KR20100013157A - Tag clustering apparatus based on related tags and tag clustering method thereof - Google Patents

Abstract

PURPOSE: A tag clustering apparatus based on related tags and a tag clustering method thereof are provided to form a tag cluster by gathering tags, thereby overcoming deterioration of a search result. CONSTITUTION: A related tag mapping module(132) extracts and maps related tag pairs each other. A frequency extracting module(134) extracts an appearance frequency of the same tags. A weight matrix generating module(136) generates a weight matrix. A tag clustering module(138) extracts only related tags with a frequency which is larger than a threshold value.

Description

Tag clustering apparatus based on related tags and tag clustering method

The present invention relates to an internet environment based on web 2.0, and more particularly, to a method and apparatus for using an associated tag to improve search accuracy when searching for content using a tag in a web service environment.

 With the development of the Internet, as the number of users is increasing rapidly, the web service environment is changing in various ways. While conventional web services are statically passive, web services are gradually changing dynamically and actively, and web 2.0 has been introduced to reflect the flow of web service changes.

Web 2.0 refers to a series of movements developed to induce information sharing and participation among Internet users through the opening of information, and to continuously increase the value of created information. In other words, in Web 2.0, netizens can freely participate, produce, recreate and share content based on an open web environment.

In Web 2.0, information is produced by the user, and the information is organized by tags that the user attaches. Users can easily share this information, so that various resources are correlated. As such, the Web 2.0 phenomenon has become an essential strategy for all Internet sites, and various techniques have been introduced to implement Web 2.0 successfully.

One of these techniques is tagging. Tagging is widely used in web documents such as blogs and multimedia contents such as images and videos, so that users can easily tag and search their own contents.

However, since this tagging is randomly attached by the producer of the content, the precision of information retrieval is low. In other words, a tagging technique may be useful for a user to include information in a wide category, but the process of retrieving useful information is not efficient because the category is too wide. For example, a user can tag a picture taken of a computer with his name. In this case, those pictures will not be exposed to anyone trying to retrieve pictures from the computer. That is, among tags tagged in a resource, there are many tags that are incorrectly tagged.

In addition, information retrieval navigation is inefficient because tagged tags are unstructured metadata. For example, if a user who wants to search a monitor picture of a computer performs a tag search using the name of the monitor, it is impossible to navigate using the feature that the monitor constitutes a part of the computer. It is treated entirely individually.

Therefore, there is an urgent need for a system that not only improves the accuracy of content search results using tags, but also enables information navigation between tags using tag interrelationships.

SUMMARY OF THE INVENTION It is an object of the present invention to provide an apparatus and method for forming a tag cluster by removing an incorrect tag and collecting only relevant tags to overcome the deterioration of the search result due to the incorrect tag.

Another object of the present invention is to identify the interrelationships between tags by applying a semantic model (ontology model) to the interrelationship between the tags, and by using the user to facilitate navigation between a plurality of tags An apparatus and method for generating a topic map are provided.

One aspect of the present invention for achieving the above object relates to a tag clustering device, the tag clustering device according to the present invention is associated tag pairs from each other from the two tags associated with the same content among the contents included in a predetermined population a related tag mapping module for extracting related tag pairs and mapping each other, a frequency extracting module for extracting a frequency of occurrence of the same tag in a tag mapping process, and a frequency of related tag pairs. A weight matrix generating module for generating a weight matrix based on the tag matrix, and a tag clustering for generating a tag cluster by extracting only related tags having a frequency greater than or equal to a threshold among the pairs of related tags from the weight matrix; It includes a module (tag clustering module). In particular, the association tag mapping module extracts the association tag pair by considering tags representing different pronunciations for the same foreign language, tags representing the main language and the abbreviation, and tags representing the synonyms as the same tag. .

An apparatus for tag clustering according to an aspect of the present invention includes a topic generating module for extracting tags in the tag cluster as a topic, and topics related to the pair of related tags among the topics in order of increasing frequency from smallest to smallest. A topic pair extracting module for extracting a pair, a semantic relation generating module for extracting a semantic relation of the topic pair by applying a predetermined lexical knowledge model to the extracted topic pair, and the topic pair And an occurrence generating module for assigning an address of a content suitable for the topic pair.

The tag clustering module included in the tag clustering apparatus according to an aspect of the present invention selects a base association tag pair among the association tag pairs, and includes another association tag including respective tags included in the selected base association tag pair. The tag cluster may be generated by repeatedly selecting a pair.

The tag clustering module included in the tag clustering apparatus according to an aspect of the present invention selects an associative tag pair having the highest weight as the basic associative tag pair, and includes the number of tags included in the generated tag cluster and the associative tag pair. The threshold value is determined in consideration of an average of weights.

Alternatively, the tag clustering module included in the tag clustering apparatus according to an embodiment of the present invention may determine the basic association tag pair and the threshold value according to a user selection.

The lexical knowledge model included in the tag clustering apparatus according to an aspect of the present invention includes a Resource Description Framework (RDF), Knowledge Query and Manipulation Language (KQML), Darpa Agent Markup Language-Ontology Inference Layer (DAML-OIL), and Ontology Web Language) and a topic map.

The occurrence generation module included in the tag clustering apparatus according to an aspect of the present invention is characterized by assigning a URL (Uniform Resource Locator) of the content including all the associated tag pairs corresponding to the topic pair to the topic pair.

Another aspect of the present invention for achieving the above object is, the association associated with each other to extract the relevant tag pairs (tag pairs) from each of the tags included in the predetermined population related to the same content (map) A tag mapping step, a frequency extraction step of extracting the appearance frequency of the same tag in the tag mapping process, a weighting matrix generation step of generating a weighting matrix for generating a weighting matrix based on the frequency of association tag pairs and the association tag from the weighting matrix A tag clustering method includes generating a tag cluster by extracting only related tags having a frequency greater than or equal to a threshold among pairs. In particular, the associative tag mapping step includes the steps of considering the tags representing different pronunciations for the same foreign language, the tags representing the main language and the abbreviation, and the tags representing the synonyms as the same tag.

The tag clustering method according to another aspect of the present invention is a topic generation step of extracting the tags in the tag cluster as a topic, extracting the topics related to the associated tag pair of the topics as the topic pair in order from the highest frequency A topic pair extraction step, a semantic relationship generation step of extracting a semantic relationship of the topic pair by applying a predetermined lexical knowledge model to the extracted topic pair, and an occurrence generation step of giving an address of content suitable for the topic pair to the topic pair It characterized in that it further comprises.

Tag clustering step included in the tag clustering method according to another aspect of the present invention, selecting a basic association tag pair of the association tag pair, another association tag pair including each of the tags included in the selected basic association tag pair And repeatedly selecting another association tag pair including respective tags included in the selected other association tag pair.

According to the present invention, since the tag cluster is formed using only the highly related tags by removing the incorrect tag, deterioration of the search result due to the incorrect tag is overcome and the quality of the search result is improved.

In addition, according to the present invention, by generating a topic map by applying a semantic model (ontology model) to the interrelationship between the tags, the user can know the interrelationship between the tags, and users can use the interrelationship to You can easily navigate.

In order to fully understand the present invention, the operational advantages of the present invention, and the objects achieved by the practice of the present invention, reference should be made to the accompanying drawings which illustrate preferred embodiments of the present invention and the contents described in the accompanying drawings.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In addition, in order to clearly describe the present invention, parts irrelevant to the description are omitted, and the same reference numerals in the drawings indicate the same members.

Throughout the specification, when a part is said to "include" a certain component, it means that it may further include other components, without excluding the other components unless otherwise stated. In addition, the terms "... unit", "... unit", "module", "block", etc. described in the specification mean a unit for processing at least one function or operation, which means hardware, software, or hardware. And software.

1 is a block diagram conceptually illustrating a tag clustering apparatus according to an embodiment of the present invention.

The tag clustering apparatus 100 illustrated in FIG. 1 includes a tag manager 110 and a display 190. The tag manager 110 includes a tag reader 120, a tag clustering unit 130, and a topic map generator 150. The tag clustering unit 130 includes an association tag mapping module 132, a frequency extraction module 134, a weight matrix generation module 136, and a tag clustering module 138. The topic map generation unit 150 includes a topic generation module 152, a topic pair extraction module 154, a semantic relationship generation module 156, and an occurrence generation module 158.

A tag refers to metadata that a user directly creates about any information, resource, or content. That is, the user can record metadata in the form of a tag for all kinds of web resources such as articles, images, videos, and favorites on the web. Metadata refers to a set of keywords or terms that the user considers relevant and appropriate for the information.

In the present specification, clustering refers to a technique for grouping data having similar characteristics and extracting common features of these data. That is, tags simultaneously appearing among tags included in the same content or information may be regarded as tags having high correlation with each other. For convenience of description, this tag pair is referred to as a related tag pair in the present specification. Finding relevant tag pairs among a large number of tags can eliminate meaningless tags, which can improve the accuracy of search results based on tags.

Ontology in the present invention is a set of vocabulary used in a specific field, and refers to a technique for conceptualizing and specifying knowledge in an application field. The purpose of the ontology is to provide improved information retrieval and semantic navigation by defining, integrating and sharing semantic relationships among distributed information resources.

The tags 180 are read by the tag reader 120 of the tag manager 110. The read tags are transferred to the tag clustering unit 130. In this process, the associative tag mapping module 132 does not simply compare the read tags, but considers similar tags that are conceptually determined to refer to the same object as the same tag, thereby excessively increasing the dimension of the weight matrix described below. Not only can it be prevented, it can reduce information throughput and improve search results. For example, the associated tag mapping module 132 may consider tags representing different pronunciations for the same foreign language as the same tag. For example, a computer, a computer, a computer, a computer, or the like can be regarded as representing a computer that is the same object. In addition, the associated tag mapping module 132 may regard the tags representing the main language and the abbreviation as the same tag. For example, Korea and Korea can be considered to be the same tag. In addition, the associated tag mapping module 132 may regard tags representing synonyms as the same tag. For example, the associated tag mapping module 132 may determine summer and summer as the same tag. In addition, the associative tag mapping module 132 may convert meaningless character strings due to mistranslation of the Korean-English keyboard into corresponding languages. For example, the tag 'zmffjtmxjfld' can be considered the same as the tag 'clustered', using the most widely used two-to-English keyboard conversion when the tag 'zmffjtmxjfld' is present.

The related tag mapping module 132 extracts a set of related tag pairs consisting of two tags related to the same content among a plurality of contents included in a predetermined population. That is, the association tag mapping module 132 maps tags that are all related to the same content as association tag pairs. See FIG. 2A and FIG. 2B to understand this mapping process.

2A and 2B are diagrams conceptually describing an operation of the association tag mapping module 132 illustrated in FIG. 1.

Multiple tags may be input to one content. Referring to FIG. 2A, the first tag group 200 includes tags such as woman, teacher, and school. Then, the associated tag mapping module 132 generates three association tag pairs by mapping each other because the tags woman, teacher, and school are included in the same content. In other words, three association tag pairs are created: woman-teacher, teacher-school, and school-woman.

Referring to FIG. 2B, the second tag group 300 includes tags school, people, children, and boy. That is, since the tags school, people, children, and boy are related to the same content, four related tag pairs are generated by mapping them. In addition, the first tag group 200 and the second tag group 300 include a tag called school in common. Therefore, the school generates association tag pairs in association with the tags (teacher, woman) included in the first tag group 200 and the tags (people, boy, children) included in the second tag group 300. It can be seen that.

Referring again to FIG. 1, a frequency extracting module 134 extracts a frequency, which is the number of times associated tag pairs occur in the contents. In other words, a frequency of 2 means that an associated tag pair occurs twice among contents belonging to a given population. The higher the frequency, the higher the association of the associated tag pair.

As such, the process of forming an association tag pair and obtaining a frequency is performed by using Equations 1 and 2 below.

T _G (i, j) = 0; When there is no association between tag i and tag j

T _G (i, j) = k; When tag i and tag j are both found k times

In Equations 1 and 2, T _G (i, j) represents an element of the row i column j columns of the weight matrix.

The frequency extracted by the frequency extraction module 134 is transferred to the weight matrix generation module 136, and the weight matrix generation module 136 generates a weight matrix using the received frequency for each pair of associated tags. 3A illustrates an example of a weight matrix generated by the weight matrix generation module 136.

Referring to FIG. 3A, the weight matrix includes the tags extracted from the content group in rows and columns, respectively. Each element of the weight matrix is the frequency of an associated tag pair. For example, an association tag pair called teacher-classroom has a weight of 2 and an association tag pair called classroom-school has a weight of 4. For ease of understanding, an associated tag pair having a weight of 4 is indicated by a circle, and an associated tag pair having a weight of 2 is indicated by a triangle.

When the weight matrix is configured as described above, the tag clustering module 138 generates a tag cluster by extracting only an associated tag pair having a frequency above a threshold. At this time, the larger the threshold (threshold) applied to generate the tag cluster, the fewer tag pairs are included in the tag cluster, and the smaller the threshold is applied, the larger number of tags are included in the tag cluster. The larger the number of tags included in the tag cluster, the higher the recall, but the lower the precision. The accuracy indicates how close the search result is to the desired search purpose when searching for a certain content using a predetermined tag. For example, if you try to search an Apple computer, and you see apples as fruits as a result, this is less accurate. The recall rate indicates how many percent of the search results contain the correct result.

It is very important for the tag clustering module 138 to set thresholds to generate tag clusters, which will be described in detail below with reference to FIGS. 5A and 5B.

The pseudo-code of the tag clustering algorithm is as follows.

// i: cluster number

// C (i): i-th cluster

// T (i, j): frequency between tag i and tag j, i.e. elements in row i column j of weight matrix T _G

// Max (i, j): the element with the maximum weight among the elements of T _G

// A _i : Weight matrix of tags in cluster C (i)

i = 1

// repeat until all tags with weight greater than threshold are included in cluster

A _i Initialization

Repeat {

// Select two tags i, j of element Max (i, j) with maximum weight in T _G and add them to cluster C (i)

select Max (i, j)

Add tag i and tag j to C (i)

Add element Max (i, j) to A _i

While (T (i, j)> = threshold) {

// tag element is greater than the weighted average of the size and the threshold value associated with the tag i and j both tags added to the weighting matrix Ai of the cluster C (i) T (i, j) in the selected weight matrix T _G C (i Add to)

Add tag i and tag j of T _G to C (i)

Add element T (i, j) to A _i

}

i = i + 1

} until (All (T (i, j))> = threshold)

This clustering process will be described in detail with reference to FIG. 3B.

First, a tag i (school) and a tag j (classroom) having the maximum weight among the weighting matrix are added to the cluster C (i). Then, from the weighting matrix T _G , T (i, j) is selected from elements weighted equal to or greater than the threshold among teachers, me, female, and women, tags related to tag i and j, added to C (i). Add to (i). This operation is repeated until all weights are smaller than the threshold. Referring to FIG. 3B, first, a school-classroom having the highest weight is extracted (indicated by a circle), and then, a classroom-teacher and teacher-school having a weight of 2 are extracted (indicated by a triangle).

As shown, the tag clustering module 138 extracts associative tag pairs that are related to each other from any tag group, and generates a tag cluster based on the frequency of the extracted associative tag pair, thereby associating tag pairs belonging to the tag cluster. They are closely related to each other.

In the tag clustering module 138, a threshold to be applied to perform tag clustering may be selected by a user, or an optimal threshold may be selected as follows.

5A and 5B are graphs for describing a process of determining a threshold in the tag clustering module 138 illustrated in FIG. 1. 5A illustrates the number of average tags included in a tag cluster when different thresholds of 2 to 12 are applied. As shown, it can be seen that as the threshold increases, the number of tags included in the tag cluster decreases. 5B illustrates the cohesion of tags included in a tag cluster when the threshold is increased. The cohesion degree of FIG. 5B represents a clustering degree of association tag pairs weighted in a predetermined tag cluster, and represents a weighted average of association tag pairs included in the tag cluster.

The degree of cohesion of any tag cluster C (i) is calculated using the following equation (3).

In Equation 3, Ai (j, k) means a tag j and a tag k connected by weight in the cluster C (i), n represents the total number of tags belonging to C (i).

A threshold may be selected based on the evaluation of the cohesion of the individual clusters using Equation 3, and for this purpose, the mean of the cohesion of the entire clusters is calculated as in Equation 4 below.

In Equation 4, m represents the number of tag clusters C (i) selected according to different thresholds.

Referring to FIG. 5B, the degree of cohesion also increases as the threshold increases, but after a certain value, there is little change in the degree of cohesion. In the graph shown in FIG. 5B this value will be 9. That is, it can be seen that when the threshold value is 9 or more, there is no change in the degree of cohesion. That is, the user may select the threshold as 9 with reference to the result as shown in FIG. 5B.

Referring back to FIG. 1, the tag cluster generated by the tag clustering module 138 is transferred to the topic map generator 150. Topic generation module 152 extracts the tags in the tag cluster as topics. The extracted topics are passed to the topic pair extraction module 154. The topic pair extraction module 154 extracts topics related to the associated tag pair among the topics as topic pairs in order from the highest frequency to the lowest frequency. Then, the extracted topic pairs are transferred to the semantic relationship generation module 156. The semantic relationship generation module 156 extracts the semantic relationship of the topic pair by applying a predetermined lexical knowledge model to the extracted topic pair. Then, the occurrence generation module 158 reflects the extracted semantic relationship and gives the topic pair an address of a content suitable for each topic pair. Hereinafter, the operation of each component of the topic map generator 150 will be described in detail.

The topic generation module 152 may provide a UI for constructing a topic map. The UI provides the user with information such as the name of the topic map, description, creator, distributor, and creation date. The topic generation module 152 automatically uses the tags included in the tag clustering module 138 as topics. The base name of a topic allows the user to understand each topic among the various topics. The base topic of each topic map may be selected by the user, or may be selected from the pair of most frequently associated tags.

Topic generation module 152 extracts the topic, topic pair extraction module 154 extracts the topic pair from the tag cluster. After all, it is a tag that constitutes a tag cluster, and these tags become topics when used in a topic map. These terms refer to terms that are used in a general sense in the art. Briefly introducing the pseudo-code used in the topic pair extraction module 154 as follows.

// C (i): Cluster selected by the user

// A (i, j): Weight matrix of tags in C (i)

// Max (A (i, j)): the element with the maximum value among the elements of the weight matrix A (i, j)

// T (k): set of all tags associated with tag of Max (A (i, j))

// B (l, m): Weight matrix of tags in T (k)

// Max (B (l, m)): the element with the maximum value among the elements of the weight matrix

// repeat until A (i, j) becomes empty

Repeat {

// extract maximum value among elements of weight matrix

Extract Max (A (i, j))

// construct T (k) by retrieving all tags associated with tags of Max (A (i, j)) from C (i)

Find T (k)

// repeat until B (l, m) becomes empty

Repeat {

// extract Max (B (l, m)) from the elements of the weight matrix

Extract Max (B (l, m))

// remove Max (B (l, m)) from B (l, m)

Remove Max B (l, m)) from B (l, m)

} until (B (l, m) == empty)

// delete B (l, m) from A (i, j)

Remove B (l, m) from A (i, j)

} until (A (i, j) == empty)

The pseudo code described above will be described with reference to FIGS. 4A to 4E.

4A through 4E are diagrams for conceptually describing an operation of extracting a topic pair in the topic pair extracting module 154 illustrated in FIG. 1.

First, among the generated tag clusters (see FIG. 4A), Max (A (i, j)) having the maximum weight among the elements of the weight matrix A (i, j) of the cluster C (i) selected by the user is selected, The selected tag i and tag j are extracted as associative tag pairs (see FIG. 4B). Then, all tags associated with tag i and tag j of Max (A (i, j)) are searched from C (i) to construct T (k) (see FIG. 4C). Then, the element Max (B (l, m)) having the maximum value among the elements of the weight matrix B (l, m) included in T (k) is selected (Fig. 4D). Thereafter, Max (B (l, m)) is deleted from B (l, m), and this process is repeated continuously (FIG. 4E).

When the topic pair extraction module 154 extracts a topic pair, the semantic relationship generation module 156 assigns a semantic relationship by applying a web-based ontology to the extracted topic pair.

The basics of ontology are the concepts that exist in the realm. For example, a topic called a book may have attributes such as author, publisher, number of pages, and price, and a topic of bidding may have attributes such as object, date, method, and condition. Topics can also be related to each other, the most basic of which is up and down relationships. For example, a topic called a fairy tale book becomes a sub-concept included in the book. As the ontology develops, it can contain rich contents by defining the characteristics of attributes and more complex forms of relationships. Viewing the ontology as an independent central component and placing it at the center of development and operation is an ontology-driven system, and the concept of web ontology is introduced. Web Ontology is a definition or specification of a vocabulary or concept and represents the components corresponding to the contents of the system in the information system field. In other words, the ontology describes a common vocabulary so that a computer can process a task that a person intuitively or semantically judges or processes to construct a semantic web. However, since the representation of all shapes is very difficult, the web ontology language was designed based on the extensibility generation language (XML) and the resource description framework (RDF) in the W3C.

Ontology Web Language (OWL) is a semantic web generation language for publishing and sharing ontologies that provide advanced web search, software agent and knowledge management functions on the web. The ultimate goal of the semantic web is to create a web that is a source of knowledge that can be understood by a computer. Today's Web, which is made up of documents in the form of HTML, provides information to humans, but computer programs We start with the consciousness of not being able to grasp exactly the content. OWL was developed as an extension of the Resource Description Framework (RDF) and originated from the DAML + OIL language. OWL is a language that defines Web ontology and its related knowledge. It defines propositions accumulated in the inference system, describes the relationships between classes and their members, and allows classes and attributes that enable logical inference of facts that are not syntactically defined. And a set of constraints applicable to this.

There are various systems using ontology. Among them, KQML-Knowledge Query and Manipulation Language and KIF-Knowledge Interchange Format are defined. In particular, DARPA's DAML-OIL (DARPA Agent Markup Language-Ontology Inference Layer) is accepted as a representative ontology expression language and format.

Alternatively, WordNet, a lexical knowledge model based on English, may be applied to the extracted pair of related topics. WordNet is information related to semantic patterns or usage patterns on words and can be referred to as a database that establishes associations between words. WordNet is openly distributed through the Java-based WordNet Library (JWNL), which can derive the relationship between two words, upper words, lower words, and synonyms.

The semantic relation generation module 156 included in the tag clustering apparatus 100 according to the present invention may extract an association relation between two selected topics from WordNet to give an automated semantic relation to a topic pair. For example, WordNet provides associations such as has kind, is a kind of, has members, is a member of, has particulars, is a particulars, has part, is a part of. For example, in the case of a topic pair school-classroom, since a relationship of "is part of" may be derived using WordNet, a semantic relationship of "classroom is part of school" may be established.

FIG. 6 illustrates an example of a topic map generated by the semantic relationship generation module 156 illustrated in FIG. 1.

The topic map shown in FIG. 6 is constructed by using computer as a basic topic. In Fig. 6, the relationship of (1) represents the relationship of "has part of", and the relationship of (2) represents the relationship of "has company of". In addition, the relationship of (3) shows the relationship of "has kind of". As shown, each topic included in the topic map shown in FIG. 6 is not simply listed, but a relationship between these topics can be known. Therefore, search performance is improved.

The occurrence generating module 158 included in the topic map generator 150 of FIG. 1 assigns an address of content suitable for the extracted topic pair to the topic pair. That is, the occurrence generation module 158 attaches URL information of content corresponding to each extracted topic. In this process, since the content corresponding to the topic is related to the topic, the search performance is improved later. For example, content that includes the tag "apple" could have a fruit apple, and an Apple computer. In this case, the occurrence generation module 158 assigns the URL address of the content corresponding to the fruit apple to a topic (ie, a tag) included in the topic map representing the fruit apple. In addition, a topic included in a topic map representing an apple computer is given a URL address of content corresponding to the computer apple. As a result, the fruit apple and the computer apple are clearly distinguished and searched.

7 is a flowchart of a tag clustering method according to another aspect of the present invention.

First, two pieces of related tag pairs are extracted from two tags related to the same content among the contents included in a predetermined population and mapped to each other (S710). Then, a frequency that is the number of times that the extracted association tag pairs occur in the contents is extracted (S720).

The extracted frequency is used to generate a weight matrix (S730). The manner of generating the weight matrix is as described above with respect to the weight matrix generation module 136 of FIG.

Then, only tag tags having a frequency greater than or equal to a threshold among pairs of related tag pairs are extracted to generate a tag cluster (S740). As described above, when determining a threshold indicating a threshold in this process, the number of tags included in the tag cluster and the degree of aggregation thereof may be referred to.

When the tag cluster is generated, a topic and an associated topic pair are extracted from the generated tag cluster (S750). Topic and association topic pairs are terms used in the ontology and correspond to the tag and association tag pairs used in the tag cluster as described above.

Then, a semantic relation is given to the extracted related topic pair by using an ontology (S760). Finally, the address of the content suitable for the topic pair is allocated (S770).

According to the tag clustering method and apparatus according to the present invention, the deterioration of search results due to incorrect tags and the inefficiency of navigation due to unstructured tags are overcome. For example, as a result of comparing the search results of the tag-based site Flickr with the tag clustering method according to the present invention, the following results are obtained.

The comparison test was performed using tags assigned to 120 images each searched using the keywords computer, apple, and jaguar. As a result, the accuracy and reproducibility of the Flickr site was 45.8% on average, whereas the accuracy of the system to which the present invention was applied was 90.4% on average and the reproducibility was 42.8% on average.

Although the present invention has been described with reference to the embodiments shown in the drawings, this is merely exemplary, and it will be understood by those skilled in the art that various modifications and equivalent other embodiments are possible. For example, in addition to the CAST (Complexity Analysis of Sequence Tracts) algorithm applied to extract a highly relevant association tag pair in the tag clustering module 138, an application in GCG (SEQ) used in bio-informatics. Of course, techniques such as Sequence Clustering, BLASTCLUST, PROCLUST, TribeMCL, and GeneRAGE may be applied. That is, all techniques that can form a tag cluster by extracting an associated tag pair having a predetermined weight or more from the associated tag pair may be applied to the tag clustering module 138.

Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

The present invention can be applied to a tag based search system.

1 is a block diagram conceptually illustrating a tag clustering apparatus according to an embodiment of the present invention.

2A and 2B are diagrams conceptually describing an operation of the association tag mapping module 132 illustrated in FIG. 1.

3A and 3B are diagrams illustrating a process of extracting tags having a high frequency based on a frequency in a weight matrix.

4A through 4E are diagrams for conceptually describing an operation of extracting a topic pair in the topic pair extracting module 154 illustrated in FIG. 1.

5A and 5B are graphs for describing a process of determining a threshold in the tag clustering module 138 illustrated in FIG. 1.

FIG. 6 illustrates an example of a topic map generated by the semantic relationship generation module 156 illustrated in FIG. 1.

7 is a flowchart of a tag clustering method according to another aspect of the present invention.

Claims (14)

In the tag cluster device, A related tag mapping module for extracting two related tag pairs from two tags related to the same content among the contents included in a predetermined population and mapping them to each other; A frequency extracting module for extracting a frequency of appearance of the same tag in a tag mapping process; A weight matrix generating module for generating a weight matrix based on the frequency of the associative tag pairs; And A tag clustering module for generating a tag cluster by extracting only relevant tags having a frequency greater than or equal to a threshold among the associated tag pairs from the weight matrix, wherein the associated tag mapping module includes: And a tag representing different pronunciations for the same foreign language, tags representing the main language and the abbreviation, and tags representing the synonym, respectively, and extracting the association tag pairs as the same tag. The method of claim 1, A topic generating module for extracting tags in the tag cluster as topics; A topic pair extracting module for extracting topics related to the associative tag pair from among the topics as topic pairs in order of increasing frequency; A semantic relation generating module for extracting semantic relations of the topic pairs by applying a predetermined lexical knowledge model to the extracted topic pairs; And And an occurrence generating module for assigning an address of content suitable for the topic pair to the topic pair. The tag clustering module of claim 2, wherein the tag clustering module comprises: The tag cluster is generated by selecting a base association tag pair among the association tag pairs and repeatedly selecting another association tag pair including respective tags included in the selected base association tag pair. Tag clustering device based on the associated tag. The tag clustering module of claim 3, wherein the tag clustering module comprises: Selecting a pair of tag with the highest weight as the base associated tag pair, And determine the threshold value in consideration of the average number of tags included in the generated tag cluster and the weight of the associated tag pair. The tag clustering module of claim 3, wherein the tag clustering module comprises: The apparatus for tag clustering based on an association tag, wherein the basic association tag pair and the threshold are determined according to a user selection. The method of claim 2, The lexical knowledge model includes at least one of a Resource Description Framework (RDF), Knowledge Query and Manipulation Language (KQML), a Darpa Agent Markup Language-Ontology Inference Layer (DAML-OIL), Ontology Web Language (OWL), and a topic map. Tag clustering device based on the associated tag, characterized in that. The method of claim 2, wherein the occurrence generation module, The tag clustering device based on the association tag, characterized in that to give the topic pair a Uniform Resource Locator (URL) of the content including all of the association tag pair corresponding to the topic pair. In the tag cluster method, An association tag mapping step of extracting two related tag pairs from two tags related to the same content among the contents included in a predetermined population, and mapping each other; A frequency extraction step of extracting an appearance frequency of the same tag in a tag mapping process; A weight matrix generation step of generating a weight matrix based on the frequency of the associative tag pairs; And A tag clustering step of generating a tag cluster by extracting only the associated tags having a frequency greater than or equal to a threshold among the pair of association tags from the weight matrix, wherein the association tag mapping step, And a tag representing different pronunciations for the same foreign language, tags representing the main language and the abbreviation, and tags representing the synonyms, respectively, as the same tag. The method of claim 8, A topic generation step of extracting tags in the tag cluster as topics; A topic pair extraction step of extracting topics related to the associative tag pair among the topics as topic pairs in descending order of frequency; Generating a semantic relationship by applying a predetermined lexical knowledge model to the extracted topic pairs; And And generating an occurrence of assigning an address of a content suitable for the topic pair to the topic pair. The method of claim 9, wherein the tag clustering step, Selecting a basic association tag pair among the association tag pairs; Selecting another associated tag pair including respective tags included in the selected basic associated tag pair; And And repeatedly selecting another association tag pair including respective tags included in the selected other association tag pair. The method of claim 10, wherein the tag clustering step, Selecting an associated tag pair having the highest weight as the base associated tag pair; And And determining the threshold value in consideration of the average of the number of tags included in the generated tag cluster and the weight of the associated tag pair. The method of claim 10, wherein the tag clustering step, And determining the basic association tag pair and the threshold according to a user selection. The method of claim 9, The lexical knowledge model includes at least one of RDF, KQML, DAML-OIL, OWL, and topic map. The method of claim 9, wherein the occurrence generating step comprises: And assigning to the topic pairs a content including all of the associative tag pairs corresponding to the topic pairs.

Images