KR20090017830A - Apparatus for providing aspect-based documents clustering that raises reliability and method therefor - Google Patents

Abstract

An apparatus for providing aspect-based documents clustering and a method thereof are provided to improve reliability by correcting an error of a clustering result by a feedback of a user. A document clustering part(100) extracts a keyword of documents, calculates importance of each keyword, executes a document clustering with the keyword and the importance, and corrects a result of the document clustering by using a user feedback. A keyword extracting unit(110) classifies the documents, and extracts a keyword from a classified documents. An importance calculating unit(120) calculates an importance of each keyword based on a text content of the documents. A similarity calculating unit(130) calculates a similarity between the documents. A clustering number decision unit(140) groups the documents, determines the number of document clustering, and generates a document clustering result according to the number of clustering. A post processing unit(150) corrects an error generated in the clustering result by a cooperative clustering.

Description

Apparatus for providing Aspect-based Documents Clustering that raises Reliability and Method therefor}

The present invention relates to a document structure-based clustering device and method for improving reliability, and more particularly, to perform document clustering by reflecting the internal structure of a document and correcting errors that may occur in the clustering result by using user feedback. The present invention relates to a document structure-based clustering device and a method for improving the reliability.

Document clustering is the discovery of groups of documents with similar characteristics from a set of documents by a clustering algorithm. Document clustering is an important document analysis method used in many applications of information retrieval, such as organizing data, browsing web search results, or multiple document summaries.

In document clustering, the distribution of data sets, internal structure, or the type of clustering desired by users have a significant effect on clustering results. There have been studies to improve the quality of clustering by providing limited information to users.

In addition, various document clustering techniques have been proposed for the purpose of providing convenience to a user's document access. For example, techniques for automatically classifying documents covering the same event or the same subject.

However, these techniques do not take into account the fact that documents can be created based on different contents from different viewpoints according to the authors within one subject, so that it is difficult for the user to comprehensively and objectively recognize the subject based on various viewpoints. there was.

The problem to be solved by the present invention is to solve the above problems, the document that improves the reliability to induce the user to recognize that even if the document on the same event or the same topic, the point of view that the document creator is important may vary. It is to provide a structure-based clustering device and method.

Another problem to be solved by the present invention is to solve the above problems, a document structure-based clustering device and method for improving the reliability that the user can objectively and comprehensively recognize a specific subject or event based on various viewpoints. To provide.

Another object of the present invention is to solve the above problems, and to provide a document structure-based clustering device and method for improving reliability by using a gradual error correction of the clustering result based on user feedback.

The present invention relates to a document structure-based clustering device with improved reliability. The present invention classifies input documents into subjects, extracts keywords reflecting key contents of classified documents, calculates importance of extracted keywords, extracts keywords and A document clustering unit which executes document clustering using the calculated importance and supplements the results of the document clustering using the cooperative clustering; And an input / output unit configured to receive document information that is a target of document cluster, output a document cluster result, and receive error information of the output document cluster result. It includes.

Preferably, the document grouping unit, the keyword extraction means for classifying the input documents by subject, extracting a keyword from the classified documents; Importance calculation means for calculating the importance of the extracted keyword based on the body content of the document; Similarity calculating means for calculating a similarity between documents using the extracted keyword and the calculated importance; Cluster number determining means for determining a suitable number of document clusters using the similarity calculation result and generating a document cluster result according to the determined number of clusters; And post-processing means for correcting errors of the cluster result by the cluster number setting means by the cooperative cluster. Characterized in that it comprises a.

Also preferably, the similarity calculation may be performed by cosine measurement.

Also preferably, the received documents may be news articles or newspaper articles.

Also preferably, the document group extracts a keyword from a title, subtitle, and lead of a news article or newspaper article.

Also, preferably, the importance of the keyword may be calculated using one or more variables among the position of the paragraph including the keyword, the number of occurrences of the keyword, the length of the paragraph including the keyword, and the length of the sentence including the keyword. It features.

Also preferably, the importance of the keyword may be calculated by the following equation.

Where v is the importance value of the keyword, La (i) is the length of the i-th document of the document set, and Lp (i, j) is the length of the j-th paragraph in the i-th document of the document set, Ls (i , k) is the length of the kth sentence in the i-th document of the document set, d _Pj is the diminishing factor of the j-th paragraph in the i-th document of the document set, and d _Sk is the i-th of the document set Refers to the deminishing factor of the kth sentence in the document.)

Also preferably, the document clustering may be performed using a Hierarchical Agglomerative Clustering method.

Also preferably, the number of suitable document clusters may be automatically determined using Elbow Criterion.

Preferably, the input / output unit may include any one or more of a tagging interface and a drag and drop interface.

On the other hand, the present invention relates to a document structure-based clustering method to improve the reliability, comprising the steps of: (a) receiving document information that is the target of the document cluster; (b) extracting keywords from the input documents; (c) calculating the importance of the extracted keyword; (d) calculating similarity between documents using the extracted keyword and the calculated importance; (e) generating document clustering results using the similarity calculation results; (f) outputting the generated document clustering result; And (g) complementing the output document clustering result by the cooperative clustering. It includes.

Preferably, the step (b) comprises: (b-1) classifying the received documents by subject; Characterized in that it comprises a.

Also preferably, the step (b) may include: (b-2) removing the stopwords from the received documents; And (b-3) extracting roots from documents from which stop words have been removed; Characterized in that it comprises a.

Also preferably, the step (c) may include: (c-1) calculating a diminishing factor to consider the position of the paragraph or sentence containing the keyword in the document; (c-2) calculating a ratio of the length of the paragraph or sentence including the keyword to the total document length; And (c-3) calculating a sum of importance for each occurrence of each keyword in the document; Characterized in that it comprises a.

Also preferably, the step (g) may include: (g-1) receiving a keyword of a document from a user through a tagging interface; Characterized in that it comprises a.

Also preferably, the step (g-1) may include: (g-1-1) searching for documents having a word set that does not include a keyword input through the tagging interface; (g-1-2) adding a keyword input through the tagging interface to the searched word set; And (g-1-3) calculating importance of the added keyword; Characterized in that it comprises a.

Also preferably, the step (g) may include: (g-2) receiving a movement signal of a document determined to be a wrongly clustered document by a user through a drag and drop interface ; Characterized in that it comprises a.

And, preferably, the step (g-2), (g-2-1) the words included in the document to be moved overlaps the word set of the cluster after the movement and the word set of the cluster before the movement Categorizing into words; And (g-2-2) assigning different weights to words that overlap with the word set of the cluster after the movement and words that overlap with the word set of the cluster before the movement; Characterized in that it comprises a.

According to the present invention, there is an effect that a user can recognize that even a document related to the same event or the same subject may have various points of view that the document creator places importance on.

According to the present invention, there is an effect that the user can objectively and comprehensively recognize a document dealing with a subject of interest based on various points of view with minimal time or effort.

According to the present invention, there is also an effect that can improve the reliability by using the continuous error correction of the clustering result by the user feedback.

Before describing the details for carrying out the present invention, it should be noted that configurations that are not directly related to the technical gist of the present invention are omitted within the scope of not distracting the technical gist of the present invention.

In addition, the terms or words used in the present specification and claims are consistent with the technical spirit of the present invention on the basis of the principle that the inventor can appropriately define the concept of the term in order to explain the invention in the best way. It should be interpreted as meaning and concept.

The present invention proposes a document structure-based clustering device and method for classifying and grouping documents based on a structure based on a viewpoint that a document focuses, and an apparatus and method for correcting errors that may be included in a clustering result using feedback of users. Initiate.

Hereinafter, a document structure based clustering device having improved reliability according to an exemplary embodiment of the present invention will be described with reference to FIGS. 1 to 3.

1 is an overall configuration diagram of a document structure based clustering device having improved reliability according to a preferred embodiment of the present invention, FIG. 2 is a schematic diagram of a general article structure, and FIG. 3 is a drag according to a preferred embodiment of the present invention. An illustration of an end drop interface.

As shown in FIG. 1, the document structure based clustering device having improved reliability according to the preferred embodiment of the present invention includes a document clustering unit 100 and an input / output unit 200.

The document cluster unit 100 extracts the keywords of the input documents, calculates the importance of each extracted keyword, executes document clustering using the extracted keywords and the importance, and uses the user feedback as a result of the document clustering. It performs a function of correcting, and includes a keyword extraction means 110, importance calculation means 120, similarity calculation means 130, cluster number determining means 140 and post-processing means 150.

The keyword extracting unit 110 classifies the input documents into subjects and extracts keywords from the classified documents.

In a preferred embodiment of the present invention, the received document information may be implemented as Internet newspaper articles, Internet news articles, various articles, various books, and the like.

Techniques used for subject classification of documents may use conventional techniques, and thus detailed description thereof will be omitted.

The keyword extraction by the keyword extraction means 110 focuses on the fact that most documents are composed of a title, subtitle, and text. In other words, noun equivalent phrases extracted from the titles or subtitles of documents classified by subjects using morphological analysis are selected as keywords of articles. The keywords selected represent important content covered in each document.

For example, when a document targeted for a document cluster is a news article or a newspaper article, as shown in FIG. 2, the document includes a head, a sub-head, a lead, and a main text. It is common to be. This is due to the inverted pyramid type writing rules typically used in news reporting, where journalists are trained to place news information or facts to be conveyed in articles in order of importance. Therefore, if you use this to extract keywords, it is possible to grasp the important aspects of the article.

In news articles or newspaper articles, headings and subtitles contain important information that can tell you the direction of the article, since it is an important factor in determining whether to read the article. And, a lead usually means the first sentence or the first two sentences of an article, and an important content that covers the entire content of the article comes out. In the following text, the information supporting the contents of the lead is arranged in order of importance.

Therefore, when the document to be a document cluster is a news article or a newspaper article, it is desirable to extract keywords from the title, subtitle and lead to improve the accuracy of the document cluster.

Next, the importance calculating means 120 calculates the importance of each extracted keyword based on the body content of the document.

Hereinafter, the importance calculation of keywords according to a preferred embodiment of the present invention will be described.

The higher the importance of the keyword appearing in the front part of the main body and the repeated occurrence of the keyword in the main body, and the longer the paragraph or sentence in which the keyword appears, the higher the importance of the keyword.

Accordingly, the importance of the keyword is calculated by four variables consisting of the position of the paragraph containing the keyword, the number of occurrences of the keyword, the length of the paragraph containing the keyword and the length of the sentence containing the keyword, The present invention is not limited thereto.

Table 1 below summarizes the variables used in the importance calculation according to the preferred embodiment of the present invention.

sign Explanation A (i) I-th document in the document set La (i) Length of A (i) Lp (i, j) The length of the jth paragraph in A (i) Ls (i, k) Length of kth sentence in A (i) W (i) Set of keywords in A (i) w _n Nth keyword of W (i) v _n Importance value of wn Va (i) Set of keyword, importance pairs d _pj Diminishing Factor of the jth paragraph in A (i) d _sk Diminishing Factor of the kth sentence in A (i)

In this embodiment, the four variables described above were applied to the importance calculation of the keyword in the following manner.

First, a diminishing factor is used to consider the importance according to the position of a paragraph or sentence including a keyword. Second, use the ratio of the length of the paragraph containing the keyword to the total article length. Third, the ratio of the length of the sentence including the keyword to the total article length is used. Fourth, to reflect the number of occurrences of keywords in the article, the sum of the importance of each keyword is calculated.

In the following two articles, an example in which the above-described importance calculation is applied to a news article will be described.

As shown in the two examples above, first, noun equivalent phrases are extracted from the title and the lead as keywords. Examples of extracted keywords are shown in squares.

The importance of each keyword is calculated in proportion to the length of the paragraph or sentence in which the keyword appears over the entire length of the article. For example, if the length of a paragraph in which a keyword appears is 1/4 of the total length of the article, it is given an importance of 0.25. If the length of the sentence in which the keyword appears is 1/8 of the length of the article, it adds 0.125. The way is possible.

In addition, in order to change the importance of the keyword according to the position of the paragraph or sentence containing the keyword, it reflects the diminishing factor for each paragraph or sentence.

The deminimization factor may be set to decrease as the paragraph or sentence is located later in the article, for example, if the length of the first paragraph is one quarter of the length of the article, Reflecting 1, and the second paragraph reflecting the deminishing factor 0.75.

An example of such importance calculation has been described in the case of a news article. However, the above-described importance calculation is not limited to this, and it is obvious that the article can be applied to documents such as Internet newspaper articles, various articles, and various books. If it is arranged, more accuracy can be achieved.

The above-described importance calculation process can be clearly expressed by the following [Equation 1].

Since the mathematical meaning of each variable included in [Equation 1] is described in [Table 1], it will be omitted here.

Next, the similarity calculating means 130 calculates the similarity between the documents using the keywords extracted from each document and the calculated importance.

The set Va (i) of keyword and importance pairs extracted from each document can be viewed as a vector on Euclidean space, and the similarity is calculated based on this vector.

It is preferable to use a cosine measure for the calculation of the similarity, because the similarity of the document is appropriately calculated in consideration of the ratio of the importance values rather than the absolute value of the importance values.

The cosine measurement is one of the representative methods for measuring the distance between two vectors. This is a method of measuring the similarity using the difference between the directions of two vectors, that is, the angle between two vectors.

For example, if there are two two-dimensional vectors (2,2) and (1,1), these two vectors face the same direction and are judged to be the same in similarity. This cosine measure is primarily contrasted with the method of calculating the Euclidean distance between two vectors.

Next, the group count determining means 140 groups the documents using the similarity calculation result, determines the number of suitable document clusters, and generates a document cluster result according to the determined number of clusters.

As a way to group documents, you can use the Hierarchical Agglomerative Clustering method that automatically determines the number of article groups.

Since it is impossible to predict how many kinds of document groups the documents targeted for document grouping are to be classified, the number of document groups cannot be determined in advance in performing the document grouping according to the present embodiment.

The Hierarchical Agglomerative Clustering method is one of the typical hierarchical clustering methods and generates a document group by combining the nearest cluster pairs.

In the Hierarchical Agglomerative Clustering method, all individual elements are set to one cluster, and clustering is started. The clustering is performed until all the elements are combined into one cluster while combining the nearest vector. If there are several vectors constituting a cluster, the average value of each vector becomes the representative value of the cluster, and the two nearest clusters are combined by calculating the distance using the representative value.

The number of suitable document clusters is automatically determined using Elbow Criterion.

When applying the Hierarchical Agglomerative Clustering method, a clustering result is generated according to the number of clusters. That is, a variety of results are possible from the case where all elements form one cluster to the case where all elements each individually form a cluster.

The Elbow Criterion is one of methods for determining the number of clusters among the number of formation of these various clusters, and uses the degree to which the dispersion (the degree of separation of vectors from the center of the cluster) is reduced.

For example, if all the elements are in a cluster, there are more variances than if the elements formed two clusters because several vectors were spread from the center of the cluster. By using this property, the number of clusters is determined at a point where the variance does not decrease even if the number of clusters is increased while increasing the number of clusters.

Finally, the post-processing means 150 corrects errors that may occur in the clustering result by the cluster number setting means 140 by cooperative clustering.

Hereinafter, the cooperative cluster by the post-processing means 150 will be described.

If the document structure based clustering method described above is actually applied to document clustering, unexpected errors may occur. Keywords that are important in the document may appear later in the document, and there may be documents that use different terms for substantially the same subject. In such a case, an error may occur in extracting key keywords of the document, and the accuracy of the clustering result may be reduced.

Since embedding semantic knowledge is a very complex and difficult task, the cooperative cluster according to the preferred embodiment of the present invention corrects the above-described clustering error by utilizing the collective intelligence of users. Users can fully understand the meaning of a document and find the focus of the document. Therefore, gathering and utilizing feedback from users can be an appropriate way to improve the accuracy of cluster results.

The cooperative clustering according to the present embodiment may be performed by using feedback of users collected by a tagging interface and a drag and drop interface of the input / output unit 200. Users provide words that reflect the focus of the document through the tagging interface and reflect the provided words in the extracted word set. In addition, it is desirable to consider the words provided through the tagging interface more importantly when calculating the similarity.

More specifically, documents with a word set that does not include a keyword entered through the tagging interface are searched, a keyword entered in the word set is added, and importance is calculated in the same manner as other keywords. Through this, it is possible to compensate for the case that the important keyword is not automatically extracted due to the characteristics of the document structure or the limitation of the stemmer.

Preferably, users can select misclassified documents through a drag and drop interface and move them to the appropriate cluster. The document moved by the drag and drop is set in a semantic relationship close to the cluster after the move, and is set in a semantic relationship far from the cluster before the move. 3 illustrates an example of document movement by the drag and drop interface.

The semantic relationship may be reflected in the clustering result by adjusting weights assigned to words.

Specifically, words included in the moved document are classified into words overlapping with the word set of the cluster after the movement and words overlapping with the word set of the cluster before the movement. The words that overlap with the word set of the cluster after the movement are given weights similar to the weights of the cluster words after the movement, and the words overlapping with the word set of the cluster before the movement are set to be close to zero by reducing the weight. In this manner, in the above-described distance calculation between the vectors, it is possible to move away from the cluster before the movement and to approach the cluster after the movement.

The post-processing means 150 performs post-processing of the document clustering result through the cooperative cluster described above, and supplements the document clustering result by reflecting the performed post-processing result.

In addition, the input / output unit 200 receives document information that is a target of document cluster, outputs a result of document cluster, and receives error information of a cluster result by using user feedback.

Document information that is a target of document clustering may be set to be input by connecting to a wired Internet network using a TCP / IP protocol, a wireless Internet network using a WAP protocol, or the like, or to be input from a USB memory, or the like. It is not.

In addition, the input / output unit 200 includes an Internet browser (eg, Netscape, Internet Explorer) capable of displaying document information in the form of Hyper Text Markup Language (HTML), and the web using the provided Internet browser. It is preferable to include a function of displaying document information or document clustering results on a page, but is not limited thereto.

In addition, the input / output unit 200 preferably includes a tagging interface and a drag and drop interface.

The key word input through the tagging interface may be reflected in the document clustering process, or the user may directly modify the document clustering result through the drag and drop interface to compensate for the document clustering result.

Hereinafter, a document structure based clustering method of improving reliability according to a preferred embodiment of the present invention will be described with reference to FIG. 4.

4 is a flowchart illustrating a document structure based clustering method of improving reliability according to a preferred embodiment of the present invention.

First, as shown in the figure, the input and output unit 200 receives the document information that is the target of the document group (S10).

Next, the keyword extraction means 110 extracts a keyword from the input documents (S20).

The step S20 may include classifying the received documents by subject.

The subject classification may be performed using conventional subject classification techniques, and thus detailed description thereof will be omitted.

Preferably, the step S20 may include removing stop words from the input documents and extracting roots. The removal of the stopword may be performed through a stopword list (eg, Rijsbergen, etc.).

For reference, stopwords (Stop-words or Noise Word) refers to words that are ignored at the time of search or words that search engines exclude from the index when building the database. Verbs, including, in, surveys, suffixes, conjunctions, endings, etc., and in English, including a, an, is, are, if, for, but, this, may, There are modal verbs, personal pronouns, instructional pronouns, and prepositions.

In addition, the root is a morpheme representing the actual meaning of the word among the morphemes constituting the word, is a real real morpheme synthesized in the case of a compound word, the rest of the stem except the affix in the case of derivatives. For example, the roots include the cover, the cover, the person, the person, and the generous.

Next, the importance calculation means 120 calculates the importance of each extracted keyword based on the body content of the document (S30).

In step S30, the step of calculating a diminishing factor to consider the position of the paragraph or sentence containing the keyword in the document, the length of the paragraph or sentence containing the keyword for the entire document length It is preferable to include calculating a ratio and calculating a sum of importance for each occurrence of each keyword in the document.

Next, the similarity calculating means 130 calculates the similarity between the documents using the keywords extracted from each document and the calculated importance (S40).

Next, the group count determining means 140 groups the documents using the similarity calculation result, determines the number of suitable document clusters, and generates a document cluster result according to the determined number of clusters (S50).

Next, the input / output unit 200 outputs the generated document clustering result (S60).

Finally, the post-processing means 150 corrects errors that may occur in the clustering result by the cluster number setting means 140 by the cooperative clustering (S70).

The step S70 may include receiving a keyword of a document from a user through the tagging interface or receiving a movement signal of a document determined to be a wrong group of documents by the user through the drag and drop interface. Do.

Hereinafter, a verification result of a document structure-based clustering device and method with improved reliability according to an exemplary embodiment of the present invention will be described.

News articles were used as documents to be verified, and two verifications were performed.

First, the first verification was performed using a set of news articles extracted from each of the political, economic, and social aspects of Google News. The document-based clustering was performed on the extracted news articles, and the clustering results for each set of news articles were provided to a large group of four people.

We then requested that we correct the misclassified articles in the cluster results provided. Those who participated in the verification moved the articles that they thought were misclassified to a group that they thought were appropriate, or created a new group and included them in the created group when they judged the article based on a different viewpoint than the previously classified groups. .

The verification results were compared with the results of the TF (Term Frequency) method and the structure-based clustering method in which keywords were selected only in the title and the lead, and the importance of the keywords was calculated.

The TF method is a method of extracting keywords from all articles regardless of the article structure and assigning a weight based only on the number of occurrences of the keywords. In addition, the results of the structure-based clustering method in which keywords are selected only in the title and the lead and the importance are calculated can be meaningfully compared with those in which the keyword is selected and the importance is calculated including the text.

The verification results are as shown in the following [Table 2] and [Table 3].

Average error count Structure based method (Title + Lead + Body) Structure based method (Title + Lead) TF method Politics (12 articles) 3 4 3.5 Economic aspect (19 articles) 0.25 2 5 Social aspect (18 articles) 5.75 6 8

Total error count Structure based method (Title + Lead + Body) Structure based method (Title + Lead) TF method Politics (12 articles) 4 5 4 Economic aspect (19 articles) One 2 7 Social aspect (18 articles) 7 8 14

[Table 2] and [Table 3] shows the total number of articles indicated as errors with respect to the average and the schedule of the number of articles identified as articles in which errors occurred by users participating in the verification for each subject.

In [Table 2] and [Table 3], it can be seen that the best performance is obtained when applying a structure-based clustering method in which keywords are selected from titles and leads, and the importance is calculated by reflecting the text. In the case of social articles, even better performance.

The second test was carried out by dividing 75 participants with an age distribution from 25 to 35 years into three groups of 25 people each, providing articles on the same topic. Participants had not taken a detailed approach to the subject matter prior to the verification point. In addition, participants were fairly distributed among the three groups to ensure that the habit of reading the article did not affect the verification results.

Each group was provided with articles based on structure-based classification and articles in Google News according to a preferred embodiment of the present invention, wherein the average number of articles subscribed by participants and the average number of subscription views viewed from articles were as follows. Table 4].

Average subscription article count Average Subscription Perspectives Structure-based classification 4 3 Google news 2.5 1.1

As shown in [Table 4], when articles were provided by the structure-based classification, participants could see a relatively large number of articles and could see various viewpoints on the same subject. In particular, in the case of the average number of subscription viewpoints, it can be seen that the structure-based classification method provides three times more various views than the Google News.

As described above and described with reference to a preferred embodiment for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as shown and described as described above, it is a deviation from the scope of the technical idea It will be understood by those skilled in the art that many modifications and variations can be made to the invention without departing from the scope of the invention. Accordingly, all such suitable changes and modifications and equivalents should be considered to be within the scope of the present invention.

1 is an overall configuration diagram of a document structure based clustering device with improved reliability according to a preferred embodiment of the present invention.

2 is a schematic diagram of a general article structure.

3 is an exemplary diagram of a drag and drop interface in accordance with a preferred embodiment of the present invention.

4 is an overall flowchart of a document structure based clustering method with improved reliability according to a preferred embodiment of the present invention.

Claims (18)

In the document structure based clustering device with improved reliability, The input documents are classified by subject, the keywords reflecting the core contents of the classified documents are extracted, the importance of the extracted keywords is calculated, the document cluster is executed using the extracted keywords and the calculated importance, and the results of the document cluster are displayed. A document cluster unit 100 that supplements using a cooperative cluster; And An input / output unit 200 which receives document information that is a target of document cluster, outputs a result of document cluster, and receives error information of the output document cluster result; Document structure based clustering device that includes improved reliability. The method of claim 1, The document group unit 100, Keyword extracting means (110) for classifying the input documents into subjects and extracting keywords from the classified documents; Importance calculation means 120 for calculating the importance of the extracted keyword based on the content of the main text of the document; Similarity calculating means (130) for calculating the similarity between the documents using the extracted keyword and the calculated importance; Cluster number determining means (140) for determining a suitable number of document clusters using the similarity calculation result and generating a document cluster result according to the determined number of clusters; And Post-processing means 150 for correcting errors of the cluster result by the cluster number setting means 140 by the cooperative cluster; Document structure-based clustering device to improve the reliability, characterized in that it comprises a. The method of claim 2, And the calculation of the similarity is performed by cosine measurement. The method of claim 1, And the input documents are news articles or newspaper articles. The method of claim 4, wherein The document clustering unit (100) is a document structure-based clustering device with improved reliability, characterized in that for extracting a keyword from the title, subtitle and lead of the news article or newspaper article. The method of claim 1, The importance of the keyword, Improved reliability based on document structure, which is calculated using one or more variables among paragraph position including keyword, frequency of occurrence of keyword, length of paragraph including keyword, and sentence length including keyword. Device. The method of claim 1, Importance of the keyword is a document structure based clustering apparatus improved reliability, characterized in that calculated by the following equation.

Where v is the importance value of the keyword, La (i) is the length of the i-th document of the document set, and Lp (i, j) is the length of the j-th paragraph in the i-th document of the document set, Ls (i , k) is the length of the kth sentence in the i-th document of the document set, d _Pj is the diminishing factor of the j-th paragraph in the i-th document of the document set, and d _Sk is the i-th of the document set Refers to the deminishing factor of the kth sentence in the document.) The method of claim 1, And the document clustering is performed using a Hierarchical Agglomerative Clustering method. The method of claim 8, A document structure based clustering device with improved reliability, characterized in that it automatically determines the number of suitable document clusters using Elbow Criterion. The method of claim 1, The input / output unit 200 includes a tagging interface and a drag and drop interface, any one or more of the document structure-based clustering device with improved reliability. In the document structure based clustering method with improved reliability, (a) receiving document information targeted for document clustering; (b) extracting keywords from the input documents; (c) calculating the importance of the extracted keyword; (d) calculating similarity between documents using the extracted keyword and the calculated importance; (e) generating document clustering results using the similarity calculation results; (f) outputting the generated document clustering result; And (g) complementing the printed document clustering result by the cooperative clustering; Document structure based clustering method improved reliability including a. The method of claim 11, In step (b), (b-1) classifying the received documents by subject; Document structure based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 11, In step (b), (b-2) removing the stopwords from the received documents; And (b-3) extracting roots from documents from which stopwords have been removed; Document structure based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 11, In step (c), (c-1) calculating a diminishing factor to consider the position of the paragraph or sentence containing the keyword in the document; (c-2) calculating a ratio of the length of the paragraph or sentence including the keyword to the total document length; And (c-3) calculating a sum of importance for each occurrence of each keyword in the document; Document structure based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 11, Step (g) is (g-1) receiving a keyword of a document from a user through a tagging interface; Document structure based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 15, In step (g-1), (g-1-1) searching for documents having a word set that does not include a keyword entered through the tagging interface; (g-1-2) adding a keyword input through the tagging interface to the searched word set; And (g-1-3) calculating importance of the added keyword; Document structure based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 11, Step (g) is (g-2) receiving a movement signal of a document determined to be a wrongly clustered document by a user through a drag and drop interface; Document structure-based clustering method to improve the reliability, characterized in that it comprises a. The method of claim 17, Step (g-2) is, (g-2-1) classifying the words included in the moved document into words that overlap with the word set of the cluster after the movement and words that overlap with the word set of the cluster before the movement; And (g-2-2) assigning different weights to words that overlap with the word set of the cluster after the movement and words that overlap with the word set of the cluster before the movement; Document structure based clustering method to improve the reliability, characterized in that it comprises a.

Images