JP6173848B2 - Document classification device - Google Patents

Description

  Embodiments described herein relate generally to a document classification apparatus.

  In recent years, with the increase in performance of computers, the increase in capacity of storage media, and the spread of computer networks, it is possible to store and manage a large amount of electronic documents in a computer system. The document here refers to, for example, business documents such as forms, planning documents, design documents, minutes, manuals, patents, technical literature, laws, regulations, news articles, e-mails, web pages, books, and the like. If only a large number of documents are stored in a computer file system or database without being organized, it becomes impossible to know where and what information is present, which causes a problem that information cannot be used.

  In order to deal with such problems, it has been attempted to promote effective utilization and sharing of information by classifying and organizing documents according to contents and uses. For example, analyze and investigate reports such as daily and weekly reports that are created and accumulated daily, inquiries sent from customers, defect information on products, intellectual property such as patents and technical literature, etc. In order to grasp the tendency of contents or to obtain new knowledge, grouping of documents having similar contents is performed. Since it takes labor to perform such document classification work manually, a document classification technique for automatically classifying documents has been developed.

As an example of document classification technology, a user automatically prepares a group (ie, cluster) of documents that have something in common or similar to each other from a set of document data to be classified without preparing a rule or category as a classification reference in advance. There is a clustering to generate automatically. There are two typical methods for this clustering.
In the first method of clustering, the features of each document are represented by feature vectors, and the similarity between documents is defined by the similarity (for example, inner product or cosine) between feature vectors of each document. Then, a cluster is generated as a group of documents having a high degree of similarity defined (see, for example, Patent Document 1). This method is called document clustering. Document features are often expressed using the frequency of words that appear in the text of the document, and the feature vectors in this case are called word vectors.

  The second method of clustering is a method that focuses on words appearing in a document. In this method, first, based on the number of documents in which each word appears (that is, the appearance frequency) and the number of documents in which a plurality of words appear in common (ie, the co-occurrence frequency) in the document data set to be classified, Then, important words that well represent the contents of the document, relations between words, or groups of words having similar appearance tendencies are extracted. Based on the extraction result, a group of documents is automatically generated by associating each word with the document in which it appears (see, for example, Patent Document 2). This method is called word clustering.

  Further, there is a biaxial map (also referred to as cross tabulation) as an analysis method for grasping a tendency of contents or obtaining new knowledge using a classification structure. According to this method, two classification axes are selected, a product set of documents (that is, a document data set classified in both categories) is obtained for each category as an item of each classification axis, and the number of documents is calculated. Display in a matrix. As a result, an overall image of the document data set can be grasped, and knowledge about the correlation of each category can be obtained. As a conventional technique related to a biaxial map, clustering is performed for each item included in a document (for example, “summary” or “claim” for a patent document), and a biaxial map is performed using the clustering result ( For example, refer to Patent Document 3), and there is one in which a user can select an arbitrary part on a classification hierarchy created by different viewpoints and classification methods and perform a biaxial map (for example, refer to Patent Document 4). .

JP 7-36897 A JP 2000-231560 A JP 2007-108867 A JP 2004-86350 A

In word clustering and document clustering as in Patent Literature 1 and Patent Literature 2, a classification structure is generated for a certain document set based on word significance and content similarity. The classification axis is not considered. Therefore, even if a two-axis map is configured using the clustering result as in Patent Document 3, it is not always a classification map suitable for the user's purpose. For example, when a user wants to see a two-axis map of a collection of patent documents with the horizontal axis as “by filing year”, the user needs to grasp the time-series application trend and analyze the technology trends. it is conceivable that. Further, when “By Applicant” is set on the horizontal axis, it is considered that there is a demand for the user to analyze technical trends such as strengths and weaknesses of each company. For such needs, it is necessary to perform clustering in consideration of the classification axis in the biaxial map according to each case, but this cannot be done with the prior art.
And in patent document 3 and patent document 4, it is the process of one direction of displaying a biaxial map using a clustering result and the classification structure at that time (however, patent document 4 can narrow down a classification axis. ). For this reason, although there is a need to generate a classification structure in accordance with the purpose of analysis and the user's recognition on the biaxial map, this cannot be done with the prior art.

  The problem to be solved by the present invention is to provide a document classification device that generates a classification structure suitable for the user's viewpoint and enables classification and analysis suitable for the user's purpose.

  A document classification device according to an embodiment includes a document storage unit, a category storage unit, a category operation unit, a feature calculation unit, a trend vector generation unit, a clustering unit, a category generation processing unit, and a biaxial map display unit It comprises. The document storage unit stores document data. The category storage unit stores a hierarchical structure of categories and classification rules for classifying document data into categories. The category operation unit receives an input of a category to be classified and a target category that is a category to be classified, and a set of axis categories that are subordinate categories of the category to be classified from the category storage unit. Read as. The feature degree calculation unit calculates a feature degree of a word included in the target document data set by setting a set of document data satisfying the classification rule of the target category among the document data stored in the document storage unit as the target document data set. To do. The trend vector generation unit selects a word representing the feature of the target document data set based on the feature degree calculated by the feature degree calculation unit, and sets a classification rule for each axis category in the axis category set for each selected word. A statistic based on the appearance frequency of the word in the target document data to be satisfied is calculated, and a trend vector in which the statistic is set as a value of an element corresponding to the axis category is generated. The clustering unit clusters words based on the similarity of the trend vectors generated by the trend vector generation unit. The category generation processing unit generates, for each cluster obtained as a result of clustering by the clustering unit, a feature word category having a classification rule using a target category as a higher category and a word belonging to the cluster as a filter word, and stores the category Register with the department. The biaxial map display unit uses each category included in the axis category set as a classification item for the first axis, and each cell of the biaxial map uses the feature word category generated by the category generation processing unit as the classification item for the second axis. Display information indicating the number of document data satisfying the classification rule of the axis category corresponding to the cell and the classification rule of the feature word category corresponding to the cell among the document data stored in the document storage unit .

It is a block diagram which shows the structure of the document classification device which concerns on embodiment. It is a figure which shows the example of the document data memorize | stored in the document memory | storage part which concerns on embodiment. It is a figure which shows the example of the category data memorize | stored in the category memory | storage part which concerns on embodiment. It is a figure which shows the example of the feature word category data memorize | stored in the category memory | storage part which concerns on embodiment. It is a figure which shows the example of the feature-value data memorize | stored in the feature-value data storage part which concerns on embodiment. It is a figure which shows the example of the attention word list data which the attention word setting part which concerns on embodiment memorize | stores inside. It is a flowchart which shows the flow of the document classification process which concerns on embodiment. It is a flowchart which shows the flow of the process which displays the initial biaxial map which concerns on embodiment. It is a flowchart which shows the flow of a process with respect to the category operation on the biaxial map which concerns on embodiment. It is a flowchart which shows the flow of the process which calculates the characteristic degree which concerns on embodiment. It is a flowchart which shows the flow of the process which sets the attention word which concerns on embodiment. It is a flowchart which shows the flow of the process which calculates | requires the correction | amendment feature degree which concerns on embodiment. It is a flowchart which shows the flow of the process which acquires the characteristic data which concern on embodiment. It is a flowchart which shows the flow of the process which calculates | requires the tendency vector which concerns on embodiment. It is a flowchart which shows the flow of the process of the feature word clustering which concerns on embodiment. It is a flowchart which shows the flow of the process which produces | generates the feature word category which concerns on embodiment. It is a flowchart which shows the flow of the process which displays the biaxial map which concerns on embodiment. It is a flowchart which shows the flow of a process of addition or deletion of the filter word of the feature word category which concerns on embodiment. It is a figure which shows the example of the category structure memorize | stored as an initial state of the category memory | storage part which concerns on embodiment. It is a figure which shows the example of a display of the execution screen of the biaxial map display which concerns on embodiment. It is a figure which shows the example of a display of the initial table of the biaxial map which concerns on embodiment. It is a figure which shows the example of an initial display of the biaxial map which concerns on embodiment. It is a figure which shows the example of a display of the execution screen of the feature word clustering based on embodiment, and the setting screen of an attention word. It is a figure which shows the example of a display of the category structure after the characteristic word clustering execution which concerns on embodiment. It is a figure which shows the example of a display of the biaxial map after execution of the feature word clustering which concerns on embodiment. It is a figure which shows the example of a display of the biaxial map at the time of filter word selection which concerns on embodiment. It is a figure which shows the edit operation of the feature word category in the biaxial map which concerns on embodiment, and the example of a display of the screen. It is a figure which shows the example of a display of the biaxial map after edit operation of the feature word category which concerns on embodiment. It is a figure which shows the example of a display when the biaxial map which concerns on embodiment is expressed with the line graph.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration of a document classification apparatus 100 according to an embodiment of the present invention. As shown in FIG. 1, the document classification apparatus 100 includes a document storage unit 1, a category storage unit 2, a document classification unit 3, a feature data storage unit 4, and a user interface unit 5.

  The document storage unit 1 stores document data to be subjected to document classification processing by the document classification device 100. The document storage unit 1 is realized by, for example, a file system or a document database. Alternatively, the document storage unit 1 may be configured by a plurality of storage units connected by a computer network.

The category storage unit 2 stores category data used for document data classification. The category data indicates a category name, a category hierarchy, and a category classification rule. The category hierarchy represents the upper and lower relationships between categories. The classification rule is a rule used when classifying document data into categories. For example, bibliographic information such as document data attributes, creation date, creator, and genre is used. Alternatively, the document data may be classified into categories by existing clustering.
Furthermore, the category storage unit 2 also stores feature word category data generated by the document classification unit 3 described later. The feature word category is a category in which feature words having similar appearance tendencies with respect to the axis category set, which is a category serving as a viewpoint when classifying documents, are classified as classification rules.

FIG. 19 to be described later shows a screen display example showing an overall image of the hierarchical structure of the categories stored as the initial state of the category storage unit 2, and FIG. 3 to be described later shows the categories stored in the category storage unit 2. Data are shown.
In FIG. 19, the user interface unit 5 (specifically, a biaxial map display unit 51 described later) displays the category name 303 “root” (FIG. 3) of the category data 300 in which the upper category is not set as the root hierarchy. "Root" category 1600. In addition, the user interface unit 5 (two-axis map display unit 51 to be described later) displays “category by applicant” set in the category names 313, 323, and 333 of the category data 310, 320, and 330 having the category data 300 as a higher category. ], “By application year”, and “by content” (FIG. 3) are subordinate to the “root” category 1600, “by applicant” category 1601, “by application year” category 1602, and “by content” category 1603. Display as. The following description will be given on the premise that such a whole image is displayed.

The document classification unit 3 shown in FIG. 1 receives an axis category set and a target category as inputs. The axis category set is a set of child categories (axis categories) of a category that is one axis in the 2-axis map, and as described above, is a category set as a viewpoint in classification. The child category is a category one level lower than a certain category, and the number of child categories (axis categories) included in the axis category set is one or more. The target category is a category that is the other axis in the two-axis map, and is a category that is a classification target for a category that is a viewpoint of document classification. That is, in the present embodiment, the axis category set is used as a viewpoint in classification, and the document data included in the target category is classified based on the axis category set. Here, the document classification unit 3 classifies the document data classified into the axis category using the target category. For example, if you want to classify a document data set classified in the “by content” category from a perspective by age, specify the “by content” category as the target category, and specify the “by age” category as the category of perspective. . At this time, the axis category set includes child sets of “2004”, “2005”, “2006”, “2007”, and “2008”, which are child categories of the “by age” category (the categories in FIG. 19). 1621-1625). When this content is represented by a two-axis map, one axis is in the “by age” category and the other axis is in the “by content” category.
In order to determine a feature word used for this classification, the document classification unit 3 calculates a feature degree for words appearing in the document data set (target document data set) classified into the input target category. The feature degree is an index value that quantitatively indicates how much the word represents the feature of the target document data set. Note that the number of document data included in the target document data set is one or more. Further, the document classification unit 3 generates a feature word category based on a group of feature words whose appearance tendency is similar to the axis category set as a child category of the target category.

The document classification unit 3 includes a feature calculation unit 31, a feature correction unit 32, and a feature word category generation unit 33.
The feature degree calculation unit 31 calculates a feature degree based on the statistical significance of the appearance frequency in a predetermined document data set for words appearing in the target document data set that is a set of document data classified into the target category, It is stored in the feature data storage unit 4.
The feature correction unit 32 receives, as an input, a focused word set obtained by collecting a plurality of focused words selected by the user from a focused word setting unit 53 of the user interface unit 5 described later. The word of interest is a word selected as a target of attention by the user when generating a feature word category as a child category of the target category. The feature correction unit 32 corrects the feature of each word stored in the feature data storage unit 4 based on the co-occurrence of the word and the word set of interest.

  The feature word category generation unit 33 generates a feature word category based on a group of feature words similar in appearance tendency to the axis category set, and stores the generated feature word category data in the category storage unit 2. The feature word category generation unit 33 includes a trend vector generation unit 34, a clustering unit 35, and a category generation processing unit 36. The trend vector generation unit 34 sets a set of words whose feature degree stored in the feature degree data storage unit 4 is greater than a predetermined value as a feature word set. The trend vector generation unit 34 calculates an appearance frequency in each axis category in the target document data set for each word included in the feature word set, and calculates a statistic based on the calculated appearance frequency in each axis category. A tendency vector as an element value corresponding to the category is generated and stored in the feature data storage unit 4. The clustering unit 35 performs clustering based on the similarity of the generated trend vectors, and extracts feature word clusters that are groups of closely related words. The category generation processing unit 36 generates a feature word category that uses a feature word that is a word included in the feature word cluster as a filter word and uses the filter word as a classification condition. The category generation processing unit 36 stores the generated feature word category data in the category storage unit 2.

  The feature data storage 4 is means for storing feature data for each word included in the target document data set. The feature degree data includes the feature degree generated by the document classification unit 3 and a tendency vector for the axis category set.

  The user interface unit 5 includes two categories, a category that is the horizontal axis of the biaxial map (hereinafter referred to as “horizontal axis category”) and a category that is the vertical axis (hereinafter referred to as “vertical axis category”). Receives the input and presents the biaxial map to the user. In the actual presentation, a general-purpose device such as a browser that communicates via the Internet and is displayed on a display of a personal computer (PC) is often used. In addition, the user interface unit 5 accepts an execution request for feature word clustering, a setting of a focused word set, and an operation for editing a feature word category from the user. The user interface unit 5 is realized by, for example, a graphical user interface (hereinafter referred to as “GUI”).

The user interface unit 5 includes a biaxial map display unit 51, a category operation unit 52, and a focused word setting unit 53.
The biaxial map display unit 51 receives input of two categories of the horizontal axis category and the vertical axis category, and displays a biaxial map with the lower categories of the categories as items of rows and columns, respectively, on a display of a personal computer or the like. (For example, FIG. 22 and FIG. 25 described later). When the biaxial map display unit 51 displays the biaxial map through the display of a personal computer (PC), each cell of the biaxial map includes a row item category and a column item category corresponding to the cell. A graph corresponding to the number of document data classified into both categories is displayed. Further, the biaxial map display unit 51 displays the filter word of the feature word category in the cell of the title row of the feature word category (for example, the cell 2101 in FIG. 25 described later) on the biaxial map. In addition, when the 2-axis map display unit 51 receives the input of the filter word selected by the user on the 2-axis map, the subordinate category of the horizontal axis category and the filter word are selected for the row in which the filter word is selected. The number of document data including the selected filter word in the document data set classified into both categories of the feature word category in the row is displayed separately from the above graph (for example, FIG. 26 described later).

  The category operation unit 52 receives a feature word clustering execution request from the user on the biaxial map, and outputs this execution request to the document classification unit 3. Further, the category operation unit 52 receives the selection of the feature word category by the user on the two-axis map, and displays the feature words of the selected feature word category (for example, the feature word addition screen 2310 in FIG. 27 described later). Further, the category operation unit 52 accepts an editing operation such as addition / deletion of a filter word to / from the feature word category from the user, and updates the data of the feature word category stored in the category storage unit 2 according to the editing operation. .

  The focused word setting unit 53 accepts a plurality of words focused on for classification as a focused word set from the user and outputs them to the document classifying unit 3. The document classification unit 3 receives the input of the target word set from the target word setting unit 53, corrects the feature level of the feature level data of each word stored in the feature level data storage unit 4, and corrects the corrected feature. A feature word category is generated using a word selected based on the degree as a feature word. As a result, the document classification unit 3 can make the generated feature word category and the category feature word presented by the category operation unit 52 in accordance with the word focused by the user.

  The document classification device 100 is configured as described above, and is realized by, for example, a personal computer (PC). The PC includes, for example, a central processing unit (CPU), a memory, a hard disk drive (HDD), a liquid crystal display, a keyboard, a mouse, and the like.

FIG. 2 is a diagram illustrating an example of document data stored in the document storage unit 1. The document storage unit 1 stores a plurality of document data. Each document data includes data of a document number 201 which is a unique identifier as shown in document data 200a in FIG.
Further, the document data 200a includes attribute data according to the purpose and format. For example, FIG. 2A shows an example of a document describing patent information, and includes attribute data such as an application date 203 and an applicant 204. The document data 200a includes text such as a document name 202 and a body 205 as text data of the document, that is, data described in a natural language such as Japanese or English. In addition, the document data 200a holds data of a word vector 206 whose element is the number of appearances (appearance frequency tf) of each word included in the document data 200a. The word vector 206 is calculated by the document classification unit 3. Note that the word vector 206 may be stored in a database different from the document database in which the original document data is stored in a form corresponding to the document data 200a.

  The document data 200b shown in FIG. 2B is an example when the document data 200a shown in FIG. 2A is described in the XML (extensible markup language) format. In this case, in the document data 200b, data corresponding to the document number, each attribute, each text, and the word vector in the document data 200a is described using XML elements (tags) and attributes.

FIG. 3 is a diagram illustrating an example of category data stored in the category storage unit 2. Note that FIG. 19 described later shows an example of a hierarchical structure of categories.
In FIG. 3, six examples of category data 300, 310, 320, 330, 340, 350 are shown. Each category data includes data of category numbers 301, 311, 321, 331, 341, and 351, which are unique identifiers. Here, categories are created for each applicant and each application year based on the applicant 204 and the application date 203 indicating the attributes of the document data 200a (or document data 200b) stored in the document storage unit 1, respectively.

  In the document classification device 100 of the present embodiment, a hierarchical structure in a tree format is configured by a plurality of categories (for example, FIG. 19 described later). Therefore, each category data has data of upper categories 302, 312, 322, 332, 342, and 352 as data representing the relationship between the upper and lower categories. However, since the category data 300 represents a category corresponding to the root of the hierarchical structure, “(none)” is set in the upper category 302. The upper category of the “by applicant” category is the “root” category. Therefore, the set value “C000” of the category number 301 of the category data 300 corresponding to the “root” category is set in the upper category 312 of the category data 310 corresponding to the “by applicant” category. Each category data 300, 310, 320, 330, 340, 350 includes data of category names 303, 313, 323, 333, 343, 353, respectively.

  Further, the category classification criteria are described in classification rules 304, 314, 324, 334, 344, and 354. For example, since the category data 300 of the “root” category represents the category of the root of the hierarchical structure, the classification rule 304 that is the classification criterion is “true”. The classification rule 304 of “true (true)” indicates that the rule that all document data satisfies the condition is set. For example, the classification rule 344 of the category data 340 is “./Applicant=“Company A” ”. This classification rule 344 is a rule on condition that the applicant 204 having the attribute of the document data 200 a stored in the document storage unit 1 is “Company A”.

  The classification rules 304, 314, 324, 334, 344, and 354 are examples in which a description of a condition for the document data 200b described in the XML format, that is, an inquiry called XQuery (or XPath) is described using a language. is there. The description format of the classification rules 304, 314, 324, 334, 344, and 354 may be determined according to the method for realizing the document storage unit 1 and the format of the document data. For example, SQL may be used.

  The classification rule 354 of the category data 350 is also an example described using the XQuery conditional expression, and is a rule on condition that the “application date” attribute of the document data is “2008”. This classification rule acts as an AND condition for the category hierarchy. For example, although different from the example illustrated in FIG. 3, it is assumed that the “2008” category indicated in the category data 350 exists as a lower category of the “Company A” category indicated in the category data 340. In this case, the document data classified into the “2008” category includes a classification rule “./applicant=“Company A” ”of the“ Company A ”category and a classification rule of the“ 2008 ”category which is a lower category thereof. AND condition “(./applicant=“Company A”) ”and (./application) with“ ./application date> = ”2008/01/01” and ./application date <= “2008/12/31” ” Date> = “2008/01/01” and ./application date <= “2008/12/31)”, that is, the applicant is “Company A” and the application year is “2008”. Document data of “Year”.

FIG. 4 is an example of feature word category data that is category data of the feature word category generated by the document classification unit 3.
The feature word category data is stored in the category storage unit 2 in the same manner as the other category data shown in FIG. FIG. 4 shows two examples of feature word category data 400 and 410. The feature word category data 400 and 410 are data of category numbers 401 and 411, upper categories 402 and 412, category names 403 and 413, classification rules 404 and 414, respectively, similarly to other category data as shown in FIG. including. Further, the feature word category data 400 and 410 includes data of filter words 405 and 415 indicating feature words included in the feature word cluster extracted by the category generation processing unit 36.

  The classification rules 404 and 414 included in the feature word category data are generated by the category generation processing unit 36 based on the filter words 405 and 415. For example, the category generation processing unit 36 generates a condition indicating that filter words 405 and 415 are included in the text information of the document data as a classification rule. “Classification”, “knowledge”, and “shared” are set in the filter word 405 of the feature word category data 400 shown in FIG. Accordingly, the classification rule 404 of the feature word category data 400 includes a condition that these filter words 405 are included in the body 205 that is text information of the document data 200a, that is, “contains (./body,“ analysis ”) and “contains (./text,“ knowledge ”)” and “contains (./text,“ share ”)” are set.

  FIG. 5 is an example of feature data generated by the document classification unit 3 and stored in the feature data storage unit 4. The feature data 500-1, 500-2,... Shown in the figure correspond to each word extracted from the target document data set in the document classification unit 3, and include a word 510, a document frequency 501, and a feature 502. , Correction feature 503 and frequency vector / trend vector 504 data for the axis category (category number). The feature data 500-1, 500-2,... Are collectively referred to as the feature data 500.

  The document frequency 501 indicates the number of documents (document frequency df) that is the number of document data in which the word 510 appears in the target document data set. The feature degree 502 indicates the feature degree calculated by the feature degree calculation unit 31. The corrected feature 503 indicates a value obtained by correcting the feature by the feature correcting unit 32 based on the focused word set designated by the user. The frequency vector / trend vector 504 for the axis category (category number) includes data of the frequency vector 511 and the trend vector 512. Let the frequency vector 511 and the trend vector 512 of the feature data 500-i (i = 1, 2,...) Be a frequency vector 511-i and a trend vector 512-i, respectively. The frequency vector 511 uses the number of documents (cf), which is the number of document data in which the word 510 appears, in the common set of the target document data set and each axis category in the axis category set as the value of the element (component). The trend vector 512 uses the appearance ratio (cp) of the word 510 in the common set with respect to the target document data set as the value of each element. However, the trend vector generation unit 34 does not obtain a frequency vector and a trend vector because words whose correction feature degrees are equal to or less than a predetermined value are excluded from the feature word cluster. FIG. 5 is an example in which the words 510 “search” and “mail” are excluded from the feature word cluster, and the frequency vectors 511-3 of the feature data 500-3 and 500-5 of these words, 511-5 and the trend vectors 512-3 and 512-5 are empty.

  FIG. 6 is a diagram illustrating an example of attention word list data stored in the attention word setting unit 53. In the target word candidate 601 of the target word list data 600 shown in FIG. 6, a filter word selected by the user as the target word on the two-axis map or a list of character strings input by the user as the target word is set. The attention word candidate 601 is given a attention word setting 602 that is a flag indicating whether or not the document classification unit 3 actually applies the attention word as a attention word. In this embodiment, if the attention word setting 602 is “1”, it is applied as a attention word, and if it is “0”, it indicates that it is not applicable.

Hereinafter, an example of document classification processing performed by the document classification device according to the embodiment of the present invention will be described with reference to FIGS.
FIG. 7 is a flowchart showing an example of the flow of document classification processing performed by the document classification device 100 of this embodiment. First, the biaxial map display unit 51 of the user interface unit 5 receives an input of the horizontal axis category and the vertical axis category of the biaxial map by a user operation, and performs an initial biaxial map display process (step S1). In the initial biaxial map display process, the biaxial map display unit 51 acquires a set of child categories of the horizontal axis category and a set of child categories of the vertical axis category from the category storage unit 2. The 2-axis map display unit 51 uses the horizontal axis category child categories (hereinafter referred to as “horizontal axis category”) as the horizontal axis items, and the vertical axis category child categories (hereinafter referred to as “vertical axis category”). 2 axis map is displayed with each item on the vertical axis. When there is no child category in the horizontal axis category, the horizontal axis category is the horizontal axis item, and when there is no child category in the vertical axis category, the vertical axis category is the vertical axis item (for example, a diagram to be described later). 22).

  Subsequently, the category operation unit 52 receives a clustering request and a target category input from the user (step S2). For example, the user selects a target category from the horizontal axis category or vertical axis category of the biaxial map displayed in step S1. The category operation unit 52 outputs the input target category and axis category set to the document classification unit 3. The axis category set is a set of subordinate categories of the horizontal axis category not selected as the target category or a set of subordinate categories of the vertical axis category. The feature calculation unit 31 of the document classification unit 3 extracts a word that has a predetermined part of speech and is not an unnecessary word from the target document data set that is a set of document data classified into the target category, and calculates the feature level. Calculate and write to the feature data storage unit 4 (step S3). The category operation unit 52 receives an input of a word of interest from the user (for example, a word of interest setting form 1910 in FIG. 23 described later), and outputs it to the document classification unit 3 (step S4). The feature correction unit 32 of the document classification unit 3 determines the feature of each word stored in the feature data storage unit 4 in step S3 based on the co-occurrence of the word and the target word set in the target document data set. (Step S5).

  The trend vector generation unit 34 generates a trend vector for words whose corrected feature values are larger than a predetermined value, and stores them in the feature data storage unit 4 (step S6). The clustering unit 35 performs clustering of words based on the similarity of tendency vectors, and extracts feature word clusters that are groups of closely related words (step S7). The category generation processing unit 36 generates a feature word category that uses a feature word that is a word included in the feature word cluster as a filter word and uses the filter word as a classification condition. The category generation processing unit 36 stores the generated feature word category data (for example, the feature word category data 400 and 410 in FIG. 4) in the category storage unit 2 (step S8). Thereby, the category structure is updated (for example, FIG. 24 described later).

  The biaxial map display unit 51 displays a biaxial map having the axis category set and the feature word category set generated in step S8 as two axes (step S9). At this time, the biaxial map display unit 51 displays the filter words of each feature word category (for example, FIG. 25 described later).

  The category operation unit 52 accepts editing operations such as addition or deletion of filter words from the user (for example, FIG. 27 described later), and updates the feature word category data stored in the category storage unit 2 (step S10). . The biaxial map display unit 51 displays a biaxial map (for example, FIG. 28 to be described later) having two axes, the axis category set and the feature word category set based on the feature word category data updated in step S10 ( Step S11).

Next, detailed processing in each processing step of FIG. 7 will be described.
FIG. 8 is a flowchart showing the flow of processing for displaying the initial biaxial map. The flowchart shown in the figure shows an example of detailed processing of the initial biaxial map display processing in step S1 of FIG.

  First, the biaxial map display unit 51 receives an input of a horizontal axis category xAxisCat and a vertical axis category yAxisCat of the biaxial map by the user (step S1001). For this input, in this embodiment, the biaxial map display unit 51 displays a screen representing the category structure based on the category data and feature word category data stored in the category storage unit 2. Here, it is assumed that the feature word category data 400 and 410 shown in FIG. 4 has not been generated yet.

  FIG. 19 is a diagram illustrating a screen display example of the category structure stored as the initial state of the category storage unit 2. The biaxial map display unit 51 displays the category name 303 “root” of the category data 300 in which the upper category is not set as the “root” category 1600 of the root hierarchy. Further, the biaxial map display unit 51 identifies the category data 310, 320, 330 in which the category number 301 “C000” of the category data 300 is set as the upper category. The biaxial map display unit 51 sets “by applicant”, “by application year”, and “by content” set to the category names 313, 323, and 333 of the category data 310, 320, and 330 to “route”. It is displayed as “by applicant” category 1601, “by application year” category 1602, and “by content” category 1603, which are lower layers of the category 1600.

  Further, the biaxial map display unit 51 identifies category data in which the category number 301 “C0001” of the category data 310 is set as the upper category. The biaxial map display unit 51 displays the category names “Company A”, “Company B”, “Company C”, “Company D”, and “Company E” indicated by the identified category data in the “by applicant” category 1601. The lower hierarchy “Company A” category 1611, “Company B” category 1612, “Company C” category 1613, “Company D” category 1614, and “Company E” category 1615 are displayed.

Similarly, the biaxial map display unit 51 identifies category data in which the category number 321 “C0002” of the category data 320 is set as the upper category. The biaxial map display unit 51 displays the category names “2004”, “2005”, “2006”, “2007”, “2008” indicated by the identified category data in the “by application year” category 1602. The lower layer “2004” category 1621, “2005” category 1622, “2006” category 1623, “2007” category 1624, and “2008” category 1625 are displayed. The “Company A” category 1611 and the “2008” category 1625 correspond to the category data 340 and 350 shown in FIG. 3, respectively.
Since there is no category data having the category number 331 “C0003” of the category data 330 as the upper category, the biaxial map display unit 51 does not display the lower category in the “by contents” category 1603.

  Further, the biaxial map display unit 51 reads out classification rules from the category data (or feature word category data) of each category stored in the category storage unit 2. The biaxial map display unit 51 counts the number of document data classified into each category using the read classification rule, and displays the count number. Note that the biaxial map display unit 51 may store the category data (or feature word category data) of each category in association with the document number of the document data classified into the category.

Next, an example of a method for inputting the horizontal axis category xAxisCat and the vertical axis category yAxisCat will be described using the display of FIG.
FIG. 20 is a diagram illustrating a display example of the execution screen of the biaxial map display, and illustrates a display example when the horizontal axis category xAxisCat and the vertical axis category yAxisCat are input using the display of FIG. First, the user selects two categories for the horizontal axis and the vertical axis in the biaxial map for the category structure displayed by the biaxial map display unit 51. Here, the user selects the “by application year” category 1602 and the “by content” category 1603 in FIG.

  Upon receiving the input of the two selected categories, the biaxial map display unit 51 displays a screen 1710 for selecting which of the two categories is the horizontal axis category xAxisCat that is a classification viewpoint. The user selects a category radio button 1711 to be classified and presses an execute button 1712. Accordingly, the biaxial map display unit 51 receives an input of the “by application year” category 1602 selected by the user as the horizontal axis category xAxisCat. The vertical axis category yAxisCat becomes the “content-specific” category 1603 that is not selected by the user.

  As described above, the biaxial map display unit 51 receives input of the horizontal axis category xAxisCat and the vertical axis category yAxisCat selected by the user. In this embodiment, an example in which a GUI as shown in FIGS. 19 and 20 is used as an input method to the biaxial map has been shown. However, the category storage unit 2 stores the biaxial map in order to display the biaxial map. It is only necessary to input which of the categories indicated by the category data or the feature word category data is the horizontal axis category xAxisCat and the vertical axis category yAxisCat. Therefore, not only the use of the GUI but also the input from the command line of the computer system may be used.

  In FIG. 8, the biaxial map display unit 51 includes a horizontal axis category set xCats, which is a set of child categories of the horizontal axis category xAxisCat, based on the category data and feature word category data stored in the category storage unit 2. The vertical axis child category set yCats, which is a set of child categories of the vertical axis category yAxisCat, is acquired (step S1002). A child category of the horizontal axis category xAxisCat is a horizontal axis child category xCat, and a child category of the vertical axis category yAxisCat is a vertical axis child category yCat. The horizontal axis category xCat corresponds to category data or feature word category data in which the category number of the horizontal axis category xAxisCat or the category number of the feature word category data is set as the upper category. Similarly, the vertical axis child category yCat corresponds to the category data or the feature word category data in which the category number of the vertical axis category yAxisCat or the category number of the feature word category data is set as the upper category.

  In the case of the example illustrated in FIG. 20, the biaxial map display unit 51 sets the horizontal axis category set xCats as a set of child categories of the “by application year” category 1602 that is the horizontal axis category xAxisCat {{2004} category 1621, The “2005” category 1622, the “2006” category 1623, the “2007” category 1624, and the “2008” category 1625} are acquired. The horizontal axis category xCat corresponds to the category data in which the category number 321 of the category data 320 corresponding to the “by application year” category 1602 is set as the upper category. Further, since the “by contents” category 1603 which is the vertical axis category yAxisCat has no child category, the biaxial map display unit 51 acquires an empty set as the vertical axis child category set yCats. That is, category data or feature word category data in which the category number 331 of the category data 330 corresponding to the “by content” category 1603 is set as a higher category is not stored in the category storage unit 2.

  The biaxial map display unit 51 sets the vertical axis category yAxisCat and each vertical axis category yCat included in the vertical axis category set yCats as rows, and sets each horizontal axis category xCat included in the horizontal axis category set xCats. A biaxial map initial table is created and displayed as a column (step S1003).

  FIG. 21 shows the biaxial map initial table 1800 created in step S1003. Since the biaxial map initial table 1800 also includes title rows and title columns for displaying categories, the number of rows is (1 + vertical category number + vertical child category number) and the number of columns is (1 + horizontal child category number). It is a table. In this embodiment, since there is no vertical axis child category, the number of rows is (1 + the number of vertical axis categories). Similarly, when there is no horizontal axis category, the number of columns is (1 + the number of horizontal axis categories). The first row (row including cell 1802) of the biaxial map initial table 1800 is a title row, and the first column (row including cell 1801) is a title column.

  In FIG. 8, the biaxial map display unit 51 selects all the cells (hereinafter, referred to as “cell”) in the created biaxial map initial table 1800 one by one, and steps S1005 to S1010 for the selected cell. The above process is repeated (step S1004-NO).

  First, the biaxial map display unit 51 determines whether or not the cell is the first row (first row) or the first column (first column) (step S1005). When it is determined that the cell is the first row (first row) or the first column (first column) (YES in step S1005), the biaxial map display unit 51 performs the processing from step S1006 to step S1008. In the processing from step S1006 to step S1008, the category name and filter word of the category corresponding to the cell are displayed with the first row or first column as the title row or title column of the table.

  That is, when the cell being processed is the first row or the first column, the biaxial map display unit 51 has a category cat corresponding to the cell (vertical axis category yAxisCat, vertical axis category yCat, or horizontal axis category xCat). The category name is displayed (step S1006). The biaxial map display unit 51 reads the category name from the category data or the feature word category data of the category corresponding to the cell. Furthermore, the biaxial map display unit 51 determines whether or not the category cat corresponding to the cell has a filter word set filter (step S1007). Specifically, the biaxial map display unit 51 determines whether or not the category cat corresponds to the feature word category data and whether or not the filter word set filters are included depending on whether or not the filter word is set. .

  When it is determined that the category cat has the filter word set filters (step S1007—YES), the biaxial map display unit 51 displays the filter words included in the filter word set filters on the cell (step S1008). The filter word set filters is a set of filter words set in the feature word category data of the category cat corresponding to the cell. When it is determined that the category cat does not have the filter word set filters (step S1007-NO), or after the process of step S1008, the biaxial map display unit 51 returns to step S1004 and selects an unselected cell. Repeat the process.

  When it is determined in step S1005 that the cell is neither the first row (first row) nor the first column (first column) (step S1005-NO), the biaxial map display unit 51 displays the vertical axis corresponding to the cell row. The number of documents dn which is the number of document data classified into both the category yAxisCat or the vertical axis category yCat and the horizontal axis category xCat corresponding to the cell column is obtained (step S1009).

  When the document data is XML, the document number dn can be obtained by using a product of conditional expressions in XQuery. For example, in the case of a cell 1803 in FIG. 22 described later, document data that satisfies the logical product of the classification rule of the corresponding vertical axis category yAxisCat = “by content” category and the classification rule of the horizontal axis category xCat = “2004” category. Count the number of

  The classification rule 334 set in the category data 330 of the “by content” category is “true”, and the classification rule 304 set in the upper category “root” category data 300 is “true”. Therefore, the classification rule of the vertical axis category yAxisCat “by content” category is “(true) and (true)”.

  On the other hand, the classification rule set in the category data of the “2004” category is “./application date> =” 2004/01/01 ”and ./application date <=“ 2004/12/31 ””, The classification rule 304 set in the category data 300 of the upper category “root” is “true”. Therefore, the classification rule of the horizontal axis category xCat “2004” category is “(true) and (./application date> =“ 2004/01/01 ”and ./application date <=” 2004/12/31 ”) It is.

  Accordingly, the biaxial map display unit 51 calculates the logical product “{(true) and (true)} of the classification rule of the vertical axis category yAxisCat“ by contents ”category and the classification rule of the horizontal axis category xCat“ 2004 ”category. and {(true) and (./application date> = “2004/01/01” and ./application date <= “2004/12/31”)} ”are counted, and the number of documents dn And The number of documents dn can be calculated by using the XQuery count () function.

  Next, the biaxial map display unit 51 displays a circle chart having a size corresponding to the number of documents dn calculated in step S1009 on the cell of the table created in step S1003 (step S1010). The biaxial map display unit 51 returns to step S1004, selects an unselected cell, and repeats the process.

  And the biaxial map display part 51 will complete | finish the process of FIG. 8, if the process of step S1005-step S1010 is complete | finished with respect to all the cells in the biaxial map initial table 1800 (step S1004-YES).

  FIG. 22 is a diagram illustrating an initial display example of the biaxial map displayed by the biaxial map display unit 51 at the end of the above processing. In this case, since the feature word category is not generated below the “by content” category, as shown in the figure, only the “by content” category is displayed in the first column of the vertical axis. It is displayed.

  FIG. 9 is a flowchart showing a flow of processing for category operations on the biaxial map executed by the category operation unit 52. By the processing shown in the figure, the category operation unit 52 controls the flow of processing in steps S2 to S11 in FIG. The user requests execution of the feature word clustering by inputting the target category via the category operation unit 52. For example, the category operation unit 52 receives a feature word clustering execution request from the user on the initial biaxial map displayed by the biaxial map display unit 51 by the initial biaxial map display process of FIG. The category operation unit 52 receives a selection of feature word categories generated by feature word clustering. The category operation unit 52 displays a feature word used as a filter word in the selected feature word category (for example, a feature word addition screen 2310 in FIG. 27 described later), and adds a filter word to the feature word category from the user. -Accept editing operations such as deletion. The category operation unit 52 updates the category data of the feature word category stored in the category storage unit 2 in accordance with the received editing operation.

  Therefore, when the category operation unit 52 receives an input of the category cat selected by the user in the biaxial map (step S1101-YES), the category operation unit 52 further receives an input of a feature word clustering execution request from the user (step S1102). ), Whether a filter word addition request is input (step S1105) or a filter word deletion request is input (step S1111). Hereinafter, the process of FIG. 9 will be described in detail.

First, the category operation unit 52 receives an input of a category cat selected by the user (step S1101-YES).
FIG. 23 is a diagram illustrating a display example of a feature word clustering execution request screen and a focused word setting screen. Here, the user has selected the clustering target category on the biaxial map initial table displayed by the biaxial map display unit 51 as shown in FIG. 22 in the initial biaxial map display processing. In the figure, the user selects a cell 1801 to select the “content-specific” category, which is the vertical axis category, as the target category. Thereby, the category operation unit 52 receives an input of the “by contents” category selected by the user as the category cat. Further, upon receiving an input of a feature word clustering execution request from the user, the category operation unit 52 displays a feature word clustering execution confirmation screen 1930.

  In FIG. 9, upon receiving an input indicating that the user has selected the “execute” button 1931 on the execution confirmation screen 1930 (step S1102—YES), the category operation unit 52 performs the process of step S1103. That is, the category operation unit 52 outputs the category cat input as the target category and the axis category set, which is a set of child categories of other axes not input as the target category in the 2-axis map, to the document classification unit 3. The execution of feature word clustering is instructed (step S1103). In this embodiment, since the vertical axis category is input as the target category, the category operation unit 52 outputs the horizontal axis category set xCats, which is a set of the horizontal axis categories xCat, to the document classification unit 3 as the axis category set. To do. The horizontal axis category set xCats is an axis category set as a viewpoint in classification. As shown in FIG. 23, when the user selects the “by content” category and selects the execution of feature word clustering, the category operation unit 52 outputs the “by content” category as the category cat to the document classification unit 3. , {“2004” category, “2005” category, “2006” category, “2007” category, “2008” category} are output as the horizontal axis category set xCats. The processing in steps S1101 to S1103 corresponds to the category operation processing in step S2 in FIG.

  The document classification unit 3 that has received the input of the category cat and the horizontal axis category set xCats in step S1103 executes the feature word clustering, and ends the processes from step S3 to step S8 in FIG. The category operation unit 52 outputs the horizontal axis category xAxisCat and the vertical axis category yAxisCat of the current biaxial map to the biaxial map display unit 51, and updates the display of the biaxial map (step S1104). For example, as shown in FIG. 23, when the user selects a cell 1801 and inputs an execution request for feature word clustering, the horizontal axis category xAxisCat is the “by application year” category, and the vertical axis category yAxisCat is “by content”. Category. Thereby, the result of the feature word clustering by the document classification unit 3 is reflected on the biaxial map. The process of step S1104 corresponds to the process of step S9 in FIG. The category operation unit 52 repeats the processing from step S1101.

  Then, the category operation unit 52 receives an input of the category cat selected by the user (step S1101-YES), and further receives an input of a filter word addition request (step S1102-NO, step S1105-YES). ), The processing of step S1106 to step S1110 is performed, and when the filter word deletion request is received (step S1102, step S1105-NO, step S1111-YES), the processing of step S1112 is performed. The category operation unit 52 outputs the horizontal axis category xAxisCat of the displayed biaxial map and the vertical axis category yAxisCat to the biaxial map display unit 51 after the process of step S1110 or step S1112 is completed, and outputs the biaxial map. The display is updated (step S1104). The processing in steps S1105 to S1112 corresponds to the processing in step S10 in FIG. 7, and the subsequent processing in step S1104 corresponds to the processing in step S11 in FIG. The category operation unit 52 repeats the processing from step S1101. Details of these processes will be described later.

  When category selection is not input (step S1101-NO), or after category selection is input (step S1101-YES), the category operation unit 52 adds feature word clustering execution requests and filter words from the user. When neither a request, a filter word deletion request, or an end request is input (step S1102, step S1105, step S1111, step S1113-NO), the processing from step S1101 is repeated, and an end request is input ( Step S1102, Step S1105, Step S1111-NO, Step S1113-YES), the process is terminated.

  Next, details of the processing in steps S3 to S9 in FIG. 7 will be described. This processing corresponds to the processing in steps S1103 to S1104 in FIG.

  FIG. 10 is a flowchart showing a flow of processing in which the feature calculation unit 31 calculates the feature. The flowchart shown in the figure is an example of detailed processing of the feature calculation processing in step S3 of FIG. In this process, the feature calculation unit 31 extracts a word by performing morphological analysis on the text information of the target document data set stored in the document storage unit 1, and sets the calculated feature for the extracted word. Data is stored in the feature data storage unit 4.

  First, the document classification unit 3 receives the classification target category tgtCat from the category operation unit 52 of the user interface unit 5 (step S1201). That is, the document classification unit 3 receives the category cat output from the category operation unit 52 in step S1103 of FIG. 9 as the target category tgtCat. For example, as illustrated in FIG. 23, when the user selects the cell 1801 and inputs a feature word clustering execution request, the target category tgtCat becomes the “by content” category.

  The feature degree calculation unit 31 acquires a document data set tgtDocs classified into the target category tgtCat. Specifically, the feature calculation unit 31 refers to the category storage unit 2 and reads out a classification rule from the category data corresponding to the target category tgtCat and the higher category data of the category data. The feature calculation unit 31 selects document data d satisfying all the read classification rules from the document data stored in the document storage unit 1, and sets a set of the selected document data d as a document data set tgtDocs. . The feature degree calculation unit 31 acquires text information to be analyzed from all the document data d included in the document data set tgtDocs (step S1202). In the present embodiment, the document data stored in the document storage unit 1 has the same data format as the document data 200a or 200b shown in FIG.

  The feature calculation unit 31 performs morphological analysis on the body 205 that is the text information acquired in step S1202 (step S1203). The feature degree calculation unit 31 selects all the words (morpheme) t obtained as a result of the morpheme analysis one by one, and performs the processing from step S1205 to step S1209 for the selected word t (step S1204-NO).

  The feature degree calculation unit 31 determines whether or not the part of speech of the word t is a word of a predetermined part of speech to be included in the word vector and is not an unnecessary word (step S1205). For example, the feature degree calculation unit 31 performs word selection such that a word whose part of speech of the word t is a noun, a strange noun, a proper noun, or the like is included in the word vector and a conjunction or adverb is not included in the feature vector. Further, the feature degree calculation unit 31 compares the word t with an unnecessary word registered in advance as a word that does not indicate the feature of the document data. For example, when a patent document is to be processed, words such as “device” and “means” are unnecessary words because they do not represent the characteristics of the document. When it is determined that the part of speech of the selected word t is not a word of a predetermined part of speech to be included in the word vector or is an unnecessary word (step S1205—NO), the feature calculation unit 31 returns to step S1204, The selected word t is selected and the process is repeated.

  On the other hand, if it is determined that the part of speech of the selected word t is a word of a predetermined part of speech to be included in the word vector and is not an unnecessary word (step S1205—YES), the feature calculation unit 31 acquires in step S1202 The appearance frequency tf that is the number of occurrences of the word t in the text information of the document data set tgtDocs is calculated (step S1206). Further, the feature degree calculation unit 31 calculates the document frequency df, which is the number of document data in which the word t appears in the text information (body 205) among the document data included in the document data set tgtDocs (step S1207). The feature degree calculation unit 31 calculates the feature degree s (t) of the word t based on the following calculation formula (1) (step S1208).

  s (t) = tf × (log (| tgtDocs | / df) +1) (1)

  In the calculation formula (1), | tgtDocs | is the number of document data d (number of documents) included in the target document data set tgtDocs. This calculation formula is generally called TF / IDF and has been widely used in the fields of information retrieval and document classification. That is, the more frequently the word t appears in the document data d (tf is larger), or the fewer the documents that include the word t (all df is smaller) of all document data, the more the word t is in the document data d. It is considered to be a word that well describes the feature. In the present invention, this TF / IDF is used to calculate the word feature degree for the target document data set.

The feature calculation unit 31 generates feature data in which the word t, the document frequency df and the feature s (t) calculated for the word t are set as the word 510, the document frequency 501, and the feature 502, respectively. And it memorize | stores in the characteristic data storage part 4 (step S1209). The feature calculation unit 31 returns to step S1204, selects an unselected word t, and repeats the process.
And when it determines with having processed the process of step S1205-step S1209 about all the words t obtained as a result of the morphological analysis (step S1204-YES), the characteristic degree calculation part 31 complete | finishes a process.

  FIG. 11 is a flowchart showing a flow of processing in which the attention word setting unit 53 sets the attention word. The flowchart shown in the figure shows an example of detailed processing of the attention word setting processing in step S4 of FIG. In this processing, the focused word setting unit 53 accepts a plurality of focused words from the user as focused words when classifying the feature word category, and outputs it to the document classifying unit 3.

  First, for example, as shown in FIG. 23, the attention word setting unit 53 displays a attention word setting form 1910 for the user to set a attention word (step S1301). The attention word setting form 1910 includes a attention word input field 1911 and a attention word list 1913 in which a list of contents of the attention word list data 600 is displayed. Further, the attention word setting form 1910 sets an “add to list” button 1912 for adding the character string input in the attention word input field 1911 to the attention word list, and the word selected in the attention word list 1913 as the attention word. A “set to focus word” button 1914 for canceling, and a “cancel” button 1915 for canceling the focus word setting operation.

  In the target word candidate 601 of the target word list data 600, a filter word selected as a target word by the user on the biaxial map or a list of character strings input by the user in the target word input field 1911 is set. However, the initial value of the focused word list data 600 is an empty list. Note that the focused word setting unit 53 may set a list of words set in the feature data stored in the feature data storage unit 4 as the initial value for the focused word candidate 601. In this case, the initial values of the attention word setting 602 may all be “0” or all “1”.

  In the attention word list 1913, a check box indicating whether or not to actually set the attention word is displayed before the word registered as the attention word candidate 601. When the attention word setting form 1910 is initially displayed, the attention word setting unit 53 defaults to a check box corresponding to the attention word candidate 601 in which “1” is set in the attention word setting 602 of the attention word list data 600. Set and display the check.

  Next, in FIG. 11, the attention word setting unit 53 receives input of the attention word or selection of the attention word from the user (step S <b> 1302). Here, the user can also select a word that is actually used as a focused word from the focused word candidates displayed in the focused word list 1913 of the focused word setting form 1910 (check the check box). If the focused word is not in the focused word list 1913, the word (character string) can be directly input in the focused word input field 1911.

  When the focused word setting unit 53 receives a request for adding a focused word from the user, that is, when the user inputs a character string in the focused word input field 1911 and selects the “add to list” button 1912 (step (S1303-YES), the character string input in the attention word input field 1911 is additionally displayed in the attention word list 1913 (step S1304), and the process proceeds to step S1302. When there is no request for adding a focused word to the list from the user, the focused word setting unit 53 transitions to the process of step S1305 (step S1303-NO).

When there is no request for adding a focused word to the list (step S1303-NO), and when a setting request for the focused word is received from the user, that is, when the user selects the “set as focused word” button 1914, the focused The word setting unit 53 performs the processing of step S1306 and step S1307 (step S1305-[target word setting]).
That is, the attention word setting unit 53 stores the contents of the attention word list 1913 as the attention word list data 600 (step S1306). Specifically, the attention word setting unit 53 sets a word included in the attention word list 1913 as the attention word candidate 601. At this time, the attention word setting unit 53 sets a flag in the attention word setting 602 for the words whose check boxes are checked by the user in the attention word list 1913 (“1” in this embodiment), and is not checked. The word is stored in the focused word list data 600 without setting a flag (“0” in this embodiment). The attention word setting unit 53 reads the word set in the attention word candidate 601 in which the flag is set in the attention word setting 602 in the attention word list data 600 updated in step S1306. The focused word setting unit 53 outputs the set of read words as the focused word set to the document classification unit 3 (step S1307), and ends the focused word setting process of FIG.

  If neither the attention word setting request nor the cancellation request is input in step S1305, the attention word setting unit 53 transitions to the process of step S1302 (step S1305 [no request]). Alternatively, when there is a cancel request in step S1305, that is, when the user selects the “cancel” button 1915, the focused word setting unit 53 ends the focused word setting process of FIG. 11 (step S1305- [Cancel ]).

  FIG. 12 is a flowchart showing a flow of processing in which the feature correction unit 32 obtains a corrected feature. The flowchart shown in the figure shows an example of detailed processing of the feature correction processing in step S5 of FIG. In this processing, the feature correction unit 32 receives the attention word set from the attention word setting unit 53 of the user interface unit 5 and sets the correction feature in the feature data stored in the feature data storage unit 4.

  First, the feature correction unit 32 receives a focused word set ats that is a set of focused words at specified by the user via the focused word setting unit 53 of the user interface unit 5 (step S1401). Specifically, the feature correction unit 32 receives the attention word set output from the attention word setting unit 53 in step S1307 in FIG. 11 and sets it as the attention word set ats. The input of the focused word set ats by the user may be performed simultaneously with the input of the target category (category cat) in step S1103 in FIG. 9 or may be performed at another timing.

  The feature correction unit 32 selects all the attention words at in the received attention word set ats one by one, and repeats the process of step S1403 (step S1402-NO). That is, the feature correction unit 32 performs feature data acquisition processing for acquiring feature data corresponding to the attention word at (step S1403). By this processing, the feature correction unit 32 acquires the feature data of the attention word at if it is registered in the feature data storage unit 4, and if not registered, the feature degree of the attention word at. Data is generated and registered in the feature data storage unit 4. The attention word at which the feature data is not registered is, for example, a word such as a compound word that has not been acquired by morphological analysis in the feature calculation processing illustrated in FIG. Details of the feature data acquisition processing will be described later with reference to the flowchart of FIG.

  When the feature degree data acquisition processing is completed for all the attention words at in the attention word set ats (step S1402-YES), the feature degree correction unit 32 stores all the feature degree data registered in the feature degree data storage unit 4. The word t is selected one by one, and the following steps S1405 to S1410 are repeated for the selected word t (step S1404-NO).

  First, the feature correction unit 32 performs feature data acquisition processing for acquiring the feature data kd corresponding to the word t from the feature data storage unit 4 (step S1405). Details of the feature data acquisition processing will be described later with reference to the flowchart of FIG. The feature correction unit 32 acquires the feature s (t) from the feature data kd acquired in step S1405, and the initial correction feature ms (t) of the word t as shown in the following calculation formula (2). A value is set (step S1406).

  ms (t) = s (t) (2)

Subsequently, the feature correction unit 32 selects all the attention words at included in the attention word set ats one by one, and repeats the processing of step S1408 and step S1409 for the selected attention word at (step S1407—NO). ).
First, the feature correction unit 32 calculates the co-occurrence degree co (t, at) between the word t and the attention word at in the document data set tgtDocs classified into the target category (step S1408). The document data set tgtDocs can be acquired by the same processing as step S1202 in FIG. 10 or can be obtained by the feature calculation unit 31 in step S1202 in FIG. Here, the co-occurrence degree co (t, at) of the word t and the attention word at in the document data set tgtDocs is a value calculated by any of the following calculation formulas (3) to (7).

Number of co-occurrence = | t∩at | (3)
Dice coefficient D = | t∩at | / (| t | + | at |) (4)
Jaccard coefficient J = | t∩at | / | t∪at | (5)
Simpson coefficient S = | t∩at | / min (t, at) (6)
Cosine coefficient C = | t∩at | / sqrt (| t | × | at |) (7)

  In the above description, in the document data set tgtDocs, the number of document data d including the word t in the text information (hereinafter referred to as “occurrence number”) is | t |, and the occurrence number of the attention word at in the document data set tgtDocs is | At | In the document data set tgtDocs, the number of document data d including both the word t and the attention word at in the text information (hereinafter referred to as “co-occurrence number”) is | t∩at |, and the word t and the attention word at. Let | t | at | be the number of document data d that includes at least one of them in the text information. Min (t, at) indicates the smaller of the number of occurrences of the word t and the number of occurrences of the attention word at, and sqrt indicates that the square root is obtained.

  The feature correction unit 32 uses the co-occurrence co (t, at) between the word t calculated in step S1408 and the word of interest at (co), based on the following calculation formula (8), and the corrected feature ms (t ) Is updated (step S1409).

  ms (t) = ms (t) × co (t, at) (8)

  After the process of step S1409, the feature correction unit 32 returns to the process from step S1407, selects an unselected target word at in the target word set ats, and repeats the process. When the repetitive processing of step S1408 and step S1409 is completed for all the attention words at (step S1407—YES), the feature correction unit 32 corrects the corrected feature to the corrected feature 503 of the feature data kd acquired in step S1405. Insert degrees ms (t). The feature correction unit 32 updates the feature data of the word t currently stored in the feature data storage 4 with the feature data kd set with the correction feature 503 (step S1410).

  After step S1410, the feature correction unit 32 returns to step S1404, selects an unselected word t, and repeats the process. When the processing of steps S1405 to S1410 is completed for all words t (step S1404—YES), the feature correction unit 32 ends the feature correction processing.

  In the above, the feature correction unit 32 receives the attention word set from the attention word setting unit 53 and performs the feature correction processing. However, one attention word is input to the attention word setting unit 53. The feature correction processing may be executed sequentially by receiving the attention word each time.

  FIG. 13 is a flowchart showing a flow of processing in which the feature correction unit 32 acquires feature data. The flowchart shown in the figure shows an example of the feature data acquisition processing in steps S1403 and S1405 in FIG. In this process, the feature correction unit 32 acquires feature data of a predetermined word k from the feature data storage unit 4. The word k is the attention word at in the case of the process of step S1403, and is the word t in the case of the process of step S1405.

  When the feature correction unit 32 receives the feature data acquisition request for the word k (step S1501), the feature correction unit 32 determines whether the feature data for the word k exists in the feature data storage unit 4 (step S1502). When it is determined that the feature data for the word k is not stored in the feature data storage unit 4 (step S1502-NO), the feature correction unit 32 performs the following steps S1503 to S1507, and the word k The feature degree data kd for is generated.

  Steps S <b> 1503 to S <b> 1507 are processes for using words that are not obtained by morphological analysis by the feature calculation unit 31 to generate a feature word category. If the word t is extracted by morphological analysis (the target part-of-speech and not an unnecessary word), the feature data for the word t is generated in the process of the feature calculation unit 31 shown in FIG. However, the user can set an arbitrary character string as the focused word at in the focused word setting unit 53 of the user interface unit 5. At this time, the attention word at specified by the user is not necessarily included in the word t extracted by the morphological analysis. For example, when the user sets the character string “internal control” as the attention word at, and the morphemes extracted by the characteristic calculation unit 31 are “internal” and “control”, the characteristic corresponding to these two words Data is generated, but no feature data is generated for the word “internal control”. Such a problem occurs particularly in a compound word that handles a plurality of words as one word such as “internal control”. The processes in steps S1503 to S1507 are processes for dealing with this problem.

  Specifically, the feature correction unit 32 generates an appearance frequency that is the number of occurrences of the word k in the text information (body text 205) of all the document data d included in the document data set tgtDocs classified into the target category. kf is calculated (step S1503). The document data set tgtDocs can be acquired by the same processing as step S1202 in FIG. 10 or can be obtained by the feature calculation unit 31 in step S1202 in FIG. Here, since the word k is a word that is not extracted by the morphological analysis as described above, the feature correction unit 32 does not count the appearance frequency from the morphological analysis result, but counts it using a character string search or the like.

  Next, the feature correction unit 32 calculates a document frequency df that is the number of document data d in which the word k appears in the text data (body 205) in the document data set tgtDocs (step S1504). The feature correction unit 32 calculates the feature s (k) of the word k using the following calculation formula (9) similar to the calculation formula (1) in step S1209 (step S1505).

  s (k) = kf × (log (| tgtDocs | / df) +1) (9)

  The feature correction unit 32 generates feature data kd in which the word k, the calculated document frequency df, and the feature s (k) are set to the word 510, document frequency 501, and feature 502, respectively (step S1506) and stored in the feature data storage unit 4 (step S1507). The feature correction unit 32 outputs the generated feature data kd to the feature data acquisition request source (step S1509).

  On the other hand, if it is determined in step S1502 that the feature data of the word k is stored in the feature data storage unit 4 (YES in step S1502), the feature correction unit 32 reads the word from the feature data storage unit 4. The feature data kd of k is acquired (step S1508). The feature correction unit 32 outputs the acquired feature data kd to the feature data acquisition request source (step S1509).

  When the feature correction processing (step S5) shown in FIG. 7 is completed by the processing of FIG. 12 and FIG. 13, in the feature word category generation unit 33, the trend vector generation unit 34 performs the trend vector generation processing (step S6). The clustering unit 35 performs clustering processing (step S7), and the category generation processing unit 36 performs feature word category generation processing (step S8). In this manner, in step S6 to step S8, the feature word category generation unit 33 uses the feature data stored in the feature data storage unit 4 to use feature words having similar appearance tendencies with respect to the axis category set. A feature word category based on the group is generated.

  FIG. 14 is a flowchart showing a flow of processing in which the trend vector generation unit 34 of the feature word category generation unit 33 obtains a trend vector. The flowchart shown in the figure shows an example of detailed processing of the trend vector generation processing in step S6 of FIG. In this process, the trend vector generation unit 34 extracts a feature word set based on the feature data stored in the feature data storage unit 4, and the trend vector for each feature word included in the extracted feature word set. Is stored in the feature data storage unit 4.

  First, the trend vector generation unit 34 generates an empty feature word list tl (step S1601). The trend vector generation unit 34 selects all the words t stored with the feature data in the feature data storage unit 4 one by one, and performs the processing from step S1603 to step S1605 on the selected word t (step S1602). -NO). In steps S1603 to S1605, the trend vector generation unit 34 extracts feature words, and stores a set of extracted feature words in the feature word list tl.

  Specifically, the trend vector generation unit 34 reads the feature data of the word t stored in the feature data storage unit 4 and the corrected feature set in the corrected feature 503 from the read feature data. ms (t) is acquired (step S1603). When the acquired corrected feature value ms (t) is larger than a predetermined threshold value minms (minms <ms (t)) (step S1604-YES), the trend vector generation unit 34 characterizes the word t. It adds to the word list tl (step S1605). When the acquired corrected feature value ms (t) is equal to or less than the threshold value minms (minms ≧ ms (t)) (NO in step S1604), or after the processing in step S1605, the trend vector generation unit 34 performs step S1602. Returning to FIG. 5, the unselected word t is selected and the process is repeated.

  Here, the threshold value minms is the minimum value of the correction feature degree ms (t) for the word t, and is a value set in advance on the system side. Feature words are extracted based on the threshold value minms. However, in the present embodiment, the threshold value is set as the minimum value of the corrected feature value ms (t). However, the present invention is not limited to this. Feature words may be extracted by specifying the number of feature words.

  When the iterative process is completed for all the words t for which the feature degree data is stored in the feature degree data storage unit 4 (step S1602-YES), the tendency vector generation unit 34 includes all the feature words t included in the feature word list tl. Are selected one by one, and the processing of step S1607 to step S1612 is repeated for the selected feature word t (step S1606-NO).

  First, the trend vector generation unit 34 acquires the feature data kd of the feature word t from the feature data storage unit 4 (step S1607). Further, the trend vector generation unit 34 obtains the frequency vector vcf and the trend vector vptn having the same number of dimensions as the number of categories included in the horizontal axis category set xCats input from the category operation unit 52 (the number of horizontal axis categories xCat). Generate (step S1608). Each element of the frequency vector vcf and the trend vector vptn corresponds to the horizontal axis category xCat.

The trend vector generation unit 34 selects all the horizontal axis categories xCat included in the horizontal axis category set xCat one by one, and repeats the processing of Step S1610 and Step S1611 for the selected horizontal axis category xCat (Step S1611). S1609-NO).
In other words, the trend vector generation unit 34 determines the number of document data in which the feature word t appears in the text information (the body 205) (hereinafter, “for the document data set included in common in the target category tgtCat and the horizontal axis category xCat”. Cf) (referred to as “frequency within category”) (step S1610). The document data set that is commonly included in the target category tgtCat and the horizontal axis category xCat is document data that satisfies the logical product of the classification rule of the target category tgtCat and the classification rule of the horizontal axis category xCat. It is obtained by the same processing as S1009. The trend vector generation unit 34 acquires the document frequency 501 as the document frequency df of the feature word t from the feature data kd. The trend vector generation unit 34 sets the element value corresponding to the horizontal axis category xCat of the frequency vector vcf as the intra-category frequency cf calculated in step S1610, and the value of the element corresponding to the horizontal axis category xCat of the trend vector vptn. Is set to cf / (df + 1) (step S1611). The trend vector generation unit 34 returns to step S1609, selects an unselected horizontal axis category xCat, and repeats the process.

When the trend vector generation unit 34 determines that the processes in steps S1609 to S1611 have been performed for all the horizontal axis categories xCat (step S1609-YES), the trend vector generation unit 34 arranges the elements calculated in step S1611 for each horizontal axis category xCat. The frequency vector vcf and the trend vector vptn are stored in the frequency vector / trend vector 504 for the axis category (category number) of the feature data kd. The trend vector generation unit 34 updates the feature data of the feature word t currently stored in the feature data storage unit 4 with the feature data kd storing the frequency vector vcf and the trend vector vptn (step S1612). Thereafter, the trend vector generation unit 34 returns to step S1606, selects an unselected feature word t, and repeats the process.
If it is determined in step S1606 that the processing of steps S1607 to S1612 has been performed for all words (feature words) t included in the feature word list tl (YES in step S1606), the tendency vector generation unit 34 determines the tendency. The vector generation process ends.

  In this embodiment, the value of the element of the trend vector is cf / (df + 1), that is, the “appearance ratio” of the word t in the horizontal axis category xCat with respect to the target category tgtCat, but the document frequency (df ) Or in-category frequency (cf). Alternatively, values based on statistics such as the following self-mutual information amount and Yates correction χ square value may be used. Statistics are calculated by conventional techniques. The value of the element of the trend vector becomes a weight of each feature word in the clustering process and is reflected in the clustering result.

  The self mutual information PMI is calculated by the following calculation formula (10).

  Self mutual information PMI = log (an / ((a + b) (a + c))) (10)

  The Yates correction χ square value Yates is calculated by the following calculation formula (11).

Yates' = n (| ad-bc | -n / 2) ^ 2 / ((a + b) (c + d) (a + c) (b + d))
if ((ad−bd) <0) Yates = −Yates ′
else Yates = Yates' (11)

  In the calculation formulas (10) and (11), | xCat | is the number of documents classified into the horizontal axis category xCat, | tgtCat | is the number of documents classified into the target category tgtCat, and a, b, c, d, and n are as follows.

a = cf
b = df-cf
c = | xCat | -cf
d = | tgtCat | −df) − | xCat | + cf
n = a + b + c + d = | tgtCat |

In the self-mutual information amount PMI, feature words having a large bias between the appearance probability in the target category tgtCat and the appearance probability in the horizontal axis category xCat are highly evaluated. In addition, since there is a tendency to overestimate low-frequency words, when using the self-mutual information PMI, it is necessary to perform processing such as excluding words with extremely low document frequency df from feature words.
On the other hand, the Yates correction chi-square value Yates highly evaluates words having a high appearance probability in the horizontal axis category xCat with respect to the appearance probability in the target category tgtCat. As a result, relatively low frequency feature words are strongly weighted in clustering. However, the tendency is small compared with the self-mutual information amount PMI.

  FIG. 15 is a flowchart showing a flow of feature word clustering processing executed by the clustering unit 35 of the feature word category generation unit 33. The flowchart shown in the figure is an example of detailed processing of the clustering processing in step S7 of FIG. In this process, the clustering unit 35 performs clustering of feature words based on the similarity of feature words using the trend vectors generated in the above-described trend vector generation process of FIG. 14, and is a strongly related feature word group. Generate feature word clusters.

  In the present embodiment, the document clustering method is applied as the clustering method. In the conventional document clustering method, for example, the features included in each document data are represented by feature vectors (the frequency of appearance of words in the document is used as vector elements), and the similarity (for example, inner product or cosine) of the feature vectors. ) To generate a document cluster that is a group of documents.

  The clustering unit 35 of the present embodiment generates a feature word cluster based on the similarity of the extracted feature word tendency vectors. Various clustering methods have been conventionally devised, but in this embodiment, a relatively simple clustering method called a leader-follower method is used. However, this clustering method is not limited.

  First, the clustering unit 35 generates a feature data set kds, which is a set of feature data kd in which the trend vector 512 is set in the frequency vector / trend vector 504 for the axis category (category number). (Step S1701). The clustering unit 35 sets the word t included in the acquired feature data set kds as the clustering target word set T, and sets the initial value of the feature word cluster set C as the classification destination as an empty set (step S1702). The word t included in the feature data set kds is a word set as the word 510 in the feature data kd. The clustering unit 35 selects all the words t included in the word set T one by one, and repeats the processing from step S1704 to step S1715 for the selected word t (step S1703-NO).

  First, the clustering unit 35 acquires the tendency vector vptn of the word t from the feature data set kds (step S1704). The clustering unit 35 sets the initial value of the classification target feature word cluster cmax, which is the feature word cluster of the classification target of the word t, to “none”, and sets the initial value of the maximum value smax of the similarity of the word t to 0 (step S1705). ).

  The clustering unit 35 selects all the feature word clusters c included in the feature word cluster set C one by one, and repeats the processing of step S1707 to step S1709 for the selected feature word cluster c (step S1706-NO). When the clustering unit 35 finishes the processing for all the feature word clusters c included in the feature word cluster set C (step S1706: YES), the clustering unit 35 performs the process of step S1710.

  However, in the case of processing for the first word t, since C is an empty set of initial values, the clustering unit 35 determines that the processing has been completed for all feature word clusters c included in the feature word cluster set C ( Step S1706—YES), it is determined whether or not the classification target feature word cluster cmax exists (step S1710). Since the classification target feature word cluster cmax is the initial value “none” (step S1710—NO), the clustering unit 35 newly creates a feature word cluster cnew, which is a new feature word cluster c, and the created feature word cluster cnew is added to the feature word cluster set C (step S1713). The clustering unit 35 classifies the tendency vector vptn of the word t into the created feature word cluster cnew (step S1714). Then, the clustering unit 35 sets the feature vector vc of the feature word cluster cnew as the tendency vector vptn of the word t (step S1715). That is, at this time, since the word categorized in the feature word cluster cnew is the word t1, the feature vector vc of the feature word cluster cnew is the same as the tendency vector vptn of the word t. When the process of step S1715 ends, the clustering unit 35 returns to step S1703, selects an unselected word t, and repeats the process.

  Since the feature word cluster set C is not an empty set in the second and subsequent iterations from step S1703, in step S1706, the clustering unit 35 selects all the feature word clusters included in the feature word cluster set C. c is selected one by one, and the processing of step S1707 to step S1709 is repeated (step S1706-NO).

  Specifically, the clustering unit 35 calculates the similarity s between the word t and the feature word cluster c using the tendency vector vptn of the word t and the feature vector vc of the feature word cluster c (step S1707). The feature vector vc of the feature word cluster c is generated with the initial tendency vector of the word classified into the feature word cluster c in step S1715, and additionally classified into the feature word cluster in steps S1710 to S1715 described later. This is a vector that is updated by using the trend vector of the updated word. The clustering unit 35 calculates the similarity s between the word t and the feature word cluster c based on the similarity between the tendency vector vptn of the word t and the feature vector vc of the feature word cluster c. This similarity is calculated using, for example, the cosine of a vector, that is, (vptn · vc) / (| vptn | × | vc |). Note that vptn · vc represents the inner product of the trend vector vptn and the feature vector vc, and | vptn | and | vc | represent the norms of the trend vector vptn and the feature vector vc, respectively.

  The clustering unit 35 determines that the similarity s between the word t and the feature word cluster c is greater than or equal to a predetermined threshold value smin and is greater than the maximum similarity smax of the word t (step S1708—YES). ), The feature word cluster c is set as the classification target feature word cluster cmax, and the similarity s with the feature word cluster c is set as the maximum similarity smax of the word t (step S1709). The clustering unit 35 returns to Step S1706 and repeats the process. On the other hand, the clustering unit 35 determines that the similarity s between the word t and the feature word cluster c is less than a predetermined threshold value smin, or is equal to or less than the maximum value smax of the similarity of the word t (step (S1708-NO), the process returns to step S1706 and the process is repeated.

  After the repetition process of step S1706 is completed (step S1706—YES), the clustering unit 35 determines whether or not the classification target feature word cluster cmax exists (step S1710). The clustering unit 35 has a classification target feature word cluster cmax. That is, among the existing feature word clusters c, the cluster has the maximum similarity s with the word t and the similarity is equal to or greater than a threshold value. Is present (step S1710—YES), the clustering unit 35 classifies the word t into the classification target feature word cluster cmax (step S1711). The clustering unit 35 adds the tendency vector vptn of the word t to the feature vector vc of the classification target feature word cluster cmax, and recalculates the feature vector vc (step S1712). That is, the feature vector vc of the feature word cluster c that is the classification target feature word cluster cmax is the sum of the tendency vectors of the word t classified into the feature word cluster c. Therefore, the result of adding the tendency vector vptn of the word t to the feature vector vc of the feature word cluster c that is the classification target feature word cluster cmax becomes the feature vector of the feature word cluster c after the classification of the word t. .

  On the other hand, if the classification target feature word cluster cmax does not exist in step S1710 (step S1710-NO), the clustering unit 35 performs the processing of step S1713 to step S1715 in the same manner as described above. That is, the clustering unit 35 newly creates a feature word cluster cnew and adds it to the feature word cluster set C (step S1713), and classifies the tendency vector vptn of the word t into the feature word cluster cnew (step S1714). The clustering unit 35 sets the feature vector vc of the feature word cluster cnew as the tendency vector vptn of the word t (step S1715).

  When step S1712 or step S1715 ends, the clustering unit 35 returns to step S1703 and repeats the process. When the iterative process ends for all the words t included in the word set T (step S1703—YES), the clustering unit 35 ends the clustering process of FIG.

  As described above, the tendency vectors of the words for which the trend vector is generated in the trend vector generation process shown in FIG. 14 by the clustering process of the clustering unit 35 shown in FIG. 15, that is, the words determined as the feature words are similar to each other. A feature word cluster that is a group of feature words is generated.

  In the feature degree calculation processing in FIG. 10, since words are extracted by morphological analysis, feature words (filter words) automatically generated by the document classification unit 3 are in morpheme units. On the other hand, in the focused word setting unit 53, the focused word directly input by the user does not need to be in morpheme units, and an arbitrary character string can be set as the focused word. And even if the character string set as the focused word by the user in the focused word setting process in FIG. 11 is not in morpheme units, the character string is subsequently treated as a word by the feature data acquisition process shown in FIG. Furthermore, in the feature correction processing in FIG. 12, the feature correction unit 32 of the document classification unit 3 corrects the word feature according to the target word set input from the target word setting unit 53, and in FIG. The clustering unit 35 realizes feature word clustering according to the word that the user focuses on.

  FIG. 16 is a flowchart showing a flow of processing in which the category generation processing unit 36 of the feature word category generation unit 33 generates a feature word category. The flowchart shown in the figure shows an example of detailed processing of the feature word category generation processing in step S8 of FIG. In this process, the category generation processing unit 36 generates feature word category data using the words included in the feature word clusters generated by the clustering process of FIG. 15 described above as filter words and the filter words as classification conditions. And stored in the category storage unit 2.

  The category generation processing unit 36 receives the feature word cluster set C generated by the clustering shown in FIG. 15 (step S1801). The category generation processing unit 36 acquires from the feature data storage unit 4 a feature data set kds that is a set of feature data kd of all the words included in the clusters in the feature word cluster set C (step S1802). The category generation processing unit 36 selects all the feature word clusters c included in the feature word cluster set C one by one, and repeats the processing of step S1804 to step S1807 for the selected feature word cluster c (step S1803-NO).

  First, the category generation processing unit 36 sets a set of words t included in the feature word cluster c as a filter word set ts, refers to the feature data set kds, and has the highest corrected feature in the filter word set ts. Let t be the category name cn (step S1804). For example, it is assumed that a set of words t in the feature word cluster c is {“search”, “classification”, “management”}. In this case, the category generation processing unit 36 uses {“search”, “classification”, “management”} as the filter word set ts of the feature word cluster c, and is the word having the largest correction feature degree in the filter word set ts. Let “search” be the category name cn.

  The category generation processing unit 36 generates a classification rule r based on the filter word set ts (step S1805). For example, as shown in the feature word category data 400 and 410 of FIG. 4, the classification rule r is based on the condition that “the filter word set ts is included in the body 205 (text information) of the document data”. It is generated as a rule for classifying.

  When the target document data is an XML document as shown in the document data 200b in FIG. 2, the classification rule is expressed by XQuery or XPath. In the above example, when the “body” element of the document data 200b is text information, the classification rule of the feature word cluster c is “contains (./body,“ search ”) and containers (./body,“ classification ”) and contains (contains (./text, “management”)).

  The category generation processing unit 36 acquires a document data set docs, which is a set of document data d corresponding to the generated classification rule r, from the document storage unit 1 (step S1806). The category generation processing unit 36 generates feature word category data based on the target category tgtCat, the category name cn, the filter word set ts, and the classification rule r, and stores it in the category storage unit 2 (step S1807). That is, the category generation processing unit 36 generates feature word category data in the same format as the feature word category data 400 and 410 shown in FIG. Specifically, the category generation processing unit 36 adds the category number of the feature word category data, the upper category, the category name, the classification rule, and the filter data, the arbitrary number newly assigned, the category data or the feature of the target category tgtCat, respectively. The category number of the word category data, the category name cn, the classification rule r, and the filter word set ts are set. Further, the category generation processing unit 36 writes the document data set docs in association with the generated feature word category data. For example, the document data set docs is indicated by the document number of the document data. Thereafter, the category generation processing unit 36 returns to the processing from step S1803, selects an unselected feature word cluster c, and repeats the processing.

  In step S1803, when the iterative processing of step S1804 to step S1807 is completed for all the feature word clusters c in the feature word cluster set C (step S1803-YES), the category generation processing unit 36 ends the feature word category generation processing. .

  FIG. 24 is a diagram showing a display example of the category structure after the feature word category is generated by executing the feature word clustering process of FIG. As shown in the figure, a category 2031, a “search” category 2032, and a “mining” category 2033 are generated as feature word categories in the lower categories of the “by contents” category 1603. The category data corresponding to the “analysis” category 2031 and the “mining” category 2033 are the feature word category data 400 and 410 shown in FIG.

  When the processes of FIGS. 10 to 16 are completed, the category operation unit 52 of the user interface unit 5 performs the process of step S1104 of FIG. 9 and instructs the biaxial map display unit 51 to display the biaxial map. The biaxial map display unit 51 receives an instruction from the category operation unit 52 and performs biaxial map display processing in step S6 of FIG.

  FIG. 17 is a flowchart showing a flow of processing in which the biaxial map display unit 51 displays the biaxial map. The flowchart shown in the figure shows an example of detailed processing of the biaxial map display processing in step S9 of FIG.

  The biaxial map display unit 51 receives the horizontal axis category xAxisCat and the vertical axis category yAxisCat of the biaxial map from the category operation unit 52 (step S1901). The biaxial map display unit 51 performs the processing of the horizontal axis category xCat of the horizontal axis category xAxisCat based on the category data and feature word category data stored in the category storage unit 2 by the same processing as Step S1002 of FIG. The horizontal axis category set xCats, which is a set, and the vertical axis category set yCats, which is a set of vertical axis categories yCat of the vertical axis category yAxisCat, are acquired (step S1902).

  Specifically, in the case of the category structure as shown in FIG. 24, the two-axis map display unit 51 sets a set of child categories of the “by application year” category 1602 that is the horizontal axis category xAxisCat as the horizontal axis category set xCats. {"2004" category 1621, "2005" category 1622, "2006" category 1623, "2007" category 1624, "2008" category 1625} are acquired. Further, the biaxial map display unit 51 acquires a set {“analysis” category 2031, “search” category 2032, “mining” category 2033} of child categories of the “by contents” category 1603 which is the vertical axis category yAxisCat.

  The biaxial map display unit 51 uses the vertical axis category yAxisCat and the vertical axis category yCat included in the vertical axis category set yCats as a row, and the horizontal axis category xCat included in the horizontal axis category set xCats. A two-axis map table is created and displayed (step S1903). Since the 2-axis map table also includes title rows and title columns in which categories are displayed, the number of rows is (1 + vertical category number + vertical child category number) and the number of columns is (1 + horizontal child category number). It is.

  The biaxial map display unit 51 selects all the cells in the created biaxial map table one by one, and repeats the processing of step S1905 to step S1910 for the selected cell (step S1904-NO). The processes in steps S1905 to S1910 are the same as those in steps S1005 to S1010 in FIG.

  First, the biaxial map display unit 51 determines whether or not the cell is the first row (first row) or the first column (first column) (step S1905). When it is determined that the cell is the first row (first row) or the first column (first column) (YES in step S1905), the biaxial map display unit 51 performs the processing from step S1906 to step S1908. That is, the biaxial map display unit 51 displays the category name of the category cat (vertical axis category yAxisCat, vertical axis child category yCat, or horizontal axis child category xCat) corresponding to the selected cell (step S1906). Further, the biaxial map display unit 51 determines whether or not the category cat corresponding to the cell has a filter word set filters (step S1907). When it is determined that the category cat has filter word set filters (step S1907—YES), the biaxial map display unit 51 displays the filter words included in the filter word set filters on the cell (step S1908). When it is determined that the category cat does not have the filter word set filter (step S1907—NO), or after the process of step S1908, the biaxial map display unit 51 returns to step S1904 and selects an unselected cell. Repeat the process.

  If it is determined in step S1905 that the cell is neither the first row (first row) nor the first column (first column) (NO in step S1905), the biaxial map display unit 51 displays the vertical axis corresponding to the cell row. The number of documents dn which is the number of document data classified into both the category yAxisCat or the vertical axis category yCat and the horizontal axis category xCat corresponding to the cell column is obtained (step S1909). The biaxial map display unit 51 displays a circle chart having a size corresponding to the number of documents dn calculated in step S1909 on the cell of the biaxial map table created in step S1903 (step S1910). Thereafter, the biaxial map display unit 51 returns to step S1904, selects an unselected cell, and repeats the process.

  When the biaxial map display unit 51 finishes the processing of step S1905 to step S1910 for all the cells in step S1904 (step S1904—YES), it performs the processing of step S1911.

  FIG. 25 shows the display of the biaxial map display unit 51 at the end of the processing up to step S1904 when the feature word category is generated by executing the feature word clustering process of FIG. It is a figure which shows the example of a display of the 2-axis map table to be made. The horizontal axis category (axis category) of the biaxial map shown in the figure is the same as the biaxial map shown in FIG. The vertical axis includes the “category by content” category that is the target category, and the feature word categories “analysis”, “search”, and “mining” that are subcategories generated for the “by content” category. The filter words for each word category are displayed. For example, the filter word “mining, analysis, and related words” is displayed in the cell 2101 of the feature word category “mining”. Further, when displaying the biaxial map, the biaxial map display unit 51 is classified into each category of the row item category and the column item category corresponding to each cell of the biaxial map. A graph corresponding to the number of documents is displayed.

  After the repetition of step S1904, when the user selects a filter word f from the filter word set filters of a certain category cat (step S1911-YES), the biaxial map display unit 51 performs the processes of steps S1912 to S1915. Do.

  The biaxial map display unit 51 acquires the parent category pcat of the category cat for which the filter word f is selected from the category storage unit 2 (step S1912). The biaxial map display unit 51 selects all the cells other than the title column in the row corresponding to the category cat one by one, and repeats the processing of step S1914 and step S1915 for the selected cell (step S1913-NO).

  The biaxial map display unit 51 performs classification rules based on category data (or feature category data) of the horizontal axis category xCat corresponding to the selected cell and category data (or feature category data) of the upper category of the horizontal axis category xCat. And the logical product of the read classification rules is set as the classification rule xr of the horizontal axis category xCat. Further, the biaxial map display unit 51 reads out the classification rule pr from the category data (or feature category data) of the parent category pcat. The biaxial map display unit 51 refers to the document storage unit 1 and selects the selected cell based on the classification rule xr of the horizontal axis category xCat, the classification rule pr of the parent category pcat, and the selected filter word f. The number of documents fdn, which is the number of target document data including the filter word, is obtained (step S1914). The number of documents fdn can be obtained by a product of conditional expressions as in the case of dn described above, and the conditional expressions are “xr and pr and (contains (./text, f))”.

The biaxial map display unit 51 displays, on the cell selected in the displayed biaxial map table, a circle chart having a size corresponding to the number of documents fdn with a color different from the circle chart displayed in step S1910. (Step S1915). The category operation unit 52 returns to step S1913, selects an unselected cell, and repeats the process.
When the process is finished for all cells other than the title column in the row corresponding to the category cat for which the filter word f is selected (step S1913—YES), the category operation unit 52 returns to step S1911.

  In the processes in steps S1912 to S1915 described above, when the user selects a filter word on the two-axis map, the two-axis map display unit 51 displays the filter word on the cell in each column of the row in which the filter word is selected. Are displayed separately from the circle chart displayed in step S1910.

  FIG. 26 is a diagram illustrating a display example of a biaxial map when a filter word is selected. In the example of FIG. 26, after the display of FIG. 25, the user selects the filter word “related word” in the cell 2201 in which the vertical axis category “mining” is displayed. The biaxial map display unit 51 sets the horizontal axis categories “2004”, “2005”, “2006”, “2007”, “2008” for the row including the cell 2201 in which the filter word is selected. The number of document data including the “related word” is indicated by the shaded portion in the cell of the column corresponding to each of “”. For example, in the cell 2222, a “related word” is included in the document data included in both the feature word category “mining” corresponding to the cell 2201 and the horizontal axis category “2004” corresponding to the cell 2211. Indicates the number of document data included.

  When the biaxial map display unit 51 determines in step S1911 in FIG. 17 that no filter word selection has been input (step S1911—NO), when no end request is input from the user (step S1916—NO), Returning to the process of step S1911, the process ends when an end request is received (step S1916—YES).

Next, detailed processing in step S10 and step S11 in FIG. 7 will be described using the processing flow in FIG. 9 and FIG.
In step S1104 of FIG. 9, when the biaxial map display unit 51 receives an instruction from the category operation unit 52 and displays the biaxial map as shown in FIG. 25 by the process of FIG. The process returns to step S1101. The category operation unit 52 receives an input of the category cat selected by the user on the currently displayed two-axis map (step S1101-YES), and further receives an input of a filter word addition request (step S1102-). NO, Step S1105—YES), Step S1106 to Step S1110 are performed.

  First, the category operation unit 52 reads out the feature data kd in which the tendency vector is set from the feature data stored in the feature data storage unit 4, and the feature that is a set of the read feature data kd. The degree data set kds is acquired (step S1106). The feature data set kds is a set of feature data of words determined as feature words by the document classification unit 3. The category operation unit 52 displays the words in the feature data set kds in the order of corrected features (step S1107). In the present embodiment, as described above, the display of the feature words is in the corrected feature degree order. However, the display is not limited to this, and may be simply in the document frequency order or the feature degree order.

  FIG. 27 is a diagram illustrating an editing operation of a feature word category in a biaxial map and a display example of the screen. In the figure, the category operation unit 52 displays a feature word addition screen 2310 of the feature word category “mining” corresponding to the cell 2303 selected in step S1101. The category operation unit 52 causes the feature word list display field 2311 of the feature word addition screen 2310 to display a list of feature words that are words included in the feature data set kds acquired in step S1106. At this time, the category operation unit 52 displays each feature word with a check box. Further, the category operation unit 52 displays the check box corresponding to the feature word that is the filter word of the selected category “cat” as an initial display. The feature word addition screen 2310 includes an input field 2312 for the user to input an arbitrary character string as the filter word of the selected cat, and “add to filter word” for requesting the user to execute addition of the filter word. A button 2313 is included.

  The category operation unit 52 accepts selection or input of a filter word from the user for the display of the feature word (step S1108). Specifically, the user inputs a character string to be added as a filter word in the input field 2312 or checks a check box corresponding to a feature word to be added as a filter word in the feature word list display field 2311. If the user does not receive an additional execution request for the filter word f (NO in step S1109), the category operation unit 52 repeats the processing from step S1105.

  In step S1109, when the execution request for adding the filter word f is received from the user, specifically, when the user selects the “add to filter word” button 2313 in FIG. 27 (step S1109—YES), the category operation is performed. The unit 52 performs a process of adding the filter word f to the feature word category data of the selected category cat (step S1110). The filter word f is a character string input in the input field 2312 by the user in step S1108 or a feature word checked in the feature word list display field 2311. The details of the process of adding the filter word f will be described later in the flow of the filter word adding / deleting process of FIG.

  After the process of step S1110, the category operation unit 52 performs the process of step S1104 described above to instruct the biaxial map display unit 51 to update the biaxial map. The biaxial map display unit 51 receives the input of the horizontal axis category xAxisCat and the vertical axis category yAxisCat of the biaxial map currently displayed from the category operation unit 52, performs the process of FIG. Reflect in the axis map. The category operation unit 52 returns to the process of step S1101.

  The category operation unit 52 receives an input of the category cat selected by the user (step S1101-YES), and further receives a request to delete the filter word f from the user (step S1102, step S1105-NO, step S1111-). YES), a process of deleting the filter word f from the feature word category data of the selected category cat is performed (step S1112). Details of the process of deleting the filter word f will be described later in the flow of the process of adding / deleting a filter word in FIG.

  After the process of step S1112, the category operation unit 52 performs the process of step S1104 described above to instruct the biaxial map display unit 51 to update the biaxial map. The biaxial map display unit 51 receives the input of the horizontal axis category xAxisCat and the vertical axis category yAxisCat of the biaxial map currently displayed from the category operation unit 52 and performs the process of FIG. Reflect in the axis map. The category operation unit 52 returns to the process of step S1101.

  FIG. 18 is a flowchart showing the flow of processing for adding or deleting filter words in the feature word category in the category operation unit 52. The flowchart shown in the figure shows an example of detailed processing in steps S1110 and S1112 of FIG.

  First, the category operation unit 52 receives a word t to be added or deleted and a category cat as inputs (step S2001). The category cat is the category selected in step S1101 in FIG. 9, the word t to be added is a filter word selected or input by the user in step S1108 in FIG. 9, and the word to be deleted is step S1111 in FIG. Is the filter word f requested to be deleted. The category operation unit 52 specifies the feature word category data of the category cat stored in the category storage unit 2, and acquires the filter word set fs which is a set of filter words set in the specified feature word category data. .

  In the case of adding a filter word (Step S2002- [Add]), the category operation unit 52 adds the word t to the filter word set fs (Step S2003). However, if the word t already exists in the filter word set fs, the category operation unit 52 does nothing. On the other hand, in the case of deleting a filter word (step S2002-[delete]), the category operation unit 52 deletes the word t from the filter word set fs (step S2004). However, if the word t does not exist in the filter word set fs, the category operation unit 52 does nothing.

  After the process of step S2003 or step S2004, the category operation unit 52 generates a classification rule r based on the filter word set fs and updates the classification rule of the category cat (step S2005). For the classification rule generation method, refer to the description of step S1805 in FIG. The category operation unit 52 includes the classification rule r set in the feature word category data of the category cat stored in the category storage unit 2 and the classification rule r updated in step S2005 and the step S2003 or step. The filter word set fs updated in S2004 is updated (step S2006).

  In this embodiment, filter word addition / deletion processing has been described, but by combining addition / deletion, the category operation unit 52 can move or copy a filter word of a certain category as a filter word of another category. It becomes possible.

FIG. 28 is a diagram illustrating a display example of the biaxial map after the feature word category editing operation. In the biaxial map illustrated in FIG. 27, the user performs the following editing operations (1) and (2). It is a later display example.
(1) The user moves the filter word “management” in the “search” category displayed in the cell 2302 to the “analysis” category displayed in the cell 2301.
(2) The user deletes the filter word “related word” in the “mining” category displayed in the cell 2303.

  By the processing in steps S1105 to S1112 of FIG. 9 described above, the user can easily edit the feature word category automatically generated by the document classification unit 3 on the biaxial map. Therefore, the document classification apparatus 100 can configure a category structure in accordance with the user's intention and purpose of classification / analysis. In addition, when the user selects a filter word at this time, the document classification apparatus 100 performs display as shown in FIG. By this display, the user can grasp the appearance tendency related to the selected filter word. Therefore, the document classification device 100 can support the classification / analysis work based on the horizontal axis category designated by the user (the “by application year” category in FIG. 26).

  In this embodiment, each cell of the biaxial map is expressed by a so-called bubble chart in which the number of documents dn is displayed in a circle size, but the display of the number of documents dn is not limited to this. For example, the number of documents dn corresponding to each cell may be expressed by a line graph or a bar graph.

  FIG. 29 is a diagram illustrating a display example when the biaxial map is expressed by a line graph. As shown in the figure, when a line graph is used, a patent document is effective for grasping a time-series trend change, such as grasping a state of transition of application tendency. At this time, as shown in the figure, by expressing the number of documents dn with a graph of different line types for the filter word, it is possible to further easily grasp the appearance tendency for each filter word.

  Moreover, in the said embodiment, although the word used as a feature word is selected based on correction | amendment feature degree, you may select based on feature degree. In this case, the feature data does not include correction feature data, and the feature correction unit 32 does not perform the processing from step S1404 to step S1410 in FIG. Then, in step S1603 of FIG. 14, the tendency vector generation unit 34 acquires the feature degree of the word t from the feature degree data stored in the feature degree data storage unit 4. In step S1604, the trend vector generation unit 34 performs the process of step S1605 if the acquired feature value is larger than a predetermined threshold value, and returns to step S1602 if the feature value is smaller than the threshold value.

  Moreover, in the said embodiment, although the input of the attention word from a user is received, there is no need of input of a attention word. When there is no input of the word of interest, in the process of FIG. 12, the feature correction unit 32 does not repeat the processes of steps S1401 to S1403 and steps S1407 to S1409. After calculating the corrected feature value ms (t) in step S1406, the feature value correcting unit 32 performs the process of step S1410 for storing the calculated corrected feature value ms (t) in the feature data, and returns to step S1404. It becomes processing.

  In addition, when the target word is not input, the document classification apparatus 100 may be configured not to include the target word setting unit 53. In this case, the process of FIG. 12 is not performed, and in step S1603 of FIG. 14, the tendency vector generation unit 34 acquires the feature degree of the word t from the feature degree data stored in the feature degree data storage unit 4. In step S1604, the trend vector generation unit 34 performs the process of step S1605 if the acquired feature value is larger than a predetermined threshold value set in advance, and proceeds to step S1602 if it is smaller than the threshold value. Return.

  In the above embodiment, an example is shown in which the horizontal axis is the axis category and the vertical axis is the target category, but the vertical axis may be the axis category and the horizontal axis may be the target category.

  According to the document classification device 100 of at least one embodiment described above, by having the category operation unit 52 and the feature word category generation unit 33, the two-axis map having two categories specified by the user as two axes. A feature word is selected based on a word appearance tendency with respect to the classification axis selected by the user, and a feature word category is generated using the selected feature word. As a result, the user can select a classification axis while looking at the two-axis map using the currently generated clustering result and classification structure. Therefore, the document classification apparatus 100 has a classification structure suitable for the user's viewpoint. It is possible to generate and support classification and analysis suitable for the user's purpose.

In addition, in conventional word clustering, the word used to generate the cluster is appropriately selected and processed with a relatively small amount of calculation. This word selection is based on the statistical appearance tendency of words appearing in the document. Was done automatically based on. For this reason, words that do not express the contents of the document set very well, or words that do not match the user's intention or the purpose of analysis, are often selected as the words used for cluster generation. In such a case, the user performs operations such as removing unnecessary words or registering important words and uses them to generate a cluster suitable for the user's purpose. It is necessary to instruct the user to select a desired word, and this work requires skill and effort.
However, according to the document classification device 100 of at least one embodiment described above, the attention word setting unit 53 is provided so that the user receives the designation of the attention word, and the feature word is strongly related to the designated attention word. Generate feature word categories. Further, the document classification apparatus 100 has the category operation unit 52, and receives the editing of the filter word for the generated feature word category, and updates the biaxial map using the filter word of the edited word. As a result, it is possible to realize generation and correction of a classification structure based on feature words according to the user's interest while reducing the effort for the user to select an important word for cluster generation.

  Note that a program for realizing the function of the document classification apparatus 100 of FIG. 1 in each of the above-described embodiments is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read by a computer system. By executing this, the document classification apparatus 100 may be operated. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

  As mentioned above, although embodiment of this invention was described, this embodiment is shown as an example and is not intending limiting the range of invention. This embodiment can be implemented in various other forms, and various omissions, replacements, and changes can be made without departing from the spirit of the invention. This embodiment and its modifications are included in the scope of the present invention and the gist thereof, and are also included in the invention described in the claims and the equivalent scope thereof.

DESCRIPTION OF SYMBOLS 1 ... Document memory | storage part 2 ... Category memory | storage part 3 ... Document classification | category part 31 ... Feature degree calculation part 32 ... Feature degree correction | amendment part 33 ... Feature word category generation part 34 ... Trend vector generation part 35 ... Clustering part 36 ... Category generation process part DESCRIPTION OF SYMBOLS 4 ... Feature data storage part 5 ... User interface part 51 ... 2-axis map display part 52 ... Category operation part 53 ... Interesting word setting part 100 ... Document classification apparatus

Claims (5)

A document storage unit for storing document data;
A category storage unit that stores a hierarchical structure of categories and a classification rule for classifying the document data into the categories;
A category that receives an input of a category as a classification viewpoint and a target category that is a classification target category, and reads a set of axis categories that are lower categories of the category as a classification viewpoint from the category storage unit as an axis category set An operation unit;
Of the document data stored in the document storage unit, a set of the document data satisfying the classification rule of the target category is set as a target document data set, and a feature degree of a word included in the target document data set is calculated. A feature calculation unit;
The target document satisfying the classification rule of each axis category in the axis category set for each of the selected words is selected based on the characteristic degree calculated by the characteristic degree calculation unit. A trend vector generation unit that calculates a statistic based on the appearance frequency of the word in the data set, and generates a trend vector in which the statistic is set as a value of an element corresponding to the axis category;
A clustering unit that clusters the words based on the similarity of the trend vectors generated by the trend vector generation unit;
For each cluster obtained as a result of clustering by the clustering unit, a feature word category having a classification rule using the target category as a higher category and a word belonging to the cluster as a filter word is generated in the category storage unit A category generation processing unit to be registered;
Of the document data stored in the document storage unit, each cell of the biaxial map having the axis category as the first axis classification item and the feature word category as the second axis classification item A biaxial map display unit for displaying information representing the number of the document data satisfying the classification rule of the axis category corresponding to and the classification rule of the feature word category corresponding to the cell;
A document classification apparatus comprising: The biaxial map display section
When the filter word used in the classification rule of the feature word category is displayed and the filter word selected from the displayed filter words is received, the filter word is stored in the document storage unit. Displaying information indicating the number of the document data satisfying the classification rule of the axis category and including the selected filter word among the document data.
The document classification apparatus according to claim 1, wherein: The biaxial map display section
When the filter word used in the classification rule of the feature word category is displayed and the filter word is edited with respect to the feature word category, the classification rule of the feature word category subjected to the editing operation is displayed. Based on the editing operation,
Of the document data stored in the document storage unit in each cell of the two-axis map, the classification rule of the axis category corresponding to the cell and the characteristic word category corresponding to the cell are changed. Displaying information indicating the number of document data satisfying the classification rule;
The document classification device according to claim 1, wherein the document classification device is a document classification device. A focused word setting unit that receives an input of a focused word that is a word focused on when classifying the document data;
A feature correction unit that corrects the feature of the word calculated by the feature calculation unit based on the co-occurrence of the word and the word of interest in the target document data set;
The trend vector generation unit selects the word that represents the feature of the document based on the feature degree corrected by the feature degree correction unit, and for each of the selected word, each axis category in the axis category set. Calculating a statistic based on the appearance frequency of the word in the target document data set satisfying the classification rule, and generating a trend vector in which the statistic is set as a value of an element corresponding to the axis category;
The document classification device according to any one of claims 1 to 3, wherein the document classification device is characterized in that: The biaxial map display unit receives an input of a horizontal axis category and a vertical axis category, reads out a horizontal axis category which is a lower category of the category which is a horizontal axis from the category storage unit, and a vertical axis The vertical axis category which is a lower category of the above category is read out, the horizontal axis category is read out when the horizontal axis category is read out, and the horizontal axis category is read out when the horizontal axis category is not read out. As a column item, if the vertical axis category is read, the vertical axis category is used. If not, the vertical axis category is used as a row item. Information indicating the number of the document data satisfying the classification rule of the category corresponding to the column item of the cell and the classification rule of the category corresponding to the item of the cell row Not shown,
The category operation unit receives an input as to which of the category corresponding to the item of the row and the category corresponding to the item of the column to be a category or a target category as a classification viewpoint,
The document classification apparatus according to claim 1, wherein the document classification apparatus is a document classification apparatus.

Images