KR100818553B1 - Document ranking granting method and computer readable record medium thereof - Google Patents

Abstract

The present invention relates to a method for assigning a document ranking. When a user inputs a search query through a search box of a search engine, the user action ranking of the document is considered in the weight of the document corresponding to the query, By applying a creation time, a lot of actions are performed from other users interested in the query, and if the user of the document corresponding to the subject is a document ranking grant method to be ranked higher.

The present invention has an effect of exposing a document and a user corresponding to a specific query, and showing superior search performance compared to a conventional simple keyword-based search ranking method using user feedback information.

Document, author, action, ranking, rank

Description

A computer-readable recording medium containing a document ranking method and a program capable of performing the same. {DOCUMENT RANKING GRANTING METHOD AND COMPUTER READABLE RECORD MEDIUM THEREOF}

1 is a structural diagram of a ranking calculation method applied to the present invention

Figure 2 is a table showing the type and use of the action applied to the present invention

3 shows time weights according to the document generation period (d _c -d _w ).

Figure 4 is a diagram of the application method when searching used in the present invention

5 is a search result screen example to which the present invention is applied

Figure 6 is an example of expert results screen to which the present invention is applied

The present invention relates to a method for assigning a document ranking. When a user inputs a search query through a search box of a search engine, the user action ranking of the document is considered in the weight of the document corresponding to the query, and the creation time of the document is determined. By applying, a lot of actions are performed from other users interested in the query, and the documents and users matching the specific query are exposed by allowing the user of the document to be ranked higher when the subject is a guru. By using the feedback information of the present invention, compared to the existing simple keyword-based search ranking method, the present invention relates to a document ranking granting method having an excellent search performance.

When users search for information using a search engine, ranking is the most important of the various quality measures of the search engine. In this case, the ranking is to rank appropriate documents based on specific keywords, and it is possible to reduce the effort of acquiring information to the user by placing the appropriate documents above.

Currently, domestic portals use the appearance location information or the frequency of appearance in the search result document based on the searched query. However, this alone is difficult to find an efficient search that can easily find the desired document in a large document.

In the case of foreign countries, in addition to word-based appearance information, various weight determination methods using various meta information on HTML are used, starting with page rank using a linked structure between documents. Meta search engines also use a method of selecting a search system based on a query. However, compared to such a large cost, the improvement of performance is limited, it is verified for web pages, and cannot be applied to various contents provided by a domestic portal at once.

In the literature and articles published recently, studies were conducted to assign appropriate weights to various factors that were significant for search ranking. Page ranks have the best impact on Web documents, and it is proven that each domain has strong factors.

However, these methods are not suitable for UCC (user-generated content) documents, which are the main documents currently being serviced in Korea, and they are difficult to apply due to the high cost.

The present invention is to solve the above problems of the prior art, the main element for ranking the content of the domestic portal is subject fitness and user action information. In order to serve documents and gurus that have a subject that matches a particular query, an automatic classification method is required. Autoclassification is a method of defining a list of subjects, analyzing a document, and matching it to topics that represent the document. It is also an object of the present invention to provide a method of calculating a ranking value in association with a previous subject classification suitability so as to locate a document preferred by a user using a user's action information.

According to the present invention, a document having a subject that matches or is similar to a user's input keyword and a user's various feedback information (hereinafter referred to as an action) may be ranked higher than the document that matches the subject or the user prefers. Action Ranking is called). In addition, we want to maximize the efficiency of the search by exposing the list of guru (expert) suitable for the entered keyword.

Document ranking method according to the present invention,

A document creation step of creating a document by a user; A query input step of inputting a query by a searcher; A document weight extraction step of extracting a document weight value considering a sum of plus influence, redundancy impact, and click impact force up to a search date for the document to a probability that the document created in the document creation step matches the query; The sum of document weights made by others of the documents related to the query among the documents created by the user, the sum of plus weights of other people related to the query, the sum of duplicate weights, and the sum of click weights. A user ranking extraction step of extracting a user ranking from the user; And a document ranking extraction step of extracting a document ranking from the document weight value and user ranking.

에 의해 계산되는 것이 바람직하다.The document ranking

It is preferable to calculate by.

here,

W (s, d _j ) is the document weight

AR _A (s, u _i ) is a user ranking.

에 의해 계산되는 것이 바람직하다.The document weight is

It is preferable to calculate by.

here,

에 의해 계산되는 것이 바람직하다.The user ranking is

It is preferable to calculate by.

here,

Document weight

Plus weight

Redundancy Weight

Click weights

One or both of the user ranking and the document ranking is preferably time-weighted so that the ranking decreases as time passes.

으로 정의되며, 상기 적용은 상기 사용자랭킹, 문서랭킹에 상기 시간가중치가 곱해진 것을 의미하는 것이 바람직하다.The time weighted value

Preferably, the application means that the time weighted value is multiplied by the user ranking and the document ranking.

here,

d _c is the time of action ranking calculation

d _w is the document generation time.

It is preferable to calculate a final document ranking by applying a system rank to the document ranking.

에 의해 계산되는 것이 바람직하다.The final document ranking

It is preferable to calculate by.

here,

: 질의 q에 대한 최종 문서랭킹

: Final document ranking for query q

s _j : j subject value

d _i : i-th document

q: query

ave: mean value

: 주제벡터 s_j와 쿼리벡터 q간의 코사인 값

: Cosine of subject vector s _j with query vector q

: j번째 주제에 속한 문서 i번째 문서의 Atype 액션랭크 점수

: Atype Action Rank score of the i th document in the j subject

α, β: Action Rank Score and SR Weight

SR: System rank value (ranking score of search engine)

For a document whose user is unknown, it is preferable to have a document ranking extraction step of extracting a document ranking of a document without a user, which extracts a document ranking in consideration of the influence of plus counts, click counts, and system rank values for the document. Do.

Document ranking of documents without the user

에 의해 계산되는 것이 바람직하다.

It is preferable to calculate by.

here,

It is preferable that a time weighting value is applied to the document ranking of the document without the user so that the ranking decreases as time passes.

로 정의되며, 상기 적용은 상기 사용자가 없는 문서의 문서랭킹에 상기 시간가중치가 곱해진 것을 의미하는 것이 바람직하다.The time weighted value

The application means that the document weight of the document without the user is multiplied by the time weight.

here,

d _c is the time of action ranking calculation

d _w is the document generation time.

Further, the document ranking granting method according to the present invention,

Calculating a weight of the first document according to an action of a user performed on the first document; Calculating a first goodness of fit with the first keyword and the weight of the calculated first document; When a query is generated by the first keyword, in the second document having the first keyword and a second goodness of fit of a predetermined size or more, the first document and the first document according to the size of the first goodness and the second goodness of fit. 2 Includes steps to sort documents.

The weight of the first document is preferably proportional to the ranking of the user who created the first document.

The weight of the first document is preferably proportional to the number of times the user who read the first document bookmarks the first document or the number of replies added.

The weight of the first document is preferably proportional to the number of users who read the first document.

When the first document and the second document contain the same content, and the original author of the same content is the first document, the suitability of the first document with the first keyword is greater than that of the second document. It is preferable to further include setting the height.

The weight of the first document is preferably proportional to the temporal order of actions of the user occurring on the first document.

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, with reference to the drawings will be described in detail. However, these drawings are for illustrative purposes only and the present invention is not limited thereto.

1 is largely composed of a database storing information of a document to be searched and feedback information of a user, an automatic classifier performing automatic classification, and a ranking calculation module that calculates a ranking value using classification information and user action information.

An action is any implicit or explicit activity during the information search process. These actions can be divided into two categories. The first is the Creation action as the information provider, and the second is the Delivery action as the information explorer. In the present invention, both of them will be weighted, but the actual action weight is given more significance to the Creation action than the Delivery action.

2 shows a table of the types of actions and the actions to be adopted.

The Write action is the act of writing a document. In the container, in the paper, as in the Cyworld bulletin board, it is the most important of all actions, and the Write action value, which is the value that the document is classified as a specific subject, is readjusted.

A wiki is a service that will be opened like the Search Plus service, and is a weight for the article I wrote in the wiki. Since wikis have a limited number of users who can write, it seems problematic to give them the same weight as the Write action. This violates the condition that all users must compete under the same conditions. It makes guru more guru, and the general user can reduce the chance of entering the guru, and there is a possibility that the vacancy bin side benefit phenomenon will occur here. Also, a wiki can be written by several people on the same document, so it is not easy to reflect this in action ranking. Therefore, the wiki is limited to the use of encyclopedias and proceeds separately from action rankings.

The RSS action is a score that absorbs the influence index of other users who read my archive periodically through RSS as my influence. However, RSS can be reflected only when registering only documents of a specific subject, but registering RSS as a whole cannot be mapped to a specific subject, so it is excluded from action ranking.

Plus action is a concept of bookmarking a good document found during a search or web surfing. The document you add is reflected in the original owner's action score, and the content of the original itself is also weakly reflected in my action score. Plus actions also include evaluation of the document and affect the ranking by good or bad.

The Comment action reflects the number of comments made by others about a particular document in the document owner's action ranking. Because the command system may not be built according to the type of service for which the document is generated, it is difficult to give the command value the same in all documents. In addition, it is difficult for the system to judge whether the command itself is positive or negative, so it is dangerous to directly affect the ranking, which is not cost-effective.

The Query action is the weight for the keyword that the user used as a search query. This action is useful for understanding the tendency that a user is interested in a certain topic, but it is not useful in terms of information power because a lot of queries do not have much information power.

The Click action refers to someone else's documents that the user clicked on in search results. My influence is reflected in the clicked document, and the content of the document is slightly reflected in my disposition and action score.

Duplicate action is for A type document (the best netizen information), and it is to give the original author a higher score by masking scrap or partial paragraph theft, and the higher the duplicated document, the higher the Write action value. This can be projected as a Plus action during the actual operation. For example, if user A scraps a document of user B, user A calculates that document B is Plus.

 The method of calculating the action ranking should perform automatic classification of the user's document (or actions) by the subject in order to calculate the thematic ranking. The classified documents then reflect the influence of the other person taking the action.

The calculation procedure is as follows.

-Action classification

-Reflects the influence of others on the probabilities of the classified actions

-Score the actions that reflect the influence

The basis of the action is the document and keywords. Plus, even clicks are the subject, so automatic classification of actions is the primary source. Autoclassified actions are expressed as probability values of goodness of fit for a subject. Along with this probability, we calculate the influence of the actions that others have done on this document.

The influence of the user is divided into subjects, and the influence of the user on a particular subject is primarily calculated as the sum of the document actions contained in the subject. Here, the sum of actions is calculated by calculating the probability of matching all the actions taken by the user through the automatic classification, and correcting these probability values by the action weight to determine the influence of the user.

Action Rank's algorithm is the sum of the numerical results for the aforementioned types of actions.

The following is the weight calculation method for each action.

(1) weights for generated documents

Influence on the user-created document is calculated as the sum of the weight of the user's documents included in a particular subject and the influence of others.

When the weight of the j th document for a particular subject s is W (s, d _j )

The weight of a document is basically influenced by the subject of the classified document through automatic classification. Where P (s, d _j ) is the probability value of the subject-specific document generated through the automatic classification. The weight of the document defaults to a probability value that will match the subject to which it belongs, and the probability value is corrected through the influence of the actions that other users have made on this document.

There are three main influences of other users: plus, redundancy and click. The calculation method takes into account the system resources and time, and calculates the sum of the influence of each action by the previous day (d-1) and the influence of others of the day (d) once a day. Finally, the weights for each action used in this document are used to reflect the regression coefficients evaluated through the simulation.

(2) I did plus weight

When a plus action occurs, a portion of my influence is given to the document author when calculating the document weights. And, the weight of the document is given to my influence on the subject to which the plus document belongs. The plus weight is determined by the following formula.

When the plus weight for the first document belonging to a specific subject s is W (s, p _| )

The influence of a topical document is reflected by multiplying the probability of the document belonging to the topic by the influence of the document owner.

Multiple people can be added to a document, assuming that the positive influence is equally distributed.

(3) the redundancy weights I did

Redundancy actions are calculated in the same way as plus. Redundancy weight is determined by the following equation.

When the redundancy weight for the oth document belonging to a specific topic s is W (s, r _o )

(4) the weight I clicked

Clicks also count in the same way as plus. However, the importance is expected to be superior to the plus. Clicks should apply deduplication to prevent clicks against malicious behavior (suggesting one user only one click per document per day).

Let's say W (s, c _m ) is the click weight for the mth document in a particular topic s.

The final action ranking is formulated by dividing into two types according to the existence of the author of the document by combining the three weights described above.

(1) A type that can identify the author of a document (hereinafter referred to as A type)

In Type A, Action Ranking is divided into User Ranking and Document Ranking.

The user's action ranking uses a combination of weights, plus one's plus weight, and one's click weights in the generated document. User action ranking according to A type is as follows.

)되고, 그 자신 또한, 그 주제에 일치하는 문서들에 대해서 많은 액션을 수행(

)하였을 때 순위가 올라가게 됨을 의미한다.The above equation shows that the ranking of interested users (u _j ) on a particular subject (s) is a lot of action from different users for each document written by that user.

) And itself also perform a number of actions on documents that match the subject (

) Means that the rank goes up.

Action ranking of the document is used by merging both the weighting and the user's action ranking and system ranking in the generated document. The document action ranking according to A type is as follows.

To the weight of the expression of these documents to produce document _{(W (s, d j)} ) by multiplying the user action score of the document _{(AR A (s, u j} )), the d _j documents on a specific topic (s) This is to ensure that many actions are taken by other interested users, and ranked higher if the user of the document is a guru on the subject.

(2) Type unknown to the user (hereinafter referred to as type B)

Documents of type B have no subject information, so simply use the number of action qualities considered.

(3) reflect the date

The date of creation of the document is digitized from the time point to the date of calculating the current action ranking and multiplied by the action ranking value, leading to higher ranking of the latest articles. When AR _A (s, d _j ) and AR _B (s, d _j ) are multiplied by 0.998 ^ (d _c -d _w ), a graph like FIG. However, the minimum time weight is 0.5 and multiplies 0.5 by one year to reflect timeliness. Where d _c is the time of action ranking calculation and d _w is the document generation time.

The action rank for each type considering the final time weight is as follows.

Application method is as follows.

Search is, of course, divided into A type and B type, and performs the search by the system configuration of FIG. 4 below. B type is the existing search method, but it can be ranked by B type action ranking, but A type search differs from the subject search server by searching the subject information that best matches the query and combining it with the existing ranking. The description below describes the documentation for Type A and how to apply user search.

(1) document search

In the same context as traditional search, the internal search method extracts results by combining keyword matching and topic matching.

Search algorithm

1) Mapping and finding multiple topics for different keywords (max 3).

-> Use cosine coefficients between query vector and subject vector

2) Extract documents containing query vectors into result set

-> Using Query-Document Keyword Matching

3) Calculate final ranking reflecting action ranking

here,

: 질의 q에 대한 최종 문서랭킹

: Final document ranking for query q

s _j : j subject value

d _i : i-th document

q: query

ave: mean value

: 주제벡터 s_j와 쿼리벡터 q간의 코사인 값

: Cosine of subject vector s _j with query vector q

: j번째 주제에 속한 문서 i번째 문서의 Atype 액션랭크 점수

: Atype Action Rank score of the i th document in the j subject

α, β: Action Rank Score and SR Weight

SR: System rank value (ranking score of search engine)

By policy, you can control the weight of the actions that take into consideration the subject relation with the query rather than the system rank.

∴ α will refer to the class and the coefficient β is used in the system coefficient.

A search query can be a complex query, and depending on the query or subject classification, the search results can be the result of a set of topics that span multiple topics, rather than a single topic. Of these results, we extract only up to three subjects with high cosine coefficients. This is to prevent the search reaction speed in the engine from being lowered. Finally, the documents resulting from keyword matching are searched by ranking them by action rank.

(2) people search

Similar to document search, but extracting only those within the subject.

Search algorithm

1) Mapping multiple topics for different keywords.

-> Use cosine coefficients between query vector and subject vector

2) Extract the users belonging to the mapped subject as a full result set

3) Calculate final ranking by reflecting action ranking according to query

5 and 6 illustrate a list of documents and users served by the above algorithm.

As described above, according to the present invention, the document and the user corresponding to the specific query are exposed, and the user's feedback information is used to show superior search performance compared to the existing simple keyword-based search ranking method.

Claims (19)

In the document ranking grant method, A document creation step of creating a document by a user; A query input step of inputting a query by a searcher; A document weight extraction step of extracting a document weight value considering a sum of plus influence, redundancy impact, and click impact force up to a search date for the document to a probability that the document created in the document creation step matches the query; The sum of document weights made by others of documents related to the query among the documents created by the user, the sum of plus weights of other people related to the query, the sum of duplicate weights, and the sum of click weights. A user ranking extraction step of extracting a user ranking from the user; And a document ranking extracting step of extracting a document ranking from the document weight and user ranking. The method of claim 1, The document ranking AR _A (s, d _j ) is calculated by the following formula.

here, W (s, d _j ) is the document weight AR _A (s, u _i ) is a user ranking. The method of claim 2, And the document weight value W (s, d _j ) is calculated by the following equation.

here,

The method of claim 2, And the user ranking AR _A (s, u _i ) is calculated by the following equation.

here, Document weight

Plus weight

Redundancy Weight

Click weights

The method according to any one of claims 1 to 4, One or both of the user ranking, the document ranking is applied to the document weighting method, characterized in that the ranking is reduced as time passes. The method of claim 5, The time weighting value is defined by the following equation, and the application is a document ranking granting method, characterized in that the user weighting, document ranking means that the time weighting value is multiplied.

here, d _c is the time of action ranking calculation d _w is the document generation time. The method of claim 5, wherein the final document ranking is calculated by applying the system ranking to the document ranking according to claim 5. The method of claim 7, wherein Final document ranking

The document ranking grant method, characterized in that calculated by the following equation.

here,

: Final document ranking for query q s _j : j subject value d _i : i-th document q: query ave: mean value

: Cosine of subject vector s _j with query vector q

: Atype Action Rank score of the i th document in the j subject α, β: Action Rank Score and SR Weight SR: System rank value (ranking score of search engine) The method of claim 8, For a document whose user is unknown, a document ranking extraction step of extracting a document ranking of a document without a user, which extracts a document ranking in consideration of the influence of plus counts, click counts, and system rank values for the document, is performed. Grant method. The method of claim 9, The document ranking AR _B (s, d _j ) of the document without the user is calculated by the following equation.

The method of claim 10, The document ranking granting method of claim 1, wherein a document weighting is applied to a document ranking, wherein a time weight is applied and the ranking decreases as time passes. The method of claim 11, The time weighting value is defined by the following equation, and the application is a document ranking granting method, characterized in that the time weighting value is multiplied by the document ranking of the document without the user.

here, d _c is the time of action ranking calculation d _w is the document generation time. Calculating a weight of the first document according to a user action performed on a first document, wherein the weight of the first document is proportional to the temporal order of the user's actions occurring on the first document; Calculating a first goodness of fit with the first keyword and the weight of the calculated first document; When the query is generated by the first keyword, in the second document having the first keyword and the second goodness of fit, the first document and the second document are determined according to the size of the first goodness and the second goodness of fit. And aligning the document ranking. The method of claim 13, And the weight of the first document is proportional to the ranking of the user who created the first document. The method of claim 13, And the weight of the first document is proportional to the number of times the user who read the first document bookmarks the first document or the number of replies added. The method of claim 13, And the weight of the first document is proportional to the number of users who read the first document. The method of claim 13, When the first document and the second document contain the same content, and the original author of the same content is the first document, the suitability of the first document with the first keyword is greater than that of the second document. The document ranking grant method further comprising the step of setting a high. delete A computer-readable recording medium containing a program capable of performing the method of any one of claims 1 to 17.

Images