text search program according to claim 1, wherein the computer is caused to function so as to extract a keyword.   The dependency-related keyword extracting means uses the concept dictionary storage means to view from a higher-level concept node above the hierarchy n based on the height of the hierarchy n of the node corresponding to the dependency keyword. The sentence search program according to claim 2, wherein the computer is caused to function to determine whether or not.   The dependency-related keyword extraction means uses the concept dictionary storage means based on the part-of-speech of the dependency keyword to determine the number of higher-level concept nodes to be viewed from the hierarchy n. The sentence search program according to claim 2, wherein the sentence search program is made to function.   5. The dependency-related keyword extracting unit causes the computer to function as to whether or not to extract the dependency-related keyword based on a part of speech of the dependency-related keyword. Text search program. Using the concept dictionary storage means, further comprising search related keyword extraction means for extracting a plurality of search related keywords that are lower concepts than the node corresponding to the search keyword;
The learning sentence keyword search means further searches the learning sentence information including the search related keyword using the learning sentence storage means,
6. The learning dependency analysis unit causes the computer to function so as to further extract dependency keywords for the search related keyword for learning sentence information including the search related keyword. The text search program according to item 1. A target sentence input means for inputting target sentence information to be analyzed;
Target text keyword search means for determining whether or not the target text information includes the search keyword stored in the search keyword dictionary means;
For the target sentence information including the search keyword, target dependency analysis means for extracting a dependency keyword for the search keyword;
Non-search keyword determination means for determining whether or not the extracted dependency keyword matches the non-search dependency keyword registered in the non-search keyword dictionary means;
The target sentence information determined to be false by the target sentence keyword search means, and the target sentence information determined to be true by the target sentence keyword search means and determined to be true by the non-search keyword determination means, The sentence search program according to any one of claims 1 to 6, wherein a computer is caused to function as sentence information classification means for classifying the sentence information as not belonging to the specific category. The non-search keyword dictionary means associates and registers the legitimate ratio calculated by the legitimate ratio calculator for each registered dependency keyword,
The non-search keyword determination unit matches the extracted dependency keyword with the dependency keyword registered in the non-search keyword dictionary unit, and the valid ratio of the dependency keyword is equal to or greater than a predetermined threshold. In this case, the sentence search program according to claim 7, wherein the computer is caused to function so as to determine that the dependency keyword is true. The search keyword is a keyword set for the purpose of categorization,
9. The text search program according to claim 1, wherein the computer is caused to function so that the specific category in the learning text storage means is a group of text information based on the category classification. The search keyword is an illegal or harmful keyword,
9. The text search program according to claim 1, wherein the computer is caused to function so that the specific category in the learning text storage means is a group of illegal and harmful text information. The search keyword is a keyword specified by the user,
The sentence search program according to any one of claims 1 to 8, wherein the computer is caused to function so that the specific category in the learning sentence storage unit is a group of sentence information designated by a user.   The target text information causes a computer to function as text information described by an unspecified number of users in a blog (Weblog), bulletin board, and / or word-of-mouth comment published via a network. The text search program according to any one of claims 1 to 11. A sentence analysis server that has a search keyword dictionary unit that stores search keywords, and that searches sentence information including the search keyword from sentence information acquired from another public server via a network,
A learning sentence accumulating means for accumulating a plurality of legitimate learning sentence information not belonging to a specific category and a plurality of illegal learning sentence information belonging to the specific category,
Learning sentence keyword search means for searching learning sentence information including the search keyword using the learning sentence storage means;
A learning dependency analysis unit that extracts a dependency keyword for the search keyword for learning sentence information including the search keyword;
Concept dictionary storage means in which word nodes are arranged in a tree shape;
Using the concept dictionary storage unit, a dependency-related keyword extracting unit that extracts a plurality of dependency-related keywords that are subordinate concepts when viewed from a node corresponding to the dependency keyword;
For each keyword of the dependency keyword and the dependency related keyword, the number of the legitimate learning sentence information and the number of the illegal learning sentence information are counted, and the legitimate learning sentence with respect to the number of all the learning sentence information. A legal proportion calculating means for calculating a legal proportion of the number of information;
A non-search keyword dictionary means for registering the dependency keyword having the valid ratio equal to or higher than a predetermined threshold and the dependency-related keyword as a non-search keyword, and a non-search keyword is included as a dependency keyword for the search keyword A sentence analysis server characterized in that the sentence information is not searched. A text search method having a search keyword dictionary storing search keywords and searching text information including the search keyword using a computer,
A plurality of legitimate learning sentence information that does not belong to a specific category and a plurality of illegal learning sentence information that belongs to the specific category are accumulated in the learning sentence accumulation unit,
A concept dictionary in which word nodes are arranged in a tree shape is stored in the concept dictionary storage unit,
A first step of searching for learning text information including the search keyword using the learning text storage means;
A second step of extracting a dependency keyword for the search keyword for learning sentence information including the search keyword;
A third step of using the concept dictionary storage means to extract a plurality of dependency-related keywords that are subordinate concepts when viewed from a node corresponding to the dependency keyword;
For each keyword of the dependency keyword and the dependency related keyword, the number of the legitimate learning sentence information and the number of the illegal learning sentence information are counted, and the legitimate learning sentence with respect to the number of all the learning sentence information. A fourth step of calculating a legitimate percentage of the number of information;
A fifth step of registering the dependency keyword with the valid ratio equal to or greater than a predetermined threshold and the dependency related keyword as a non-search keyword, and a non-search keyword is included as a dependency keyword for the search keyword A sentence search method characterized in that the sentence information is not searched.

Images