KR20080017686A - Method for extracting subject and sorting document of searching engine, computer readable record medium on which program for executing method is recorded - Google Patents

Abstract

A method for extracting subjects and sorting documents in a search engine, and a computer-readable recording medium storing a program thereof are provided to enable a user to access desired information conveniently/quickly by selecting atypical/various subjects not classified in a manual mode, classify the target documents into each subject, and determine whether the searched document is suitable for the subject. A relation degree representing that respective keywords are selected at the same time is measured for the keywords included in target documents. A convergence relation degree between a word set about the predetermined keyword and the word set related to other keywords is measured. The keyword is selected as a subject when the convergence relation degree is higher than a specific value. A naive Bayesian probability is calculated by performing naive Bayesian training for training documents and each keyword included in the target documents. A vector size of each keyword included in the training and target document is calculated. A distance between the vector size of each keyword of the training and target document is calculated. Similarity of each keyword is calculated by multiplying the naive Bayesian probability and the distance. A ranking value is calculated by processing the similarity of each keyword included in the target document.

Description

METHOD FOR EXTRACTING SUBJECT AND SORTING DOCUMENT OF SEARCHING ENGINE, COMPUTER READABLE RECORD MEDIUM ON WHICH PROGRAM FOR EXECUTING METHOD IS RECORDED}

1 is a block diagram of a search engine according to the present invention;

2 is a diagram illustrating an electronic document related to a specific subject and a keyword of each electronic document;

3 is a result showing the relationship between each keyword of the document A, document B, document C and the relationship between each keyword,

4 is a graph showing the relationship between the keyword distance and frequency;

5 is a flowchart illustrating a subject generation process and a document classification process according to the search engine of FIG. 1.

Explanation of symbols on the main parts of the drawings

10: search unit 20: subject generation module

21: Keyword storage unit 23: Association degree calculation unit

25: convergence degree calculation unit 30: document classification module

31: naive Bayesian part 33: keyword vector calculation unit

35: distance calculation unit 37: ranking calculation unit

The present invention relates to a method for generating a subject and a document classification method of a search engine. More particularly, the present invention relates to a search engine for classifying documents to be classified according to a qualitative similarity to a subject. It relates to the topic creation and document classification

The commercialization of the Internet has made it possible for many users to easily and quickly search for information on a variety of topics using the Internet. However, it takes time and effort to find the optimal information that users want from the vast amount of information.

Recently, search engines have introduced various search methods so that users can find the optimal information quickly. One of them is to provide a relevance ranking. Relevance ranking refers to the relative evaluation of the degree to which a searched electronic document is related to a search word designated by a user in a given network compared to other electronic documents. The relevance ranking includes a method of evaluating the number of times a predetermined search word appears in an electronic document, a position of the search word in the electronic document, a web page structure analysis method, a key word enumeration method, and the like.

The relevance ranking provided in this way provides only a probabilistic value of whether the electronic document shown by the search result matches the corresponding subject, and is determined by the frequency of the search word entered by the user. It is difficult to be a criterion for determining whether an electronic document corresponds to a subject.

On the other hand, a subject for classifying documents to be searched is generally selected and defined manually by a developer or operator of a search engine. Thus, only a limited number of topics can be classified. However, an electronic document generated by an unspecified number of users with various hierarchies, occupations, and educational attainments may cover a very wide range of topics. Therefore, the manual classification of the current subject cannot cover all the subjects, and the electronic document in the area where the subject is not classified is classified as an ambiguous subject. Accordingly, at the time of searching, the electronic document may be included in a completely wrong topic or may not be searched by a desired user. In other words, the ambiguity of classification reduces the reliability of the search results.

Accordingly, it is necessary to find a way to accurately classify electronic documents on various topics by presenting a method for generating an informal and extensive subject. In addition, it is necessary to search for a method of improving the reliability of the search by classifying the electronic documents according to the degree appropriate to the subject and showing the ranking how the classified electronic documents correspond to the desired subject.

Accordingly, an object of the present invention is to provide a method for generating a topic and classifying a document of a search engine that can generate an informal and broad topic.

On the other hand, another object of the present invention is to provide a method for generating a subject and a document classification method of a search engine that can improve the reliability of the search by classifying the electronic document according to the degree suitable for the subject.

In accordance with an aspect of the present invention, there is provided a method for measuring the degree of association of each of the keywords simultaneously selected for at least one keyword included in a document to be classified: a word set associated with any keyword and another keyword. Measuring a degree of convergence association, which is an degree of association between sets of words associated with; And selecting the keyword as a subject when the convergence connection degree is equal to or more than a predetermined schedule.

The correlation may be calculated by Equation 1.

[Equation 1]

Where LinkCount (w ₁ , w ₂ ) is the number of times the keywords w ₁ and w ₂ were tagged at the same time, WordCount (w ₁ ) is the total number of times the keyword w ₁ was used as a keyword, and WordCount (w ₂ ) was the keyword w ₂ The total number of times used as a keyword.

The convergence correlation is calculated by Equation 2;

[Equation 2]

과,

모두 미리 설정된 최소 연관도보다 크거나 같으면, PositiveVal(w₁, w₂, w_i)에 값을 대입하고, 아니면, NegativeVal(w₁, w₂, w_i)에 값을 대입하는 것이 바람직하다. here,

and,

If all are greater than or equal to the preset minimum correlation, it is preferable to assign a value to PositiveVal (w ₁ , w ₂ , w _i ), or to assign a value to NegativeVal (w ₁ , w ₂ , w _i ).

PositiveVal (w ₁ , w ₂ , w _i ) and NegativeVal (w ₁ , w ₂ , w _i ) may be calculated by Equations 3 and 4, respectively.

[Equation 3]

[Equation 4]

The convergence association degree is closer to 1 as the keywords are similar, and closer to 0 as the keywords are different.

And calculating a naive Bayesian probability by performing naïve Bayesian learning on each keyword included in the classification target document and the keywords to be classified.

And calculating a vector size of each of the keywords included in the learning document and the document to be classified.

The calculating of the vector size of each keyword may include calculating a weight of each keyword by multiplying the number of occurrences of each keyword in a corresponding document by an inverse document frequency, and calculating the weight of each keyword by the weight of each keyword. It may include the step of calculating by dividing by the sum of the weights of all the words in the included learning document or classification target document.

Calculating a distance of a keyword that is a difference between a vector size of each keyword on the learning document and a vector size of each keyword on the classification target document.

It is preferable that the distance of the keyword has a higher value as the frequency of use of the keyword in the learning document and the classification target document is similar.

Calculating a similarity value for each keyword by multiplying a naive Bayesian probability value Nb and a distance DP for each keyword; Calculating one ranking determination value through a predetermined process of the similarity value for each keyword included in the classification target document; Calculating a ranking of each classification target document by using the ranking determination value of each classification target document, and classifying the classification according to a subject.

Another object of the present invention for achieving the above object is to provide a computer-readable recording medium containing a program capable of performing the method.

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

1 is a block diagram of a search engine according to the present invention.

The search engine can be used in a variety of computing systems, including server computers, personal computers, handhelds, laptop devices, programmable electronics, and the like.

The search engine includes a search unit 10, a subject generation module 20, and a document classification module 30. The search engine automatically generates a subject, and classifies documents classified according to a subject to calculate a ranking.

The search unit 10 searches electronic documents of all web pages accessible from a computing system equipped with a search engine, and the searched documents are provided to the subject generation module 20 to generate a subject. At this time, each electronic document includes a tagging keyword (hereinafter referred to as a keyword) that the user adds to the electronic document to classify and store the document.

The subject generation module 20 automatically generates a plurality of subjects of interest from an electronic document having an unstructured and broad subject using keywords included in the electronic document.

The subject generation module 20 includes a keyword storage unit 21 for storing keywords, an association degree calculation unit 23 for calculating an association degree between keywords, and a convergence association degree for calculating an association degree between sets associated with any keyword. The calculator 25 is included.

The keyword storage unit 21 stores each keyword included in the electronic document and the number of occurrences of each keyword in the electronic document, and all the keywords stored in the keyword storage unit 21 can be generated as a subject candidate group. Becomes

On the other hand, since each keyword included in the electronic document is directly input by the user, slang, adult terms, typos, etc. may be included, and such terms may be inappropriate keywords. Because of these inadequate keywords, it is not desirable to be selected as the subject by simply the number of occurrences.

The relevance calculator 23 detects the degree to which the keywords included in the electronic document are related, thereby preventing the inappropriate keywords from being selected as the subject. For example, in a document related to 'baseball', terms related to 'baseball' are input as keywords. In FIG. 2, as an embodiment, an electronic document related to a specific subject and keywords of each electronic document are displayed.

In case of Document A, 'Park Chan-ho Baseball', 'Baseball' and 'Park Chan-ho Kim Byung-hun' are entered as keywords, and Document B is 'Baseball Lee Jong-bum', 'Baseball Japan', and 'Japanese Jong-beom' as keywords. C has the keywords 'anime anime', 'japanese manga' and 'anime manga' as keywords. Here, Document A and Document B have the same theme of 'baseball,' and accordingly, common keywords are frequently used. Therefore, Document A and Document B are highly related. On the other hand, Document A and Document C have different themes, such as 'baseball' and 'japanese manga', respectively, so there is no keyword in common with each other, and thus the correlation between the keywords in both documents is low. On the other hand, keywords such as profanity, adult terms, typos, etc. do not belong to a specific theme, and thus have low association with other keywords. Accordingly, if a topic is selected using the number of occurrences of the keyword and the degree of association between the keywords, inappropriate terms or keywords can be prevented from being selected as the topic.

The association degree calculation unit 23 calculates R (w ₁ , w ₂ ) which is the degree of association between the keywords w ₁ and w ₂ using the following equation ( ₁ ).

[Equation 1]

Where LinkCount (w ₁ , w ₂ ) is the number of times the keywords w ₁ and w ₂ were tagged at the same time, WordCount (w ₁ ) is the total number of times the keyword w ₁ was used as a keyword, and WordCount (w ₂ ) was the keyword w ₂ The total number of times used as a keyword.

Meanwhile, keywords included in each document of FIG. 2 are {Park Chan-ho, baseball, Kim Byung-hun, Lee Jong-bum, Japan, Annie, cartoon}, and the number of occurrences of each keyword is {2,4,1,2,4,2,2}. to be. Here, if you calculate the degree of association between the keywords 'Park Chan-ho' and 'baseball', the number of times 'Park Chan-ho' and 'Baseball' are tagged at the same time, the total number of times 'Park Chan-ho' is used as a keyword is 2 times, The total number of times a baseball is used as a keyword is four times.

Therefore, R (Park Chan-ho, baseball) = (1/2 + 1/4) / 2 = 3/8 = 0.375.

Similarly, the results of calculating the degree of association between the keywords of Document A, Document B, and Document C are shown in FIG.

As shown, the keyword 'Japan' has one link each with {Lee Bum, Annie, Manga, Baseball}, and the relationship between 'Japan' and the keyword connected by each link is similar. If the keywords "Japan" and "Lee Jong-bum" are included in a similar theme, the keyword set linked to "Japan" and the keyword set linked to "Lee Jong-bum" will be linked to the same keyword. However, if 'Japan' and 'Lee Jong-bum' deal with different subjects, the keyword set linked to 'Japan' and the keyword set linked to 'Lee Jong-bum' will have more links with keywords belonging to different sets. The degree of association between the keyword sets to which an arbitrary keyword belongs is convergence relation, and the degree of convergence is reflected in the association to further refine the association between keywords.

The convergence association degree calculation unit 25 calculates CR (w ₁ , w ₂ ) which is the convergence association degree between the keywords w ₁ and w ₂ using the following equations ( ₂ ) to ( ₄ ).

[Equation 2]

과,

모두 미리 설정된 최소 연관도보다 크거나 같으면, 즉, w₁과 w₂에 연결된 w_i의 연관도가 최소 연관도 이상인 경우에는 PositiveVal(w₁, w₂, w_i) 측에 값을 넣고, w₁과 w₂에 연결된 w_i의 연관도가 최소 연관도 미만인 경우인 NegativeVal(w₁, w₂, w_i) 측에 값을 넣는다. here,

and,

All pre-set minimum Relevancy greater than or equal to, i.e., w _1, and if also less than w _i Relevancy a minimum association attached to w _2, put a value on PositiveVal (w _1, w _2, w _i) side, w It is ₁ and w _i w ₂ is connected to the association of the value placed in the _{NegativeVal (w 1, w 2,} w i) is less than the minimum associated side case.

For example, if the minimum association degree is set to 0.3 for FIG. 3, w ₁ is called 'baseball', and w ₂ is called 'cartoon', 'Japan' which is a keyword w _i connected to 'baseball' and 'cartoon''Is associated with' manga 'is 0.3 or more, but' association with baseball is less than 0.3, so 'Japan' is included in the NegativeVal (w ₁ , w ₂ , w _i ).

Accordingly, if there are many words that exceed the minimum degree of relevance among the words w _i connected to the keywords w ₁ and w ₂ , the convergence correlation increases, and when there are many words that do not exceed the minimum degree of relevance among the words connected to the keyword, convergence occurs. The degree of association is small. Therefore, the convergence association degree is closer to 1 as each keyword is similar, and closer to 0 as each keyword is different.

PositiveVal (w ₁ , w ₂ , w _i ) and NegativeVal (w ₁ , w ₂ , w _i ) can be calculated by the following equations (3) and (4), respectively.

[Equation 3]

[Equation 4]

The degree of convergence can more accurately determine the association between each keyword. Accordingly, a keyword whose convergence degree is above a predetermined value can be selected as a theme.

The document classification module 30 learns a learning document having a selected subject by using a naive Baysien method, which is a statistical technique, and indicates an association of a subject with a subject to be classified as an electronic document to be classified. Classify documents by calculating search rankings. The naive Bayesian method is a learning algorithm that explicitly calculates probabilities for hypotheses. It is a simple and high-class statistical technique that is frequently used for document classification.

The document classification module 30 includes a naive Bayesian unit 31 for calculating the naive Bayesian probability, a keyword vector calculator 33 for calculating the vector size of the keyword, a distance calculator 35 for calculating the similarity between documents, and a naive. A ranking calculator 37 calculates a ranking using Bayesian probabilities and distances.

The naive Bayesian unit 31 calculates naive Bayesian probabilities of words including a keyword selected as a subject from a pre-selected learning document. At this time, the words to calculate the naive Bayesian probability may be a word extracted in advance in relation to the subject, or all words included in the learning document, which can be variously selected by the designer or the user of the search engine.

The naive Bayesian unit 31 learns the keywords and words included in the learning document using the Naive Bayesian learning method, respectively, and calculates the probability that the keywords and words become themes when the keywords and words appear. It is called (Nb). Since the naive Bayesian learning method is a well-known technique that can be easily implemented by those skilled in the art, a detailed description thereof will be omitted.

The keyword vector calculator 33 calculates the vector size of the keyword and the words by using the learned keywords and the weights of the words. The weight of a keyword and words is calculated as the product of the term frequency and the inverse document frequency of the keyword. Here, the inverse document frequency is a description for removing words that are repeatedly used, such as surveys and articles, and words appearing in various documents are frequencies calculated so as to reduce weight.

The keyword vector calculator 33 calculates the size of the vector of each keyword and word using the weight of the keyword and the word. Then, the keyword vector calculator 33 calculates the vector size of the keyword by dividing the calculated weight of the keyword by the sum of the weights of all the words in the document.

In the same manner as described above, the keyword vector calculation unit 33 calculates the vector size for the keyword and the preselected words in the classification target document.

The distance calculation unit 35 calculates the distance DP of the keyword, which is the difference between the vector size of the keyword on the learning document and the vector size of the keyword on the classification target document, for the same keyword using the following equation (5). .

[Equation 5]

Here, α is the vector size of the keyword in the learning document, and β is the vector size of the keyword in the classification target document. When the keyword of the learning document and the vector size of the keyword in the classification target document are similar, the distance DP of the keyword is close to α. That is, the distance DP of the keyword has a higher value as the frequency of use of the keyword in the learning document and the document to be classified is similar. However, if the frequency used is too high or too low, the keyword distance is rather reduced. Therefore, it is possible to remove spammy documents and the like which are used frequently.

4 is a graph showing the relationship between the keyword distance and the frequency.

In this graph, if the frequency of the keyword is used is too high, the distance value is close to zero, which removes the spammy documents in which a certain word is excessively repeated. If the electronic document to be searched can be determined that there is no spam document, the relation graph is formed so that the distance value increases as the frequency of use of the keyword increases, so that the degree of conformity with the subject of the electronic document can be determined. .

When the distance of the keyword is calculated for each keyword included in the classification target document, the ranking calculation unit 37 multiplies the naive Bayesian probability value Nb and the distance DP for each keyword to calculate a similarity value for each keyword. Calculate. The similarity values for each keyword included in one classified document are added together or averaged, and this value becomes one ranking judgment value. Each classification target document has one ranking determination value, and the ranking calculation unit 37 compares the ranking determination value of each classification target document and calculates a ranking.

Such a ranking serves as a criterion for determining whether the learning document and the classification target document are similar, and through this ranking, it is possible to determine whether the classification target document is qualitatively matched with the subject matter. Accordingly, the high-ranking classification target document is classified as the subject of the corresponding learning document, and the low-ranking classification target document can be classified into another subject by measuring similarity with other learning documents. Therefore, the documents to be classified can be easily classified.

The subject generation process and document classification process of the search engine by such a configuration will be described with reference to FIG. 5 as follows.

First, the search unit 10 searches for an electronic document, and stores the keyword included in the electronic document in the keyword storage unit 21 (S505). The association degree calculation unit 23 calculates an association degree between keywords included in each electronic document (S510), and the convergence association degree calculation unit 25 associates a word set connected to an arbitrary keyword and a word set connected to another keyword. Comparing the degree of convergence to calculate the degree of convergence (S515). As a result of the calculation, a keyword whose convergence association degree is equal to or more than a predetermined level is selected as a theme (S520).

When the subject is selected, the naive Bayesian unit 31 calculates the naive Bayesian probability, which is a probability that keywords including the subject appear in the learning document (S525), and the keyword vector calculation unit 33 applies a weight to the learning document. And a vector size for each keyword included in the document to be classified (S530). For each keyword, the distance between the keywords is calculated using the vector size in the learning document and the vector size in the classification target document (S535).

Then, the similarity value is calculated by multiplying the naive Bayesian probability by the distance between the keywords (S540), and calculating one ranking determination value by processing the similarity values for each keyword (S545). Comparing the ranking judgment value of each classified document to calculate the ranking for the corresponding subject of each classified document (S550), and through the above-described process for each subject, each classified document can be classified according to the subject. (S555).

In this way, the present search engine enables the user to automatically select a subject from the electronic document using the degree of association and convergence of the keywords input by the user. Accordingly, it is possible to select atypical and various topics that could not be sorted by hand, so that users can access the information they want more easily and quickly.

In addition, by using a keyword, which is a selected topic, a ranking of a document to be classified as a criterion of similarity between the learning document and the document to be classified, can be classified according to the subject. Therefore, it is possible to determine whether the searched classified document is qualitatively suitable for the subject, thereby improving the reliability of the search and allowing the user to collect more accurate information.

As described above, according to the present invention, the informal and various subjects which could not be classified by manual operation can be selected, and thus, the user can more easily and quickly access the desired information. In addition, the classification target document can be classified according to the subject, and it is possible to qualitatively determine whether the searched electronic document is suitable for the subject, so that the user can collect more accurate information.

Further, in the detailed description of the present invention, specific embodiments have been described, which should be taken as exemplary, and various modifications may be made without departing from the technical spirit of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below, but also by the equivalents of the claims.

Claims (13)

Measuring a degree of association indicating the degree to which each of the keywords is simultaneously selected for at least one keyword included in the document to be classified: Measuring a degree of convergence association, which is an association degree between a word set associated with an arbitrary keyword and a word set associated with another keyword; Selecting the keyword as a subject when the degree of convergence association is equal to or greater than a predetermined schedule; and generating a subject and classifying a document by a search engine. The method of claim 1, The association is calculated by Equation 1, the subject generation and document classification method of a search engine, characterized in that.

Where LinkCount (w ₁ , w ₂ ) is the number of times the keywords w ₁ and w ₂ were tagged at the same time, WordCount (w ₁ ) is the total number of times the keyword w ₁ was used as a keyword, and WordCount (w ₂ ) was the keyword w ₂ The total number of times used as a keyword. The method of claim 2, The convergence correlation is calculated by Equation 2;

here,

and,

If both are greater than or equal to the preset minimum correlation, then a value is assigned to PositiveVal (w ₁ , w ₂ , w _i ), or a value is assigned to NegativeVal (w ₁ , w ₂ , w _i ). How to create topics and classify documents in search engines. The method of claim 3, wherein PositiveVal (w ₁ , w ₂ , w _i ) and NegativeVal (w ₁ , w ₂ , w _i ) are generated by the following equations (3) and (4), respectively. Document classification method.

The method of claim 1, The convergence association degree is closer to 1 as each keyword is similar, and closer to 0 as each keyword is different. The method of claim 1, Calculating a naive Bayesian probability by performing naive Bayesian learning on each keyword included in the classification target document and a learning document as a classification standard; and a method of generating a subject and classifying a document of a search engine further comprising: . The method of claim 6, And calculating a vector size of each keyword included in the learning document and the document to be classified. The method of claim 7, wherein Computing the vector size of each keyword, Calculating the weight of each keyword by multiplying the number of occurrences of each keyword in the document by the inverse document frequency; And calculating the weight of each keyword by dividing the weight of each keyword by the sum of the weights of all words in the learning document or classification target document including the respective keywords. The method of claim 8, Calculating a distance of a keyword that is a difference between a vector size of each keyword on the learning document and a vector size of each keyword on the classification target document; and generating a subject of a search engine and a document classification method. . The method of claim 9, The distance between the keyword and the learning document, the higher the frequency of the keyword is used in the classification target document has a higher value, characterized in that the topic generation and document classification method of the search engine. The method of claim 10, Calculating a similarity value for each keyword by multiplying a naive Bayesian probability value Nb and a distance DP for each keyword; Calculating one ranking determination value through a predetermined process of the similarity value for each keyword included in the classification target document; Calculating a ranking of each classification target document by using the ranking determination value of each classification target document, and classifying the classification target document according to a subject; generating a subject and a document of a search engine further comprising Classification method. Calculating naive Bayesian probabilities by performing naïve Bayesian learning on the learning document that is a classification standard and each keyword included in the classification target document; Calculating a vector size of each keyword included in the learning document and the classification target document; Calculating a distance of a keyword that is a difference between a vector size of each keyword on the learning document and a vector size of each keyword on the classification target document; Calculating a similarity value for each keyword of the document to be classified by multiplying the naive Bayesian probability value Nb and the distance DP for each keyword; And, And calculating a ranking of each classification target document using the similarity value. A computer-readable recording medium containing a program capable of performing the method of any one of claims 1 to 12.

Images