JP2010134714A - Collaborative sorting apparatus, method, and program, and computer readable recording medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To achieve division of a set of collected tags into two groups of objective tags consensual among users and subjective tags being noise as category axes. <P>SOLUTION: Tags being sorting axes are made into feature vectors by probabilistic clustering with respect to a plurality of sorting axes stored in a database, sorting objects, and concurrence information relevant to users, and sorting axes arbitrarily given by users are divided into objective tags and subjective tags on the basis of feature vectors. Entropy values of feature vectors are calculated, and feature vectors of which the entropy values are higher than a prescribed threshold are defined as subjective tags, and feature vectors of which the entropy values are lower than the threshold are defined as objective tags. Alternatively, a set of sorting axes being a correct answer is obtained from the database, and nearby tags being at short distances from the set of sorting axes as the correct answer are defined as subjective tag candidates. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a cooperative classification method and apparatus, a program, and a computer-readable recording medium, and in particular, a cooperative classification method in which a plurality of users classify and share information such as bookmarks, photos, videos, books, and papers. The present invention relates to an apparatus, a program, and a computer-readable recording medium. Specifically, in the collaborative classification information device, an objective classification axis (objective tag) suitable as an information classification axis and a subjective classification axis (subjective tag) not suitable as an information classification axis are provided on the service side. The present invention relates to a collaborative classification method and apparatus, a program, and a computer-readable recording medium for mechanically distinguishing only from the classification information collected in the system and realizing high-level classification information presentation by a service provider.

  Recently, collaborative / social tagging (collaborative / social tagging) enables users to organize fresh information by sharing the results of each user organizing and classifying information such as URLs (bookmarks), photos, videos, and papers. System) is prosperous. In the cooperative distributed system, a user can freely assign tags, and information classification results by a user can be collected and used as a search key when acquiring information.

  However, in the collaborative classification system, since the user freely attaches tags, there is a problem that subjective tags that are not suitable as classification axes are mixed (for example, Reference 1 “The Structure of Collaborative Tagging System”, Scott Golder). , and Bernardo A. Huberman Journal of Information Science 2006). In Document 1, it is said that there are seven types of tag usage in a cooperative classification system.

● Objective tags:
1. What is the subject about? (What it is about?)
--- [Example] "web2.0", "Computer", "programming"
2. What is the target? (What it is?)
-[Example] "blog", "news", "article"
3. Who owns the subject? (Who owns it?)
-[Example] "jkondo", "dankogai", "timbray"
4). Refinement Categories
-[Example] "Computer" + "Programming", "Work" + "Delivery"
● Subjective tags:
5). Qualities or characteristics
―― [Example] “This is amazing”, “fun”, “This is terrible”, “★★★”
6). For your own reference (Self Reference)
-[Example] "Selkuma" (self-bookmark), "mystuff"
7). Task Organizing
-[Example] "Read later", "toread", "todo"
The type of these tags cannot be determined only by the face of the tag. For example, if the URL with the tag “Blog” is an article about “Blog” or the blog itself, it is not possible to understand the contents of the URL. However, among these seven types, it is considered that the types of tags differ between 1-4 and 5-7. The objective tags 1 to 4 are objective tags related to the contents of the object, and there is a possibility that an agreement can be reached among users, but the subjective tags 5 to 7 are not only for the object but also for the user. It is a subjective tag that is greatly influenced by the state, and is considered to be used differently depending on the user. In this specification, a tag considered to belong to 1 to 4 is called an “objective tag”, and a tag considered to belong to 5 to 7 is called a “subjective tag”.

<Problems of subjective and objective tags>
The problem of subjective tags and objective tags will be described below.

[Examples of subjective and objective tags]
In August 2008, examples of objective tags and subjective tags from tags used from a social bookmark service published on the Web are shown below.

● Objective tags:
⇒ "mobile", "science", "apple", "osx", "turorials", "technology", "life", "gtd"
● Subjective tags:
⇒ "fun", "cool", "this is terrible", "this is awesome"
One of the reasons why the current cooperative distributed system has gained popularity is that arbitrary keywords can be assigned as tags (classification axes). As a result, in a cooperative distributed system, objective tags and subjective tags are mixed and collected. Since the objective tag can be agreed between users, it is considered useful for searching for information or displaying a list of tags. On the other hand, since subjective tags are used differently by each person, it is useful when each person uses it personally, but it is difficult for others to use it as it is.

  For example, the reason for saying "This is amazing" varies greatly from person to person. There are many cases where a URL that has been classified many times as "This is amazing" is tagged as "This is terrible" by different users.

  However, if the service provider is prohibited from using an arbitrary tag and the user is forced to use only a predesignated tag, the advantage of the cooperative classification system is greatly impaired. If the service provider can automatically distinguish the tag type and present it effectively when reusing it, allowing the user to use any tag, the advantages of the existing collaborative classification system interface can be obtained. It can be expected that the convenience and usefulness of the service will increase without loss, and as a result, there is a possibility of gathering more users.

  For this reason, a cooperative classification system is desired to distinguish between an objective tag and a subjective tag.

As a collaborative classification system, a concept retrieval technology based on a probabilistic model PLSI (Probabilistic Latent Semantic Indexing) (see, for example, Non-Patent Document 1) or a technology in which PLSI is expanded to three sets (for example, a technology for applying social bookmarking services) Non-Patent Document 2).
"Probabilistic Latent Semantic Analysis", Thomas Hofmann, 1999, In Proc. Of Uncertainty in Artficial Intelligence, UAI'99 "Exploring Social Annotations for the Semantic Web", Xian Wu et.al, 2006, In WWW2006.

  However, the collaborative classification system itself that accumulates classification results by end users is a technology that can be realized only by the appearance of the Web, and the present invention can be solved only when data is actually accumulated and used. Challenges to be clarified have been clarified. Although the technique of the above-mentioned literature 1 etc. paid attention to the kind of tag, it does not touch on the problem of distinction between the subjective tag and the objective tag that are considered to be highly useful as a service although the level of abstraction is higher.

  As an easy method, it is conceivable to construct a dictionary manually and distinguish tag types on a dictionary basis. However, because end users continue to classify every hour in a collaborative classification system, new tags and tags with unclear meaning (emoticons, pictograms, etc.) are really subjective tags in dictionary-based methods. It is difficult to determine whether or not. Therefore, a mechanical distinction method or a technique that supports manual distinction is desired.

  As described above, in the collaborative classification system and collaborative classification system in which multiple users classify and share information such as bookmarks, photos, videos, books, and papers, problems and results of mixed objective tags and subjective tags As a result, it is difficult to reuse the classification results by a third party.

  The present invention has been made in view of the above points, and distinguishes the collected tag set into an objective tag agreed upon by users and a subjective tag that causes noise as a classification axis. Or a collaborative classification method, apparatus, program, and computer-readable record that can support manual discrimination and facilitate the use of a classification axis (tag) of a service provider The purpose is to provide a medium.

  FIG. 1 is a diagram for explaining the principle of the present invention.

The present invention (Claim 1) has a user interface for inputting / outputting data, a communication means for performing data communication, and a data processing means for executing classification processing as classification target browsing processing and classification target identifier registration processing. A collaborative classification method in a system in which users classify and share information such as bookmarks, photos, videos, books, and articles,
A feature vectorizing step (step) in which the data processing means performs feature vectorization of the tag that is the classification axis by probabilistic clustering with respect to the co-occurrence information related to the plurality of classification axes, classification targets, and users stored in the database. 1, 2),
Based on the feature vector obtained in the feature vectorization step, an analysis step (step 3) is performed for discriminating the classification axis arbitrarily given by the user into an objective tag and a subjective tag.

Further, the present invention (Claim 2) is provided in the analysis step (Step 3).
An entropy calculation step of calculating an entropy value of the feature vector obtained by the feature vectorization;
A tag having an entropy value higher than a predetermined threshold is defined as a subjective tag, and a tag having an entropy value lower than the threshold is defined as an objective tag.

In the analysis step (step 3), the present invention (claim 3)
Get the correct classification axis set from the database,
Neighboring tags that are close to the correct classification axis set are set as subjective tag candidates.

Further, the present invention (Claim 4) is provided in the analysis step (Step 3).
An entropy calculation step of calculating an entropy value of the feature vector obtained by the feature vectorization step;
A subject whose entropy value is higher than a predetermined threshold is a correct subjective tag,
Neighboring tags that are close to the correct subjective tag are determined as subjective tag candidates.

  FIG. 2 is a diagram for explaining the principle of the present invention.

The present invention (Claim 5) has a user interface for inputting / outputting data, a communication means for performing data communication, and a data processing means for executing classification processing as classification target browsing processing and classification target identifier registration processing. A collaborative classification device in a system in which users classify and share information such as bookmarks, photos, videos, books, and papers,
Data processing means
A feature vectorizing means 110 for converting the elements extracted from all or any of the information related to a plurality of classification axes, classification targets, and users stored in the database 103 by probabilistic clustering;
Based on the feature vector obtained by the feature vectorization means 110, there is an analysis means 120 that distinguishes a classification axis arbitrarily given by the user into an objective tag and a subjective tag.

Further, the present invention (Claim 6) is provided in the analyzing means 120.
Entropy calculating means for calculating the entropy value of the feature vector obtained by the feature vector converting means 110;
A unit having a higher entropy value than a predetermined threshold as a subjective tag, and a unit having an entropy value lower than the threshold as an objective tag.

Further, the present invention (Claim 7) is provided in the analyzing means 120.
A means for acquiring a correct classification axis set from the database and using a nearby tag close to the correct classification axis set as a subjective tag candidate is included.

Further, the present invention (Claim 8) is provided in the analyzing means 120.
Entropy calculating means for calculating the entropy value of the feature vector obtained by the feature vectorization;
Means that the entropy value is higher than a predetermined threshold value as a correct subjective tag,
And a means for setting a subjective tag as a subjective tag candidate that is close to the subjective tag that is the correct answer.

  The present invention (Claim 9) is a cooperative classification program for causing a computer to function as each means constituting the cooperative classification apparatus according to any one of Claims 5 to 8.

  The present invention (Claim 10) is a computer-readable recording medium storing the cooperative classification program according to Claim 9.

  As described above, according to the present invention, the tag set collected in the collaborative classification system is divided into the objective tag agreed by the users and the subjective tag that causes noise as the classification axis. It is possible to facilitate the utilization of the classification axis (tag) of the service provider by realizing this or by assisting the manual discrimination. This is because subjective tags are highly useful only for the user himself and less useful when viewed from others, but objective tags can be said to be useful by others. By classifying mechanically into two types of tags, a service provider can provide advanced classification information.

  First, terms used in this specification will be described.

・ Collaborative / Social Tagging System
A system that allows end users to freely classify information and share the classification results, such as social bookmarking services (del.iciolus, Hatena, etc.), photos (flickr), videos, papers (citeulike), etc. is there.

・ SBM (Social Bookmark Service)
Abbreviation for social bookmarking service. The classification target is a resource on the Web, that is, a URL. This is a system for sharing bookmarks of users through a network, and is a typical example of a collaborative classification system.

Classification axis Information used for classification in the cooperative classification system, and tags (keywords), images, and sounds are used as the classification axis. As shown in FIG. 3, the classification target is classified by the classification axis (tag).

・ Classifier identifiers To be classified, the classification targets must be distinguishable. The distinction between the classification targets is synonymous with the discrimination of the classification targets by identifiers because they can have a one-to-one correspondence with the identifier set that can be uniquely determined. In the case of a bookmark, the URL can be used as an identifier. That is, in this specification, there is no particular problem even if the classification target is expressed by an identifier.

・ Data collected by SBM In this specification, as shown in FIG. 4, three sets of U × R × T in which (USER), which URL (RESOURCE), and what category (TAG) are classified are assumed. is doing. In the actual SBM, comments, user scores, time stamps, and the like are also collected.

Tag A classification axis in the social bookmark service, and each user can determine an arbitrary word.

-Objective tags-Subjective tags The types of tags are divided into two types in this specification. This type is roughly divided into two types according to Golder et al. In Document 1 from the viewpoint of whether the tag is subjective or objective.

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

  FIG. 5 is a schematic block diagram of a system according to an embodiment of the present invention.

  The cooperative classification system shown in the figure is provided as a network service, and the user can use the service through a web browser or a client application.

  The collaborative classification processing system includes an application server 100 that actually performs processing, a database server 200 that stores data, a classification axis and classification target, a feature vectorization engine 300 that vectorizes user information, and calculation of entropy values of feature vectors. , Sorting, k vector acquisition engine 400, LAN 500, and router 600.

  These processing units are not only realized in a single server, but may be distributed in a plurality of units.

  Below, especially a social bookmark system is taken as an example among cooperative classification systems, and it demonstrates as what is performed with the application server 100 as a cooperative classification apparatus.

  FIG. 7 shows the configuration of the application server in one embodiment of the present invention.

  The application server 100 is connected to a plurality of user terminals 2 via a network. The user terminal 2 includes a communication unit 21 that communicates with the application server 100 via a network, a storage unit 22, a data processing unit 23, and a user interface 24 having an input / display function.

  The application server 100 is composed of a communication interface 101 for communicating with the user terminal 2 via a network, a control unit 102 for performing cooperative classification processing, and a storage unit 103 for storing calculated values and threshold values during progress. The

  The control unit 102 includes a feature vectorization unit 110 and a vector analysis unit 120.

  FIG. 6 shows an outline of processing of the control unit 102. The feature vectorization unit 110 has the function of the feature vectorization engine of FIG. 5, and performs element extraction processing for extracting co-occurrence data from the triple database, pre-filter processing for performing noise processing on the co-occurrence data, and noise is removed PLSI processing is performed on the co-occurrence data.

  The feature vectorization unit 110 uses all or any information of the classification target, tag (classification axis), and user input from the database server 200 via the input device, and extracts the characteristic features from them, for example, , Tag-tag, classification target-tag, and the distance between items can be measured. For this purpose, various existing techniques can be used. For example, techniques such as Document 1, Non-Patent Document 1, and Non-Patent Document 2 described above can be used. Typically, each item is expressed as a feature vector, and the distance between elements is measured using the similarity between the feature vectors.

  The feature vectorization unit 110 includes an element extraction unit 111, a noise removal unit 112, and a co-occurrence data indexing unit 113.

  Details of the processing of the feature vectorization unit 110 will be described below.

FIG. 8 is a flowchart of the operation of the feature vectorization unit in one embodiment of the present invention. Step 101) When raw text data (or database) is input to the element extraction unit 111, necessary elements are extracted from the raw data. Extract. The structure of raw data in social bookmarks depends on how the database schema is defined within each service. As a specific example, the act of bookmarking may be modeled as “who (U), what URL (R), when (t), what (T), what (T) and some impressions (C)”. At this time, the raw data can be regarded as five sets of co-occurrence data (a subset of U × R × t × T × C) consisting of the direct product of the respective elements. Reference 1 employs a probabilistic indexing method composed of two sets, and selects two sets of co-occurrence data from five sets of co-occurrence data so that the indexing method can be executed (U × R or U × T). Such). In addition, Non-Patent Document 1, Non-Patent Document 2 and Document 2 ("Visualization of SBM user and tag relations using PLSI", Takashi Mao, Yasuharu Eda, Masashi Yamamuro, 7th Web Intelligence and Mark In the Raction Study Group, 2006.), since the indexing can be performed on the three sets of co-occurrence data, the three sets of co-occurrence data such as (U × R × T) are selected. Information that is not selected in the process is simply ignored.

  In the following, the PLSI which is the indexing technique of Document 1 is expanded from 2 sets of co-occurrence data to 3 sets, which is the technique of Non-Patent Document 1 and the discussion holds in the same way, so the following uses 3 sets of co-occurrence data. I will explain.

  Step 102) In PLSI, it is known that items with low registration frequency in co-occurrence data become noise (Non-patent Document 1). Therefore, the noise removal unit 112 removes noise from the co-occurrence data using the minimum frequency condition of the elements in the co-occurrence data. The frequency of each item will be described.

Now, a specific example of 3 sets of data
{(U1, r1, t1), (u1, r1, t2), (u2, r2, t2), (u3, r2, t3)}
And In this case, each item frequency represents the number of times an item appears in the data set as three sets. That means
│u1│ = 3, │u2│ = 1, │u3│ = 1
│r1│ = 2, │r2│ = 3
│t1│ = 2, │t2│ = 2, │t3│ = 1
It becomes. However, | x | represents the density (number) of the item x. In this case, if 2 is specified as the lowest frequency, u2, u3, and t3 are removed as noise, and the three sets of data are {(u1, r1, t1), (u1, r1, t2)}.
Remains.

  Step 103) Using the PLSI described in Document 1 and Non-Patent Document 1, the co-occurrence data indexing unit 113 learns the co-occurrence of N sets of co-occurrence data and sets the items of the respective items as probabilities (distributions). ) Calculate as a vector. The probability vector is a vector normalized so that the sum of the values of a certain vector becomes 1, and can be regarded as a probability distribution. By using KL divergence or JS divergence as the distance between probability vectors, the distance between items can be accurately measured.

  Next, the vector analysis unit 120 of the control unit 102 will be described.

  As shown in FIG. 6, the vector analysis unit 120 performs entropy calculation processing on the PLSI vectors indexed by the PLSI processing, sorts them in descending order based on the entropy values, and based on the results, the subjective tag and the objective tag Classify into: Alternatively, a correct subjective tag is acquired by some method, a distance from the subjective tag candidate is obtained, and a K-neighbor acquisition process for classifying the subjective tag and the objective tag based on the distance is performed.

The vector analysis unit 120 acquires tags with high entropy from the PLSI vectors as subjective tag candidates in descending order. In that case, the basic method is
(1) Sorting in order of entropy and obtaining tags of the top k items (or more than the minimum entropy value e) (naive method);
(2) A correct subjective tag is prepared by some method, and a tag in the vicinity of k (or less than the minimum distance d) is developed using the tag as a key;
It consists of two. 2. above. As a method for selecting a correct tag in (1), (1) can be used. In the method (1), it is not necessary to prepare a correct tag, but in (2), a correct tag is required. In the case of (1), subjective tag candidates are calculated largely depending on the nature of the data set that has executed PLSI. However, if (2) is used, a tag similar to the usage pattern is acquired. Become. These two ideas should be selected according to the data set and the features of the correct tags that can be prepared.

  The vector analysis unit 120 includes an entropy calculation unit 121, a sorting unit 122, and a correct subjective tag acquisition unit 123 that performs K neighborhood acquisition processing.

  FIG. 9 is a flowchart (No. 1) of the operation of the vector analysis unit in the embodiment of the present invention.

  The flowchart in the figure corresponds to the above method (1).

  Step 201) First, all tag data sets T converted into feature vectors by PLSI are prepared, and array A is initialized empty for output. It is also necessary to give the number k of subjective candidate tags or e which is a threshold value for the entropy value. Both k and e are parameters depending on the data set and service, and need to be appropriately tuned.

  The entropy calculation unit 121 calculates entropy values of all the PLSI vectors of T described above. Entropy is a value calculated for each tag (= PLSI vector) in the following equation.

ステップ２０２） ソーティング部１２２は、上記で求められたエントロピー値でＴをソートする。

Step 202) The sorting unit 122 sorts T by the entropy value obtained above.

  Step 203) The correct subjective tag acquisition unit 123 stores the tags satisfying the number k or the threshold e in the array A (memory).

  Step 204) The correct subjective tag acquisition unit 123 outputs the array A as a subjective tag candidate.

  FIG. 10 shows an example of the actually calculated PLSI vector (probability distribution) in one embodiment of the present invention. This figure shows PLSI vectors (probability distributions) of several tags calculated using the actual data by the above processing. FIG. 6A shows a PLSI vector having a strong characteristic in 75 dimensions among tags with an entropy value of 0, that is, no ambiguity. You can see that all tags are related to "Apple". On the other hand, FIG. 5B shows the top 16 tags with high entropy values, that is, ambiguous tags. Of the 16 subjective tag candidates shown, 13 can actually be determined as subjective tags. It should be noted that the number of subjective tag candidates can be increased by adjusting the two arguments of the number k and the entropy value e, but the accuracy decreases as the number of candidate tags increases. FIG. 11 shows the relationship between entropy and probability distribution in an easy-to-understand manner. When the probability distribution has only one peak, the entropy value is 0, indicating that there is no ambiguity. As the number of peaks increases, the entropy increases. In theory, the entropy takes the maximum value when the distribution is uniform. In reality, it cannot be a uniform distribution, so it can be understood that entropy is higher when the number of peaks is larger.

  Next, the method (2) will be described.

  FIG. 12 is a flowchart (part 2) of the operation of the vector analysis unit in the embodiment of the present invention. In the figure, the process when there is a correct subjective tag is shown.

  Similar to the processing of FIG. 9, a PLSI vector data set T is prepared and the array A is initialized.

  Step 301) First, a correct subjective tag is prepared in advance using a dictionary and previous results. If it cannot be prepared, the process proceeds to step 302.

  Step 302) If a correct subjective tag cannot be prepared, for example, a correct subjective tag set is prepared using the processing of FIG.

  Step 303) A K neighborhood tag is acquired for the correct subjective tag set. The K vicinity tag is an operation for acquiring the top K items among tags at a short distance from a non-odd number of tags. As a distance function between tags, for example, JS divergence, which is a distance between probability distributions that have been shown to be effective in information retrieval, can be used. JS divergence is calculated as follows.

ここで、Dk(x‖y)は、ｘ、ｙ間のＫＬダイバージェンスである。全てのＫ近傍タグをＡに格納し、主観的タグ候補として出力する。

Here, Dk (x‖y) is a KL divergence between x and y. All K neighborhood tags are stored in A and output as subjective tag candidates.

  As described above, in the collaborative classification system of the present invention, it is a classification axis by performing probabilistic clustering that is unsupervised learning on the co-occurrence information of the user, the classification axis, and the classification target that is the result of classification by the user. Convert tags into feature vectors. Tags are represented as a probability distribution by probabilistic clustering. The entropy value of the tag acquired as this probability distribution is calculated, and the tags are rearranged in descending order of entropy. A high entropy value means that the probability distribution is ambiguous, and a high degree of ambiguity in the result of probabilistic clustering means that the tag cannot be grouped well into any group. . The reason why a tag cannot be grouped well indicates that the tag is used differently than other tags. This is because subjective tags are used differently for each user, and clustering cannot be performed well. This technology sorts tags in descending order of entropy values of pre-assigned tag feature vectors, and the top k items or tags that exceed the threshold for entropy values are mechanically acquired as subjective tag candidates and by some means It consists of a fully automatic subjective tag distinction method that consists of a method in which a correct subjective tag candidate is used as a key and a neighboring tag based on a distance between probability distributions as a subjective tag candidate, and a manual tag removal support method that is manual. As the correct subjective tag, a dictionary may be prepared in advance, or a tag with a high entropy value may be used.

  The operation of the control unit 102 shown in FIG. 7 is constructed as a program, installed in an application server (computer) used as the cooperative classification device, and executed by the CPU or distributed through a network. Is possible.

  In addition, the constructed program can be installed or distributed in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, an optical storage device, a magnetic storage device, or the like.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

  The present invention is applicable to a system in which end users classify various information and share the classification results.

It is a figure for demonstrating the principle of this invention. It is a principle block diagram of this invention. It is an example of the classification | category in a cooperative classification system (SBM). It is a complete tripartite graph of classification behavior. It is a block diagram of the system outline in one embodiment of this invention. It is a figure which shows the outline | summary of the feature vectorization and vector analysis process in one embodiment of this invention. It is a block diagram of the application server in one embodiment of this invention. It is a flowchart of operation | movement of the feature vectorization part in one embodiment of this invention. It is a flowchart (the 1) of operation | movement of the vector analysis part in one embodiment of this invention. It is an example of the actually calculated PLSI vector (probability distribution) in one embodiment of this invention. It is the magnitude of a PLSI vector (probability distribution) and entropy value in an embodiment of the present invention. It is a flowchart (the 2) of operation | movement of the vector analysis part in one embodiment of this invention.

Explanation of symbols

1 Collaborative classification processor (system)
2 Client 21 Communication unit 22 Storage unit 23 Data processing unit 24 User interface 100 Application server 101 Communication interface 102 Control unit 103 Storage unit 103 Database 110 Feature vectorization means, feature vectorization unit 111 Element extraction unit 112 Noise removal unit 113 Co-occurrence Data indexing unit 120 Analysis means, vector analysis unit 121 Entropy calculation unit 122 Sorting unit 123 Correct subjective tag acquisition unit 124 K-neighbor tag acquisition unit 100 Application server 200 Database server 300 Feature vectorization engine 400 Vector analysis engine

Claims (10)

A user interface for inputting and outputting data; a communication means for performing data communication; and a data processing means for executing classification processing as classification target browsing processing and classification target identifier registration processing. A collaborative classification method in a system in which each user classifies and shares information such as videos, books, and articles,
A feature vectorizing step in which the data processing means performs feature vectorization of a tag that is a classification axis by probabilistic clustering with respect to a plurality of classification axes, classification targets, and user-related co-occurrence information stored in a database; ,
Based on the feature vector obtained in the feature vectorization step, an analysis step for distinguishing the classification axis arbitrarily given by the user into an objective tag and a subjective tag;
A collaborative classification method characterized by In the analyzing step,
An entropy calculation step of calculating an entropy value of the feature vector obtained by the feature vectorization step;
The cooperative classification method according to claim 1, wherein the entropy value higher than a predetermined threshold is a subjective tag, and the entropy value lower than the threshold is an objective tag. In the analyzing step,
Obtain a correct classification axis set from the database,
The collaborative classification method according to claim 1, wherein a neighboring tag whose distance is close to the correct classification axis set is a subjective tag candidate. In the analyzing step,
An entropy calculation step of calculating an entropy value of the feature vector obtained by the feature vectorization;
The entropy value is higher than a predetermined threshold as a subjective tag that is correct,
The cooperative classification method according to claim 1, wherein a neighboring tag that is close to the subjective tag that is the correct answer is a subjective tag candidate. A user interface for inputting and outputting data; a communication means for performing data communication; and a data processing means for executing classification processing as classification target browsing processing and classification target identifier registration processing. A collaborative classification device in a system in which each user classifies and shares information such as videos, books, and articles,
The data processing means includes
A feature vectorization means for vectorizing a tag that is a classification axis by probabilistic clustering with respect to co-occurrence information related to a plurality of classification axes, classification targets, and users stored in a database;
Based on the feature vector obtained by the feature vectorization means, the analysis means for distinguishing the classification axis arbitrarily given by the user into an objective tag and a subjective tag;
A collaborative classification device characterized by comprising: The analysis means includes
Entropy calculating means for calculating the entropy value of the feature vector obtained by the feature vector converting means;
Means that the entropy value is higher than a predetermined threshold as a subjective tag, and the lower than the threshold is an objective tag;
The cooperative classification device according to claim 5, comprising: The analysis means includes
6. The collaborative classification apparatus according to claim 5, further comprising means for acquiring a classification axis set that is a correct answer from the database and using a nearby tag that is close in distance to the classification axis set that is the correct answer as a subjective tag candidate. The analysis means includes
Entropy calculating means for calculating an entropy value of a feature vector obtained by the feature vectorization;
Means for making the entropy value higher than a predetermined threshold a subjective tag that is a correct answer;
Means for setting a near tag close to the subjective tag as the correct answer as a subjective tag candidate;
The cooperative classification device according to claim 5, comprising:   The cooperative classification program for functioning a computer as each means which comprises the cooperative classification device of any one of Claims 5 thru | or 8.   A computer-readable recording medium storing the cooperative classification program according to claim 9.

Images