JPH11296552A - Document classification device, document classification method, and computer-readable recording medium recording a program for causing a computer to execute the method - Google Patents

Abstract

(57) [Summary] [Problem] To classify documents based on similarity between documents, it is an object of the present invention to efficiently and repeatedly perform document classification reflecting the intention of an operator in a short time. SOLUTION: An input unit 401 for inputting document data,
Analysis unit 4 that analyzes input document data and obtains analysis information
02 and a vector generation unit 403 that generates a document feature vector for the document data based on the obtained analysis information.
And a conversion function calculation unit 404 that calculates an expression space conversion function for projecting the generated document feature vector into a space reflecting the similarity between the document feature vectors, and a calculated expression space conversion function. A vector conversion unit 405 that converts the document feature vector generated by the vector generation unit 403, a classification unit 406 that classifies the document based on the similarity between the converted document feature vectors, and a result of the classified document classification. And a classification result storage unit 407 for storing.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a document classification device for classifying documents based on similarity between documents, a document classification method, and a computer-readable recording medium recording a program for causing a computer to execute the method.

[0002]

2. Description of the Related Art In recent years, the spread of the Internet and the like has made it possible to access a large amount of document information, classify a large amount of collected document information into meaningful categories, and grasp the structure of a group of documents. Is being carried out. When the operator manually classifies a large amount of document information, human / time costs become enormous, and only the operator who classifies the knowledge used for performing the classification is required. If the operator in charge of the classification changes, the classification standard will also change.

[0003] Therefore, an important issue is how to automatically classify a group of documents according to a classification criterion that a human classifies. That is, documents that are semantically similar are classified into the same category,
It is desired that a document classifying device configured so that each classification category generated in the classifying step is a classification category intended by an operator before executing a class is performed.

[0004] The prior art of an automatic document classification apparatus includes:
For example, as described in Japanese Patent Application Laid-Open No. H7-36897, a document is regarded as a document vector characterized by words, these document vectors are grouped using a clustering method, and based on the grouped document vectors, Some documents perform automatic classification of documents.

[0005]

However, the document classification apparatus of the prior art uses a document feature vector having a word included in a document to be classified as a feature amount, and applies a clustering method to the document feature vector. Therefore, there is a problem that it is difficult to obtain a classification result reflecting the semantic relevance of the document due to the polysemy / synonymity of the word.

As a solution to the word ambiguity / synonymous problem, as described in US Pat. No. 4,839,853, a singular value decomposition is applied to an inner product matrix between documents. In other words, into a space called a latent semantic space generated based on the co-occurrence of words between documents,
A document search that reflects semantic relevance by projecting a document and a word is performed.

[0007] Also, "Projections for
Efficient Document Cluster
ering (author name: Hinrich Schutze)
and Craving Silverstein,
Conference name: ACM, Paper name: Proceedings
of SIGIR, Page: 74-81, Publication year:
1997)), document classification is performed in the latent semantic space. Furthermore, "Rep
presenting Documents Usin
g an Explicit Model of Th
eir Similiaries (Author: Bria)
nT. Bartell, Garrison W.
Cottrell, and Richard K.
Berew, Article name: Journal of the
e American Society for In
formation Science, Society name: th
e American Society for In
formation Science, page: 25
4-271, Vol. 46 No. 4, Publication year: 19
95)), a generalization of the above-described conversion method to a latent semantic space, and using a matrix obtained by adding co-reference information generated from reference information to another document included in a document to a dot product matrix between documents. In some cases, an expression space conversion function for projecting a document or word into a space reflecting these similarities is derived.

Each dimension of the projection space generated by these conventional techniques is a conceptual one in which a plurality of words are semantically combined. Which feature dimension is used to perform document classification or document search? Is determined based only on the magnitude of the singular value calculated when applying the singular value decomposition.
For this reason, it is difficult to reflect the intention of the operator in selecting the feature dimension used at the time of performing the classification, and there is a problem that the classification result is different from the result intended by the operator. .

In another conventional document classification method, in order to perform document classification reflecting the semantic relevance of a document, a part for calculating a representation space conversion function for converting a document and a part for actually calculating the expression space conversion function are used. Although the part that performs document classification of the document converted using the expression space conversion function is continuously processed, the part that calculates the expression space conversion function is processing that requires a lot of calculation time. There is a problem that the time required for one document classification becomes enormous.

According to the present invention, in order to solve the above-mentioned problems of the prior art, a document classification apparatus, a document classification method, and a document classification method capable of efficiently and efficiently repeating a document classification reflecting the intention of an operator in a short time. It is an object of the present invention to provide a computer-readable recording medium on which a program for causing a computer to execute the above is recorded.

[0011]

Means for Solving the Problems The above-mentioned problems are solved,
In order to achieve the above object, a document classification device according to the present invention comprises an input unit for inputting document data, an analysis unit for analyzing the document data input by the input unit to obtain analysis information, and the analysis unit. And a vector generation unit that generates a document feature vector for the document data based on the analysis information obtained by the method. The document feature vector generated by the vector generation unit is projected onto a space that reflects the similarity between the document feature vectors. A conversion function calculating means for calculating an expression space conversion function for converting, and a vector conversion means for converting the document feature vector generated by the vector generation means using the expression space conversion function calculated by the conversion function calculation means. A classifier for classifying a document based on the similarity between the document feature vectors converted by the vector conversion means. When, characterized in that and a classification result storing means for storing the result of the classified documents classified by the classifying means.

According to the first aspect of the present invention, each document is converted into an expression space which can reflect the semantic relevance between the documents based on the similarity between the documents in the group of documents to be classified. By calculating an expression space conversion function for performing the operation and classifying the documents in the expression space, it is possible to realize a document classification that can reflect the intention of the operator.

The document classification device according to a second aspect of the present invention is characterized in that, in the first aspect of the present invention, the document classification device further comprises an inner product calculating means for calculating an inner product between the document feature vectors generated by the vector generating means, and Means for calculating a representation space conversion function using the inner product calculated by the inner product calculation means.

According to the second aspect of the present invention, the inner product between the document feature vectors is used as the similarity between the documents required for deriving the expression space conversion function, so that the semantic relation between the documents is obtained. It is possible to classify documents that reflect the characteristics.

According to a third aspect of the present invention, there is provided the document classification device according to the second aspect of the present invention, wherein an inter-document similarity information of the document data such as a creator and a creation date of the document input by the input means is set. A similarity information setting unit, wherein the conversion function calculation unit calculates an expression space conversion function using the inner product calculated by the inner product calculation unit and the inter-document similarity information set by the inter-document similarity information setting unit. It is characterized by the following.

According to the third aspect of the present invention, in addition to the inner product between the document feature vectors, the similarity between the documents required for deriving the expression space conversion function, the document creator and the document creation date and the like can be obtained. By using the inter-similarity information, it is possible to perform document classification reflecting the semantic relevance between documents.

According to a fourth aspect of the present invention, there is provided the document classification apparatus according to the first to third aspects, further comprising: a vector storage unit for storing the document feature vector generated by the vector generation unit; And a conversion function storage means for storing the expression space conversion function calculated by (1).

According to the fourth aspect of the present invention, by storing the document feature vector to be calculated and the expression space conversion function, the part for calculating the expression space conversion function is actually converted using the expression space conversion function. Since the document classification process is performed separately from the document classification using the document, the document classification can be performed without calculating the expression space conversion function each time, and further, the expression using the document feature vector conversion unit can be used. As a space conversion function, an expression space conversion function generated in advance based on another document feature vector can be used, so that iterative execution of document classification can be performed efficiently in a short time.

According to a fifth aspect of the present invention, in the document classification apparatus according to any one of the first to fourth aspects, before the document converting unit changes the document feature vector by the vector converting unit, Vector modifying means for modifying the document feature vector by manipulating the document feature vector and / or the feature dimension constituting the document feature vector using a rule constituted by the characteristics of the words. I do.

According to the fifth aspect of the present invention, when the document classification is repeatedly executed, the document feature vectors and the feature dimensions constituting them are operated for each individual classification execution, so that the classification differs for each classification. It is possible to change the range of the document to be classified or change the space in which the classification is performed, such as executing the document classification by deleting words.

According to a sixth aspect of the present invention, in the document classification apparatus according to the fifth aspect, when the feature dimension is changed by correcting the document feature vector in the vector correcting means, the changed feature dimension is changed. And a conversion function correction unit that corrects the expression space conversion function calculated by the conversion function calculation unit so that the vector conversion unit can appropriately convert the document feature vector.

According to the present invention, when the expression space conversion function is calculated based on the inner product of the document feature vector, the document is classified using the document converted using the expression space conversion function. In this part, the inconsistency of the expression space conversion function caused when the document feature vector or its feature dimension is manipulated can be easily corrected, so that more appropriate conversion of the document feature vector can be performed.

According to a seventh aspect of the present invention, in the document classification apparatus according to any one of the first to fifth aspects, further, there is provided a conversion function correction instructing means for instructing operation of a feature dimension of the expression space conversion function. And a conversion function correction unit that corrects the expression space conversion function based on the instruction content regarding the operation of the feature dimension specified by the conversion function correction instruction unit.

According to the seventh aspect of the present invention, a document classification that can reflect the intention of the operator by allowing the operator to perform a simple operation on the feature dimension of the space formed by using the expression space conversion function. It is possible to do.

The document classification device according to claim 8 is the invention according to claim 7, wherein the instruction content regarding the operation of the characteristic dimension instructed by the conversion function modification instructing means uses arbitrary document vector data. It is characterized in that the feature dimension of the expression space conversion function is operated.

According to the eighth aspect of the present invention, the feature dimension of the space formed by using the expression space conversion function is simplified by using arbitrary document vector data other than the classification target specified by the operator. By performing appropriate operations, it is possible to perform document classification that can reflect the intention of the operator.

According to a ninth aspect of the present invention, there is provided the document classification apparatus according to the seventh aspect, wherein the instruction content relating to the operation of the feature dimension instructed by the conversion function modification instructing means uses the document feature vector to represent the expression space. It is characterized in that the feature dimension of the conversion function is operated.

According to the ninth aspect of the present invention, a simple operation using a document feature vector designated by an operator is performed on a feature dimension of a space formed by using the expression space conversion function. It is possible to classify documents that can reflect the intention of the operator.

The document classification device according to claim 10 is
8. The invention according to claim 7, wherein the instruction content related to the operation of the characteristic dimension specified by the conversion function correction instruction unit operates the characteristic dimension of the expression space conversion function using the analysis information obtained by the analysis unit. It is characterized by being.

According to the tenth aspect of the present invention, the feature dimension of a space formed by using the expression space conversion function is
By performing a simple operation using the analysis information instructed by the operator, it is possible to perform document classification that can reflect the intention of the operator.

Further, the document classification device according to claim 11 is:
8. The invention according to claim 7, wherein the instruction content relating to the operation of the characteristic dimension instructed by the transformation function correction instructing means operates the characteristic dimension of the expression space transformation function using the classification result stored by the classification result storage means. It is characterized by that.

According to the eleventh aspect of the present invention, the feature dimension of the space formed by using the expression space conversion function is
By performing a simple operation using the pre-classified classification result instructed by the operator, it is possible to perform document classification that can reflect the operator's intention.

Further, the document classification device according to claim 12 is:
12. The classification according to any one of claims 1 to 11, further comprising: an initial center of gravity specifying unit that specifies an initial cluster center of gravity; and an initial center of gravity registration unit that registers an initial cluster center of gravity specified by the initial center of gravity specifying unit. The means classifies documents according to the initial cluster centroid registered by the initial centroid registration means.

According to the twelfth aspect of the present invention, a non-hierarchical clustering method is used as the document classification method.
The operator can arbitrarily specify the initial cluster centroid required at that time, and the document is classified according to the specified initial cluster centroid, so that the document classification reflecting the operator's intention can be performed. It is possible.

According to a thirteenth aspect of the present invention, there is provided a document classification device comprising:
In the twelfth aspect, arbitrary document vector data is specified as the initial cluster centroid specified by the initial centroid specifying means.

According to the thirteenth aspect, a non-hierarchical clustering method is used as a document classification method.
Since any document other than the classification target can be used as the initial cluster centroid required at that time, it is possible to perform document classification reflecting the intention of the operator.

The document classification device according to claim 14 is:
A twelfth aspect of the invention is characterized in that a document feature vector is specified as an initial cluster centroid specified by the initial centroid specifying means.

According to the fourteenth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method.
Since the document feature vector can be used as the initial cluster gravity center required at that time, it is possible to perform document classification reflecting the intention of the operator.

The document classification device according to claim 15 is:
In the twelfth aspect of the invention, the analysis information obtained by the analysis means is designated as the initial cluster gravity center designated by the initial gravity center designation means.

According to the fifteenth aspect, a non-hierarchical clustering method is used as a document classification method.
Since the analysis information such as words obtained as a result of applying the document to be classified to the document analysis unit can be used as the initial cluster centroid required at that time, document classification reflecting the intention of the operator should be performed. Is possible.

The document classification device according to claim 16 is:
The invention according to claim 12, wherein the classification result stored by the classification result storage means is specified as the initial cluster center of gravity specified by the initial center of gravity specification means.

According to the sixteenth aspect of the present invention, a non-hierarchical clustering method is used as the document classification method.
Since the classification result preliminarily classified can be used as the initial cluster centroid required at that time, it is possible to perform document classification reflecting the intention of the operator.

Further, according to a seventeenth aspect of the present invention, there is provided a document classification method comprising:
A first step of inputting document data; a second step of analyzing the document data input in the first step to obtain analysis information; and a document for the document data based on the analysis information obtained in the second step. Third to generate feature vector
And a fourth step of calculating an expression space conversion function for projecting the document feature vector generated in the third step onto a space reflecting the similarity between the document feature vectors,
A fifth step of converting the document feature vector generated in the third step using the expression space conversion function calculated in the fourth step, and a similarity between the document feature vectors converted in the fifth step. A sixth step of classifying documents based on the sixth step and a seventh step of storing a result of document classification classified by the sixth step classifying means.

According to the seventeenth aspect of the present invention, each document is converted into an expression space capable of reflecting the semantic relevance between the documents based on the similarity between the documents in the group of documents to be classified. By calculating an expression space conversion function for performing the operation and classifying the documents in the expression space, it is possible to realize a document classification that can reflect the intention of the operator.

The document classification method according to the eighteenth aspect is characterized in that:
18. The invention according to claim 17, further comprising an eighth step of calculating an inner product between the document feature vectors generated in the third step, wherein the fourth step uses the inner product calculated in the eighth step to represent the expression space. It is characterized in that a conversion function is calculated.

According to the eighteenth aspect of the present invention, the inner product between the document feature vectors is used as the similarity between the documents required for deriving the expression space conversion function, so that the semantic relation between the documents is obtained. It is possible to classify documents that reflect the characteristics.

The document classification method according to claim 19 is
The invention according to claim 18, further comprising a ninth step of setting inter-document similarity information of document data such as a creator and a creation date of the document input in the first step, and wherein the fourth step is the eighth step. The expression space conversion function is calculated using the inner product calculated by the above and the inter-document similarity information set in the ninth step.

According to the nineteenth aspect of the present invention, in addition to the inner product between the document feature vectors, the similarity between the documents required for deriving the expression space conversion function, the document creator and the document creation date, etc. By using the inter-similarity information, it is possible to perform document classification reflecting the semantic relevance between documents.

Further, according to a twentieth aspect of the invention, there is provided a document classification method comprising:
20. The method according to claim 17, further comprising: storing a document feature vector generated in the third step, and storing an expression space conversion function calculated in the fourth step. And 11 steps.

According to the twentieth aspect of the present invention, by storing the document feature vector to be calculated and the expression space conversion function, the part for calculating the expression space conversion function is actually converted using the expression space conversion function. Since the document classification process is performed separately from the document classification using the document, the document classification can be performed without calculating the expression space conversion function each time, and further, the expression using the document feature vector conversion unit can be used. As a space conversion function, an expression space conversion function generated in advance based on another document feature vector can be used, so that iterative execution of document classification can be performed efficiently in a short time.

Further, according to the document classification method of claim 21,
21. The invention according to any one of claims 17 to 20, further comprising, before changing the document feature vector in the fifth step, using a rule constituted by characteristics of words extracted in the second step. Modifying the document feature vector by manipulating a document feature vector and / or a feature dimension comprising the document feature vector;
It is characterized by including two steps.

According to the twenty-first aspect of the present invention, when the document classification is repeatedly executed,
By manipulating the document feature vectors and the feature dimensions that compose them, change the range of the document to be classified or change the space in which the classification is performed, such as deleting the different words for each classification and performing document classification. Is possible.

The document classification method according to claim 22 is characterized in that:
22. The invention according to claim 21, wherein when the feature dimension is changed by correcting the document feature vector in the twelfth step, the document feature vector can be appropriately converted in the fifth step by the changed feature dimension. And a thirteenth step of correcting the expression space conversion function calculated in the fourth step.

According to the twenty-second aspect, when the expression space conversion function is calculated based on the inner product of the document feature vectors, the document is classified using the document converted using the expression space conversion function. In this part, the inconsistency of the expression space conversion function that occurs when the document feature vector or its feature dimension is manipulated can be easily corrected, so that more appropriate conversion of the document feature vector can be performed.

The document classification method according to claim 23 is:
22. The invention according to claim 17, further comprising: a fourteenth step of giving an instruction relating to the operation of the characteristic dimension of the representation space transformation function; and an instruction content relating to the operation of the characteristic dimension designated by the fourteenth step. A fifteenth step of modifying the expression space conversion function based on the expression space conversion function.

According to the twenty-third aspect of the present invention, a document classification that can reflect the operator's intention by allowing the operator to perform a simple operation on the feature dimension of the space formed by using the expression space conversion function. It is possible to do.

According to a twenty-fourth aspect of the present invention, there is provided a document classification method comprising:
24. The invention according to claim 23, wherein the instruction content relating to the operation of the characteristic dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using arbitrary document vector data. I do.

According to the twenty-fourth aspect of the present invention, the feature dimension of a space formed by using the expression space conversion function is simplified by using arbitrary document vector data other than the classification target specified by the operator. By performing appropriate operations, it is possible to perform document classification that can reflect the intention of the operator.

Further, the document classification method according to claim 25,
23. The invention according to claim 23, wherein the instruction content regarding the operation of the characteristic dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using a document characteristic vector.

According to the twenty-fifth aspect of the present invention, a simple operation using a document feature vector specified by an operator is performed on a feature dimension of a space formed by using the expression space conversion function. It is possible to classify documents that can reflect the intention of the operator.

The document classification method according to claim 26 is characterized in that:
24. The invention according to claim 23, wherein the instruction content related to the operation of the characteristic dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using the analysis information obtained in the second step. There is a feature.

According to the twenty-sixth aspect of the present invention, the feature dimension of a space formed by using the expression space conversion function is
By performing a simple operation using the analysis information instructed by the operator, it is possible to perform document classification that can reflect the intention of the operator.

The document classification method according to claim 27 is characterized in that:
24. The invention according to claim 23, wherein the instruction content related to the operation of the characteristic dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using the classification result stored in the seventh step. There is a feature.

According to the twenty-seventh aspect of the present invention, the feature dimension of a space formed using the expression space conversion function is
By performing a simple operation using the pre-classified classification result instructed by the operator, it is possible to perform document classification that can reflect the operator's intention.

A document classification method according to claim 28 is characterized in that:
28. The method according to any one of claims 17 to 27, further comprising: a sixteenth step of specifying an initial cluster centroid; and a seventeenth step of registering the initial cluster centroid specified by the sixteenth step. The document is classified according to the initial cluster centroid registered in the seventeenth step.

According to the twenty-eighth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method.
The operator can arbitrarily specify the initial cluster centroid required at that time, and the document is classified according to the specified initial cluster centroid, so that the document classification reflecting the operator's intention can be performed. It is possible.

The document classification method according to claim 29 is characterized in that:
The invention according to claim 28, wherein arbitrary document vector data is specified as the initial cluster barycenter specified in the sixteenth step.

According to this invention, a non-hierarchical clustering method is used as a document classification method.
Since any document other than the classification target can be used as the initial cluster centroid required at that time, it is possible to perform document classification reflecting the intention of the operator.

A document classification method according to claim 30 is characterized in that:
A twenty-eighth aspect of the present invention is characterized in that a document feature vector is specified as the initial cluster centroid specified in the sixteenth step.

According to the thirtieth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method.
Since the document feature vector can be used as the initial cluster gravity center required at that time, it is possible to perform document classification reflecting the intention of the operator.

A document classification method according to claim 31 is characterized in that:
The invention according to claim 28, wherein the analysis information obtained in the second step is specified as the initial cluster centroid specified in the sixteenth step.

According to the thirty-first aspect of the present invention, a non-hierarchical clustering method is used as a document classification method.
Since the analysis information such as words obtained as a result of applying the document to be classified to the document analysis unit can be used as the initial cluster centroid required at that time, document classification reflecting the intention of the operator should be performed. Is possible.

A document classification method according to claim 32 is characterized in that:
The invention according to claim 28, wherein the classification result stored in the seventh step is specified as the initial cluster barycenter specified in the sixteenth step.

According to this invention, a non-hierarchical clustering method is used as a document classification method.
Since the classification result preliminarily classified can be used as the initial cluster centroid required at that time, it is possible to perform document classification reflecting the intention of the operator.

The storage medium according to the invention of claim 33 records a program for causing a computer to execute the method according to claims 17 to 32, so that the program becomes machine-readable. The operations of Items 17 to 32 can be realized by a computer.

[0076]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Referring to the accompanying drawings, preferred embodiments of a document classifying apparatus, a document classifying method, and a computer-readable recording medium storing a program for causing a computer to execute the method are described below. The form will be described in detail.

(Embodiment 1) First, the hardware configuration of the entire information processing system constituting the document classification device according to Embodiment 1 of the present invention will be described. FIG. 1 is an explanatory diagram showing the hardware configuration of the entire information processing system that constitutes the document classification device according to the first embodiment.

In FIG. 1, the information processing system constituting the document classification device according to the first embodiment is configured by a server / client system. That is, server 1
01 and a plurality of clients 102
Connected by The client 102 generates classification data, instructs the server 101, displays a classification result, and the like. On the other hand, in accordance with an instruction from the client 102, the server 101 performs processing relating to document (text) classification by enormous numerical operations, and sends the result of the processing to the client 102.

More specifically, in the server 101, text classification processing (preprocessing, clustering processing)
The client 102 performs classification data generation, processing execution instruction, text classification result display, and the like. As described above, the processing in the server 101 is divided into two, that is, “pre-processing” and “classification processing”, and the processing becomes extremely heavy depending on data. Therefore, in the server 101, the manager process creates and manages the processing reception list so that only one of each of the “preprocessing” and the “classification processing” is performed. Details of the "pre-processing" and the "classification processing" will be described later.

The server 101 and the client 10
The exchange of data between the two uses a method called file sharing. That is, by creating a file used for the classification process in the shared folder on the server 101, the two exchange data. Therefore, it is possible for the client 102 to use the shared folder of the server 101 by network sharing.

Next, the hardware configuration of the server 101 and the client 102 will be described. FIG.
FIG. 3 is an explanatory diagram showing the hardware of the server 101 in the information processing system constituting the document classification device according to the first embodiment. The server 101 is, for example, a workstation (WS).

In FIG. 2, 201 is a CPU for controlling the entire server 101, 202 is a ROM storing a boot program and the like, 203 is a RAM 203 used as a work area of the CPU 201, and 204 is a communication line 205. An interface (I / F) that is connected to the network 103 and controls an internal interface with the network 103, and a disk device 206 stores data. Reference numeral 200 denotes a bus for connecting the above components.

In addition, a display 208 for displaying document information, image information, function information, and the like, a keyboard 209 and a mouse 210 for inputting data, and the like may be similarly connected. Further, the disk device 206 is provided with a shared folder 207 for exchanging data with the client 102.

FIG. 3 is an explanatory diagram showing the hardware of the client 102 in the information processing system constituting the document classification device according to the first embodiment. The client 102 is, for example, a personal computer (P
C) and the like are used.

In FIG. 3, reference numeral 301 denotes a CPU for controlling the entire system; 302, a ROM storing a boot program and the like; 303, a RAM used as a work area of the CPU 301; A hard disk drive (HDD) 305 for controlling the reading / writing of data to / from a hard disk 305, an HD 305 for storing data written under the control of the HDD 304, and a data 306 for a floppy disk (FD) 307 under the control of the CPU 301. 307 is a floppy disk drive (FDD) for controlling read / write of a disk, and 307 is a removable FD for storing data written under the control of the FDD 306.
308 denotes a display for displaying documents, images, function information, and the like.

Reference numeral 309 is connected to the network 103 via the communication line 310, and the network
And the interface that controls the internal interface (I
/ F), 311 is a keyboard having keys for inputting characters, numerical values, various instructions, etc., 312 is moving a cursor, selecting a range, pressing an icon or button displayed on a display screen, or opening a window. A mouse 313 for moving and changing the size is an OCR (Optical).
A scanner 314 for optically reading an image having a character reader (Character Reader) function, a printer 314 for printing the contents of data including a classification result, and the like, and a bus 315 for connecting the above-described units are shown. The HD 305 stores application software 316 such as word processing software or spreadsheet software.

Next, the functional configuration of the document classification device according to the first embodiment will be described. 4 to 6 are block diagrams functionally showing the configuration of the document classification device according to the first embodiment. In FIG. 4, the document classification device includes an input unit 4
01, an analysis unit 402, a vector generation unit 403, a conversion function calculation unit 404, a vector conversion unit 405, a classification unit 406, and a classification result storage unit 407.

Further, between the input unit 401 and the analysis unit 402, a first filter unit (not shown) for absorbing fluctuations of the notation in the document data may be included. Also,
A second filter unit (not shown) that removes unnecessary words and parts of speech from the analysis information may be included between the analysis unit 402 and the vector generation unit 403. Further, a third filter unit (not shown) for removing unnecessary words and parts of speech from the document feature vector at the time of classification may be included between the conversion function calculation unit 404 and the vector conversion unit.

FIG. 5 further includes an inner product calculation section 421. In FIG. 6, the configuration further includes an inter-document similarity information setting unit 431.

The input unit 401 is for inputting document data. For example, the keyboard 209 or 311,
A scanner 313, a scanner 313 having an OCR function,
Or an I / F 204 or 309 that can obtain a document or a group of documents via the network 103. In addition to the above, the input unit 401 includes all of them that can acquire document data.
For example, if the document data is stored in a database and the medium on which the database is recorded is incorporated in the document classification device of the present embodiment, the input of the document data is also performed. Further, a document data storage unit (not shown) for storing the input document data may be included.

Here, a document is a group of one or more sentences described in a natural language, and when it is to be classified, it is called a document. Specifically, a published patent publication and a specific newspaper article are also documents, and even if a claim or a specific sentence is extracted, it is regarded as a document.

The analysis unit 402 analyzes words of the document data input by the input unit 401 to obtain analysis information. Specifically, a natural language analysis such as a morphological analysis is performed on each of the document data input by the input unit 401 to extract a word and its part of speech. Furthermore, a unique word ID is assigned to a word that has appeared in a document group, and the number of word appearances in the document and for the document group is counted.

The vector generation unit 403 includes the analysis unit 4
A document feature vector for document data is generated based on the analysis information obtained in step S03. The conversion function calculation unit 404 calculates an expression space conversion function for projecting the document feature vector generated by the vector generation unit 403 into a space reflecting the similarity between the document feature vectors. The vector conversion unit 405 converts the document feature vector using the expression space conversion function calculated by the conversion function calculation unit 404.

The vector generation unit 403 and the conversion function calculation unit 4
04, details of each processing of the vector conversion unit 405 will be described later.

The classification unit 406 classifies documents based on the similarity between new document feature vectors converted by the vector conversion unit 405. Specifically, document classification can be performed by applying a classification method such as a chi-square method, a discriminant analysis method, or a cluster analysis method to the generated classification target data. The classification unit 406 may use any classification method as long as the vector data can be applied.

Further, the classification result storage unit 407 is a storage unit that stores the results classified by the classification unit 406 in an appropriate format. For example, in addition to a predetermined area of the disk device 306 or the hard disk 316, any of the RAM 203 or 303 and any other data storage location may be used.

The inner product calculation unit 501 includes a vector generation unit 40
The calculation unit calculates the inner product between the document feature vectors generated by the two means. The contents of the process of the inner product calculation unit 501 will be described later.

The inter-document similarity information setting unit 601 is a setting unit for setting the inter-document similarity information of the document data such as the creator of the document input by the input unit 401 and the date of creation.
The inter-document similarity information includes the order in which words appear in the document, and the matching information between documents such as the date of creation and modification of the document, the creator, the modifier, the reference document, and the cited document. The operator can specify desired information from these inter-document similarity information and set it arbitrarily.

Input unit 401, analysis unit 402, vector generation unit 403, conversion function calculation unit 404, vector conversion unit 4
05, the classification unit 406, the inner product calculation unit 420, and the inter-document similarity information setting unit 430 include the ROM 202 or 302, the RAM
The functions of the respective units are realized by the CPU 201 or 301 or the like executing the instruction processing in accordance with the instructions described in a program recorded in a recording medium such as 203 or 303, or the disk device 306 or the hard disk 316.

Next, the contents of the process of generating a document feature vector by the vector generation unit 403 will be described. The vector generation unit 403 generates a document feature vector for the document data based on the analysis information obtained by the analysis unit 403. Here, the analysis information is, for example,
Information such as a word, a word ID, the number of appearances of a word, and part of speech information.

FIG. 7 is an explanatory diagram showing an example of document-word matrix data and a document feature vector. In FIG. 7, a row component 701 is a word ID, and a column component 702 is a document ID. As a matrix element, document-word matrix data is generated based on the analysis information so that the document ID is a column number and the word ID included in the document is the number of appearances of a word having a row number. Each column vector of this document-word matrix is a document feature vector. Thus, a document feature vector is generated.

Further, processing such as normalization can be simultaneously performed on the document feature vector. On this occasion,
Additional information attached to the document-word matrix data, for example, word-word ID correspondence map data describing the correspondence between word IDs, which are row components of the document-word matrix data, and the words, and word IDs for each word Word ID-speech correspondence map data describing the correspondence between the word and the part-of-speech information of the word are also generated.

Next, the contents of the conversion function calculation processing by the conversion function calculation unit 404 will be described. The generation of the document feature vector in the vector generation unit 403 is usually performed based on the number of appearances of a word in the document. This assumes that each word is semantically independent,
Each is treated as orthogonal. However, in reality, words contain ambiguity and synonymity, so the validity of the above assumption is not guaranteed, and by treating each word as orthogonal, the accuracy of classification and
It also affects the validity.

As a technique for reducing this effect, it is conceivable that this problem is regarded as a multidimensional scale problem and a statistical method is used. That is, the conversion function calculator 40
4, the expression feature conversion function for converting each document feature vector into a feature dimension between the document feature vectors, that is, a space in which the co-occurrence of words is reflected, is represented by a document feature vector generated by the vector generation unit 403. Calculated based on Note that a thesaurus or the like may be used as a method for reducing the effect of synonymity between words.

In the present embodiment, the method of calculating the expression space conversion function is described in the above “Representantat”.
ing Documents Using an Ex
Plicit Model of Their Sim
Although the calculation method of the expression space conversion function described in “ilariities” is used, the calculation may be performed using a method such as factor analysis or quantification.

That is, based on the inter-document similarity matrix and the document feature vector obtained by adding the inter-document similarity information set by the inter-document similarity information setting means to the inner product between the document feature vectors calculated by the inner product calculation unit 501. Calculate the expression space conversion function. By using this expression space conversion function, document classification can be performed in an expression space that strongly reflects semantic similarity between documents. Also,
As described above, since the operator can also freely select similar information between documents, document classification reflecting the intention of the operator can be performed.

Specifically, the number of documents is d, the number of words is t, the size of a document-word matrix (document feature vector) of size t × d is X, the inner product matrix between documents of size d × d is S, additional documents between similar information matrix of size d × d When S _a, representation space conversion function W is as equation 1.

W = M ^T CX ⁺ (Equation 1)

Note that ^T indicates transposition of a matrix.

Here, assuming that the operator for applying the singular value decomposition to the matrix is svd (),
M and X ⁺ are the following matrices.

X = svd (X) = ULA ^T (Equation 2)

[0112] S = X ^T X (Equation 3)

S + S _a = svd (S + S _a ) = C ^T C (Equation 4)

CAA ^T = svd (CAA ^T ) = MZN ^T (Equation 5)

X ⁺ = AL ^-1 U ^T (Equation 6)

To calculate the expression space conversion function using the inner product of the vectors, the additional inter-document similarity matrix S
_{Let a} be an empty matrix. In that case, the expression space conversion function is expressed by Equation 7
It becomes like.

W = U ^T (Equation 7)

[0118] Further, in order to calculate the representation space conversion function using a document between similar information set is a symmetric matrix of non-empty matrix S _a.

Further, in the present document classification apparatus, the conversion function generation unit 404 can be bypassed by using the expression space conversion function W as a unit matrix having a size of t × t.

Furthermore, the document feature vector generated by the vector generation unit 403 is usually of a very high order because the number of feature dimensions is the number of words appearing in the document group.
If the classification or the like is performed as it is, the computation cost and storage space become enormous. For this reason, a word with an extremely small number of appearances or a word with an extremely large number of occurrences can be excluded from the dimensions constituting the document feature vector, but this may reduce the classification accuracy and validity.

Since the expression space conversion function used in the present invention realizes conversion to a space in which the co-occurrence of words between document feature vectors is taken into consideration, the expression space conversion function is used as is clear from Equation 1. In the generated expression space, each feature dimension is expressed as a linear combination of a plurality of words. Therefore, the meaning of many words can be handled with a small feature dimension,
As a result, it is possible to suppress the calculation cost and the storage space when performing the classification or the like.

Next, the conversion process of the document feature vector by the vector conversion unit 405 will be described. The vector conversion unit 405 converts the document feature vector using the expression space conversion function generated by the conversion function generation unit 404, and derives data to be classified. In addition, each word can be converted using the expression space conversion function. That is, when the matrix W is used as the expression space conversion function, if the converted document feature vector is D _p , Expression 8 is obtained.

D _p = WX (Equation 8)

Further, assuming that the converted word is represented by a matrix, T _p is expressed by the following equation (9).

T _p = W ^T I = W (Equation 9)

Note that I indicates a unit matrix.

Next, a series of processing procedures of the document classification device according to the first embodiment will be described. FIG. 8 is a flowchart showing a procedure of a series of processes of the document classification device according to the first embodiment. In the flowchart of FIG. 8, first, the input unit 401 inputs document data (Step S).
810). Next, the analysis unit 402 analyzes the document data input in step S810 to obtain analysis information (step S820).

Next, the vector generation unit 403 generates a document feature vector based on the obtained analysis information in step S820 (step S830). Next, the conversion function calculation unit 404 calculates an expression space function for projecting the document feature vector generated in step 830 into a space reflecting the similarity between the document feature vectors (step S840).

Next, the vector conversion unit 405 converts the document feature vector generated in step S830 using the expression space function calculated in step S840 (step S850). Next, the classification unit 406
Classifies documents based on the similarity between the document feature vectors converted in step S850 (step S850).
860). After that, the classification result classified in step S860 is stored (step S870), and all the processing ends.

FIG. 9 is a flowchart showing another procedure of a series of processes of the document classification device according to the first embodiment. In the flowchart of FIG. 9, steps that perform the same processing as the steps of FIG. 8 are given the same numbers, and descriptions thereof are omitted.

After step S830, the inner product between the document feature vectors generated in the step is calculated (step S835). Next, it is determined whether or not there is an instruction to use the inter-document similarity information (step S).
836).

If there is no instruction in step S836 (No in step S836), step S835 is performed.
The expression space conversion function is calculated using the inner product calculated in (step S840). On the other hand, step S8
If there is an instruction in S36 (Yes at Step S836), inter-document similarity information of the document data input by the input unit 401 is set (Step S837). Thereafter, the process proceeds to step S840, and the expression space conversion function is calculated using the inner product calculated in step S835 and the inter-document information set in step S837. Hereinafter, the same processing as in FIG. 8 is performed.

As described above, according to the first embodiment, each document can reflect the semantic relevance between the documents based on the similarity between the documents in the group of documents to be classified. By calculating an expression space conversion function for converting to an expression space and performing document classification in the expression space, it is possible to realize document classification that can reflect the intention of the operator.

(Embodiment 2) The embodiment 1 described above does not describe the storage of the generated document feature vector and the calculated expression space conversion function. As in Embodiment 2, the configuration may further include a vector storage unit and a conversion function storage unit.

The functional configuration of the document classification device according to the second embodiment will be described. FIG. 10 is a block diagram functionally showing the configuration of the document classification device according to the second embodiment.
10, the same components as those in FIG. 4 of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

The vector storage unit 1001 is a storage unit for storing the document feature vector generated by the vector generation unit 403. At this time, additional information accompanying the document-word matrix data simultaneously generated in the vector generation unit 403, for example, a word describing the correspondence between the word ID which is a row component of the document-word matrix data and the word −
Word ID-corresponding map data, word ID-part-of-speech corresponding map data that describes the correspondence between word IDs of each word, and part-of-speech information of the words, syntax information data, and the like can be stored.

The conversion function storage unit 1002 is a storage unit for storing the expression space conversion function generated by the conversion function generation unit 404.

The vector storage unit 1001 and the conversion function storage unit 1002 include the ROM 202 or 302, the RAM 203
Or 303, or the CPU 201 or 30 in accordance with instructions described in a program recorded on a recording medium such as the disk device 306 or the hard disk 316.
1 and the like execute the command processing, thereby realizing the function of each unit.

By storing the document feature vector and the expression space conversion function, it is possible to convert the stored document feature vector using the stored document expression space. Therefore, the vector storage unit 1001 and the conversion function storage unit 1
002 and the vector conversion unit 405 need not be performed as a series of processes, and can be functionally separated.

Next, a series of processing procedures of the document classification device according to the second embodiment will be described. FIG. 11 is a flowchart showing a procedure of a series of processes of the document classification device according to the second embodiment. In the flowchart of FIG. 11, steps that perform the same processing as the steps of FIG. 8 of the first embodiment are given the same numbers, and descriptions thereof are omitted.

Following the processing in step S830, the document feature vector generated in the step is stored (step S831). After that, the processing shifts to step S840, and the same processing as in the first embodiment is performed. Further, following the processing in step S840, the expression space conversion function calculated in the step is stored (step S8).
41). Hereinafter, the same processing as in the first embodiment is performed.

As described above, the document classification apparatus according to the second embodiment executes the document classification without calculating the expression space conversion function each time the classification is performed by changing the number of classifications or the classification method. Therefore, a plurality of classification results can be obtained in a short time.

Further, as the expression space conversion function used in the document feature vector conversion unit, an expression space conversion function generated in advance based on another document feature vector can be used.

(Embodiment 3) Embodiments 1 and 2
However, a vector correction unit 1201 may be further added as in Embodiment 3 described below.

First, the functional configuration of the document classification device according to the third embodiment will be described. FIG. 12 shows Embodiment 3
1 is a block diagram functionally showing the configuration of a document classification device according to the present invention. 12, the same components as those in FIG. 4 of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

Before the vector conversion unit 405 changes the document feature vector, the vector correction unit 1201 converts the document feature vector and the document feature vector using a rule constituted by the characteristics of the words extracted by the analysis unit 402. The document feature vector is modified by operating one or both of the constituent feature dimensions.

FIG. 13 is a flowchart showing the procedure of the processing contents of the vector correction unit 1201. In the flowchart of FIG. 13, the vector correction unit 1201 first
The document feature vector is read (step S1).
301) Next, by specifying the word extracted by the analysis unit 402 and the part of speech information of the word (step S1302), the feature dimension of the document feature vector for performing an operation such as deletion, that is, the document group The word ID of the word that uniquely appears is determined (step S1303).

After that, the document feature vector generated by the vector generation unit 403 and the document feature vector stored by the vector storage unit 1001 are subjected to correction operations such as deletion and synthesis for the feature dimension to be operated. (Step S1304) The document feature vector is synthesized.

[0149] The vector correction unit 1201
Alternatively, the function of each unit is realized by the CPU 201 or 301 or the like executing the instruction processing in accordance with the instruction described in the program recorded in the recording medium such as 302, the RAM 203 or 303, or the disk device 306 or the hard disk 316.

FIG. 1 shows an example of a procedure for deleting t ′ feature dimensions (that is, word IDs) from a document feature vector.
It is shown in FIG. The number of documents is d, the number of words is t, the document feature vector (document-word frequency matrix) is a matrix X of size t × d, and each row (column) is a unit matrix of size t × t corresponding to the word ID. 'If a matrix of size of × t was P _t, the document feature vector X to be modified by the modification unit 1201' deleting rows corresponding to the word ID to be deleted t is as equation 10.

X ′ = P _t X (Equation 10)

Next, a series of processing procedures of the document classification device according to the third embodiment will be described. FIG. 15 is a flowchart showing the sequence of a series of processes of the document classification device according to the first embodiment. In the flowchart of FIG. 13, steps that perform the same processing as the steps of FIG. 8 of Embodiment 1 are given the same numbers, and descriptions thereof are omitted.

Following the processing in steps S830 and S831, the document feature vector is corrected (step S83).
832). After that, the processing shifts to step S840, and the same processing as in the first embodiment is performed.

As described above, in the document classification device according to the third embodiment, even after the vector correction unit 1201 generates the document feature vector by the vector generation unit 403, it is determined that the word that is unnecessary in the text Etc. can be deleted. Furthermore, the same document feature vector is efficiently classified, but the document feature vector correction unit 1201 may delete a different word for each classification and execute the document classification. it can.

(Embodiment 4) In Embodiment 3, the vector correction unit 1201 is added. However, as in Embodiment 4 described below, the conversion function correction unit 1601 is provided together with the vector correction unit. It may be added at the same time.

First, the functional configuration of the document classification device according to the fourth embodiment will be described. FIG. 16 shows Embodiment 4
1 is a block diagram functionally showing the configuration of a document classification device according to the present invention. In FIG. 16, the same components as those in FIG. 12 of the third embodiment are denoted by the same reference numerals, and description thereof will be omitted.

In the third embodiment, the vector correction unit 1
When the document feature vector is corrected in step 201, the expression space edge function is calculated based on the document feature vector before correction. Otherwise, the effect of modifying the document feature vector may be halved. Therefore, the expression space conversion function is modified based on the modified document feature vector.

That is, the conversion function correction unit 1601 in FIG. 16 corrects the expression space conversion function to W ′. When the expression space conversion function is calculated based on the inner product of the document feature vector, the expression space conversion function is given by Expression 7. At this time, if the modified representation space conversion function is W ′,
Expression 11 is expressed using Expressions 2, 7, and 10.

W ′ = L ⁻¹ U ^T P _t X (P _t X) (Equation 11)

The conversion function correction unit 1601
Alternatively, the function of each unit is realized by the CPU 201 or 301 or the like executing the instruction processing in accordance with the instruction described in the program recorded in the recording medium such as 302, the RAM 203 or 303, or the disk device 306 or the hard disk 316.

FIG. 17 is a flowchart illustrating the sequence of a series of processes performed by the document classification device according to the fourth embodiment.
In the flowchart of FIG. 17, when the document feature vector is corrected, the expression space conversion function is also corrected (step S841). The following is the same as the processing of the third embodiment.

As described above, in the document classification device according to the fourth embodiment, since the expression space conversion function can be modified along with the modification of the document characteristic vector, more appropriate conversion of the document characteristic vector can be performed. it can.

(Embodiment 5) In Embodiment 4, the conversion function correction unit 1601 is added. However, as in Embodiment 5 described below, the conversion function correction unit 161 is added.
01, a conversion function correction instructing unit 1801 that issues a correction instruction to 01
May be added.

First, the functional configuration of the document classification device according to the fifth embodiment will be described. FIG. 18 shows Embodiment 5
1 is a block diagram functionally showing the configuration of a document classification device according to the present invention. 18, the same components as those in FIG. 4 of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

The conversion function modification instructing unit 1801 is for instructing the operation of the feature dimension of the expression space conversion function. Further, the conversion function correction unit 1802 corrects the expression space conversion function by operating the feature dimension of the expression space conversion function based on the instruction content from the conversion function correction instruction unit 1801.

The conversion function correction instructing unit 1801 and the conversion function correction unit 1802 include the ROM 202 or 302, the RAM 2
The functions of each unit are realized by the CPU 201 or 301 or the like executing the instruction processing according to the instructions described in a program recorded in a recording medium such as the disk drive 03 or 303 or the disk device 306 or the hard disk 316.

In conversion function modification instructing section 1801,
One method for performing document classification that reflects the intention of the operator is to delete or combine unnecessary feature dimensions in a space formed by the expression space conversion function or feature dimensions that have an adverse effect. Or performing an operation to emphasize a certain characteristic dimension.

However, the feature dimension of the space generated by the expression space conversion function can be considered to be a combination of a plurality of words that are semantically similar among the words extracted by the analysis unit 402. The semantic interpretation of is extremely complex and ambiguous. Therefore, it is extremely difficult to present the meaning of each feature dimension to the operator.

Therefore, the operator is caused to specify information such as documents or words having contents that he does not want to reflect on the classification or contents that he wants to emphasize, and appropriately projects them into the space constituted by the expression space edge functions. By discriminating a feature dimension having a high similarity or a low feature dimension, a feature dimension on which an operation is performed can be selected.

In this embodiment, as an example of operating the feature dimension of the expression space conversion function, an example is shown in which a feature dimension having a high similarity to a certain document specified by the operator is deleted. A document specified by the operator is represented by a vector having the same dimension as the document feature vector, and an expression space conversion function is applied to the document vector to project the document vector into a space formed by the expression space conversion function. Then, by calculating the similarity between the projected document vector and each feature dimension, a feature dimension having a high similarity is determined.

At this time, as a scale for measuring the similarity, a cosine scale, an inner product scale, a Euclidean distance scale, or the like can be used. As for the discrimination, it is also possible to use a discrimination by a threshold value process in which a certain degree of similarity or more is adopted as a deletion target, or a constant process or discrimination analysis in which a certain number of similarities are adopted as deletion targets in descending order of similarity.

In this way, the expression space conversion function can be corrected by deleting the adopted feature dimension from the expression space conversion function. At this time, the operator instructs (specifies)
Any information can be applied as long as the information has the same number of dimensions as the document feature vector.

As information specified by the operator, in addition to the information that the operator can easily understand, information other than the group of documents to be classified is expressed in a vector having the same dimension as the document feature vector. Can be.
In addition, a document feature vector can be used as information specified by the operator.

In addition to the information specified by the operator, words or word parts of speech information extracted by the analysis unit 402 or manually input by the operator can be used. In addition, as the information specified by the operator, a classification representative value that is a classification result performed in advance and stored in the classification result storage unit 407 can be used.

The above-mentioned instruction information may be used alone or may be used by appropriately combining them.

FIG. 19 is a flowchart illustrating a part of a series of processes of the document classification device according to the fifth embodiment. In the flow chart of FIG. 19, first, there is an instruction to correct the conversion function (step S1901)
Next, the content of the instruction, that is, the instruction information instructed (specified) by the operator is input (step S1).
902). When there are a plurality of instructions, the same steps are repeated until all the instructions are completed, and when the instructions are completed (Yes at Step S1903), the conversion function is corrected based on the input instruction information (Step S1903). S1904), terminates all the processing.

As described above, according to the fifth embodiment, the intention of the operator is reflected by the operator performing a simple operation on the feature dimension of the space formed by using the expression space conversion function. Possible document classification can be performed.

(Embodiment 6) Embodiments 1 to 5
However, as in the sixth embodiment described below, an initial center of gravity specifying unit 2001 and an initial center of gravity registration unit 2002 may be further added.

First, the functional configuration of the document classification device according to the sixth embodiment will be described. FIG. 20 shows Embodiment 6
1 is a block diagram functionally showing the configuration of a document classification device according to the present invention. 20, the same elements as those in FIG. 4 of the first embodiment are denoted by the same reference numerals, and description thereof will be omitted.

The initial center of gravity specifying unit 2001 is a specifying unit for specifying an initial cluster center of gravity. Initial center of gravity registration unit 2002
Is a registration unit for registering the initial cluster centroid specified by the initial centroid specifying unit 2001. Classification unit 4
05 classifies documents according to the initial cluster centroid registered by the initial centroid registration unit 2002.

The initial center-of-gravity designation section 2001 and the initial center-of-gravity registration section 2002 include a ROM 202 or 302 and a RAM 203.
Or 303, or the CPU 201 or 30 in accordance with instructions described in a program recorded on a recording medium such as the disk device 306 or the hard disk 316.
1 and the like execute the command processing, thereby realizing the function of each unit.

Normally, the classification criteria used when classifying documents using the chi-square method, the discriminant analysis method, the cluster analysis method, and the like are configured based on statistical theory. However, in the present embodiment, the final evaluation of the quality of the classification at the time of classifying the document is not a statistical numerical evaluation but a subjective evaluation by an operator who analyzes the classification result. Therefore,
In the various methods for classifying the documents, by providing a room where the operator can intervene, the intention of the operator can be reflected in the classification result, and as a result, the quality of the classification result is expected to be improved.

Next, FIG. 21 is a flowchart illustrating a part of a series of processes of the document classification device according to the sixth embodiment. The non-hierarchical clustering method generally has a processing procedure as shown in the flowchart of FIG.
In the flowchart of FIG. 21, first, an initial cluster centroid is specified (step S2101), and the initial cluster centroid is registered (step S2102). Next, an initial cluster barycenter is determined (step S210).
3) The similarity between the cluster centroid and each classification target data is calculated (step S2104).

Next, each classification target data is assigned to the cluster having the highest similarity (step S2105), and the cluster centroid is calculated based on the classification target data assigned to each cluster (step S2106).

At this point, it is determined whether the repetition stop criterion is satisfied (step S2107). If the repetition stop criterion is not satisfied (step S2107: No), step S2107 is executed.
The process proceeds to steps S2104 to S21.
Steps 06 are repeatedly executed. Step S21
07, when the repetition stop criterion is satisfied (step S
2107 affirmative) ends all the processing.

It is said that the classification result strongly depends on what initial cluster centroid is selected. Therefore, a non-hierarchical clustering method such as the k-means method is used as a classification method in the classification execution unit, and the operator specifies the initial cluster centroid, thereby enabling the operator to intervene in the classification procedure. Thus, document classification reflecting the intention of the operator can be realized.

It should be noted that if the classification method is such that the similarity between each document feature vector and the center of gravity vector of the cluster is calculated and the document feature vector is assigned to the classification representative value having the highest similarity in each feature vector, Techniques other than non-hierarchical clustering can also be used. As a similarity measure for measuring the similarity between the centroid vector of the cluster and the document vector, a cosine measure, an inner product measure, a Euclidean distance measure, a Mahalanobis distance measure, or the like can be used.

An initial center of gravity designating section 2001 inputs a plurality of arbitrary document vectors having the same number of characteristic dimensions as the classification target data as an initial center of gravity for clustering. The arbitrary document vector can be specified by the operator, or the operator can indirectly specify the document vector by selecting a rule constructed based on the document feature vector to be classified.

The arbitrary document vector includes:
Any vector format having the same number of dimensions as the document feature vector can be applied. In addition, as an arbitrary document vector, a vector expressing a document other than the group of documents to be classified into a vector having the same dimension as the document feature vector can be used as an easier-to-understand one.

In addition, a document feature vector can be used as an arbitrary document vector. In addition, a word or word part of speech information extracted by the analysis unit 402 can be used as an arbitrary document vector. In addition, as an arbitrary document vector,
A classification representative value, which is a classification result performed in advance and stored by the classification result storage unit 407, can be used.

The above-mentioned instruction information may be used alone or may be used by appropriately combining them.

FIG. 22 shows an example in which five initial cluster centroids are obtained by designating two arbitrary document vectors, three document feature vectors, one word, one classification representative value, and a combination rule thereof. As shown in FIG.
In the present embodiment, document 1 is used as the initial cluster centroid 1.
, The average of documents 2 and 3 as initial cluster centroid 2, the average of document 4 and word 1 as initial cluster centroid 3, document 5 as initial cluster centroid 4, and classification representative value 1 as initial cluster centroid 5. Is specified.

If the specified document vector is less than the number of clusters specified by the operator, k-mea
The remaining cluster centroids can be obtained by using a general automatic initial centroid selection method used in the ns method or the like. Km-m based on the initial center of gravity obtained in this manner.
The document classification is executed by refining the cluster using the eans method or the like.

As described above, according to the sixth embodiment, the non-hierarchical clustering method is used as the document classification method, and the operator can arbitrarily specify the initial cluster center of gravity required at that time. Since the document classification is performed according to the designated initial cluster center of gravity, the document classification reflecting the intention of the operator can be performed.

[0195]

As described above, according to the first aspect of the present invention, based on the similarity between documents in a group of documents to be classified, each document is converted into a semantic relationship between the documents. By calculating an expression space conversion function for converting into an expression space that can be reflected, and performing document classification in that expression space,
There is an effect that a document classification device capable of implementing document classification that can reflect the intention of the operator can be obtained.

According to the second aspect of the present invention, the inner product between the document feature vectors is used as the similarity between the documents required for deriving the expression space conversion function, so that the semantic between the documents is obtained. This has the effect of providing a document classification device that can perform document classification reflecting the relevance.

According to the third aspect of the present invention, in addition to the inner product between the document feature vectors, the similarity between the documents required when deriving the expression space conversion function, the creator of the document, the date of creation, etc. The use of the inter-document similarity information has the effect of providing a document classification device capable of performing document classification reflecting the semantic relevance between documents.

According to the fourth aspect of the present invention, the document feature vector to be calculated and the expression space conversion function are stored, so that the expression space conversion function is calculated using the expression space conversion function. Since the document classification and the part for performing the document classification are performed by using the processed document, the document classification can be performed without calculating the expression space conversion function each time, and the document feature vector conversion unit is used. A document classification device capable of efficiently performing document classification repeatedly in a short time because an expression space conversion function generated in advance based on another document feature vector can be used as the expression space conversion function. Is obtained.

According to the fifth aspect of the present invention, when the document classification is repeatedly executed, each classification execution is
By manipulating the document feature vectors and the feature dimensions that compose them, change the range of the document to be classified or change the space in which the classification is performed, such as deleting the different words for each classification and performing document classification. There is an effect that a document classification device capable of performing the above is obtained.

Further, according to the invention of claim 6, when the expression space conversion function is calculated based on the inner product of the document feature vector, the document classification is performed using the document converted using the expression space conversion function. In the part to be performed, the inconsistency of the expression space conversion function that occurs when the document feature vector or its feature dimension is manipulated can be easily corrected, so that a document that can perform more appropriate document feature vector conversion There is an effect that a classification device can be obtained.

According to the seventh aspect of the present invention, a document in which the operator's intention can be reflected by the operator performing a simple operation on the feature dimension of the space formed by using the expression space conversion function. There is an effect that a document classification device capable of performing classification can be obtained.

Further, according to the invention of claim 8, regarding the feature dimension of the space formed by using the expression space conversion function, any document vector data other than the classification target specified by the operator is used. By performing a simple operation, it is possible to obtain a document classification device capable of performing document classification that can reflect the intention of the operator.

According to the ninth aspect of the present invention, a simple operation can be performed on a feature dimension of a space formed by using the expression space conversion function using a document feature vector specified by an operator. As a result, there is an effect that a document classification device capable of performing document classification that can reflect the intention of the operator can be obtained.

According to the tenth aspect of the present invention, the feature dimension of the space formed by using the expression space conversion function can be easily operated by using the analysis information specified by the operator. Thus, there is an effect that a document classification device capable of performing document classification that can reflect the intention of the operator can be obtained.

According to the eleventh aspect of the present invention, a feature dimension of a space formed by using the expression space conversion function can be easily determined by using a classification result specified in advance by an operator. By performing the operation, it is possible to obtain a document classification device capable of performing document classification that can reflect the intention of the operator.

According to the twelfth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and an operator can arbitrarily designate an initial cluster centroid required at that time. Since the document classification is performed according to the designated initial cluster centroid, an effect is obtained that a document classification device capable of performing the document classification reflecting the intention of the operator is obtained.

According to the thirteenth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and any document other than a classification target is used as an initial cluster centroid required at that time. Therefore, there is an effect that a document classifying device capable of classifying documents reflecting the intention of the operator can be obtained.

According to the fourteenth aspect of the present invention, a non-hierarchical clustering method can be used as a document classification method, and a document feature vector can be used as an initial cluster centroid required at that time. This has the effect of providing a document classification device that can perform document classification reflecting the intention of the operator.

According to the fifteenth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and a document to be classified is made to act on a document analysis unit as an initial cluster centroid required at that time. Since the analysis information such as the words obtained as a result can be used, there is an effect that a document classification device capable of performing document classification reflecting the intention of the operator can be obtained.

According to the sixteenth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and a classification result preliminarily classified is used as an initial cluster centroid required at that time. Therefore, there is an effect that a document classification method capable of performing document classification reflecting the intention of the operator can be obtained.

According to the seventeenth aspect, based on the similarity between documents in a group of documents to be classified, each document is converted into an expression space that can reflect the semantic relevance between the documents. By calculating an expression space conversion function for conversion and performing document classification in the expression space, it is possible to obtain a document classification method capable of realizing a document classification that can reflect the intention of the operator. .

According to the eighteenth aspect of the present invention, the inner product between the document feature vectors is used as the similarity between the documents required for deriving the expression space conversion function.
There is an effect that a document classification method capable of performing document classification reflecting semantic relevance between documents can be obtained.

According to the nineteenth aspect of the present invention, in addition to the inner product between the document feature vectors, the similarity between the documents required when deriving the expression space conversion function, the creator of the document, the date of creation, etc. By using the inter-document similarity information, there is an effect that a document classification method capable of performing document classification reflecting the semantic relevance between documents can be obtained.

According to the twentieth aspect of the present invention, the document feature vector to be calculated and the expression space conversion function are stored, so that the expression space conversion function is calculated and the expression space conversion function is actually used. Since the document classification and the part for performing the document classification are performed by using the processed document, the document classification can be performed without calculating the expression space conversion function each time, and the document feature vector conversion unit is used. A document classification method that can perform repetitive execution of document classification in a short time and efficiently because an expression space conversion function generated in advance based on another document feature vector can be used as the expression space conversion function. Is obtained.

According to the twenty-first aspect of the present invention, when the document classification is repeatedly performed, the document feature vectors and the feature dimensions constituting them are manipulated for each individual classification, so that each classification is performed. This has the effect of providing a document classification method that can change the range of the document to be classified and change the space in which the classification is performed, such as executing the document classification by deleting different words.

According to the twenty-second aspect, when the expression space conversion function is calculated based on the inner product of the document feature vectors, the document classification is performed using the document converted using the expression space conversion function. In the part to be performed, the inconsistency of the expression space conversion function that occurs when the document feature vector or its feature dimension is manipulated can be easily corrected, so that a document that can perform more appropriate document feature vector conversion This produces an effect that a classification method can be obtained.

According to the twenty-third aspect of the present invention, a document which can reflect the intention of the operator by allowing the operator to easily operate the feature dimension of the space formed by using the expression space conversion function. There is an effect that a document classification method capable of performing classification can be obtained.

According to the twenty-fourth aspect of the present invention, for the feature dimension of a space formed by using the expression space conversion function, any document vector data other than the classification target specified by the operator is used. By performing a simple operation, it is possible to obtain a document classification method capable of performing a document classification that can reflect the intention of the operator.

According to the twenty-fifth aspect of the present invention, a simple operation can be performed on a feature dimension of a space formed by using the expression space conversion function using a document feature vector specified by an operator. Accordingly, there is an effect that a document classification method capable of performing document classification that can reflect the intention of the operator can be obtained.

According to the twenty-sixth aspect of the present invention, a feature dimension of a space constituted by using the expression space conversion function can be easily operated by using analysis information designated by an operator. Thus, there is an effect that a document classification method capable of performing document classification that can reflect the intention of the operator can be obtained.

According to the twenty-seventh aspect of the present invention, a feature dimension of a space formed by using the expression space conversion function is easily determined by using a classification result pre-classified by an operator. By performing the operation, it is possible to obtain a document classification method capable of performing document classification that can reflect the intention of the operator.

According to the twenty-eighth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and an operator can arbitrarily designate an initial cluster centroid required at that time. Since the document classification is performed according to the designated initial cluster center of gravity, there is an effect that a document classification method capable of performing the document classification reflecting the intention of the operator is obtained.

According to the twenty-ninth aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and any document other than a classification target is used as an initial cluster centroid required at that time. Therefore, there is an effect that a document classification method capable of performing document classification reflecting the intention of the operator can be obtained.

According to the thirtieth aspect of the present invention, a non-hierarchical clustering method can be used as a document classification method, and a document feature vector can be used as an initial cluster centroid required at that time. There is an effect that a document classification method capable of performing document classification reflecting the intention of the operator can be obtained.

According to the thirty-first aspect of the present invention, a non-hierarchical clustering method is used as a document classification method, and a document to be classified is caused to act on a document analysis unit as an initial cluster centroid required at that time. Since analysis information such as words obtained as a result can be used, there is an effect that a document classification method capable of performing document classification reflecting the intention of the operator can be obtained.

According to the invention of claim 32, a non-hierarchical clustering method is used as a document classification method, and a classification result preliminarily classified is used as an initial cluster centroid required at that time. Therefore, there is an effect that a document classification method capable of performing document classification reflecting the intention of the operator can be obtained.

Further, according to the invention of claim 33, by recording a program for causing a computer to execute the method according to claim 17 to 32, the program becomes machine-readable, whereby the program can be read. ~ 32
There is an effect that a recording medium capable of realizing the above operation by a computer can be obtained.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram showing a hardware configuration of an entire information processing system constituting a document classification device according to a first embodiment of the present invention;

FIG. 2 is an explanatory diagram showing a hardware of a server in the information processing system constituting the document classification device according to the first embodiment;

FIG. 3 is an explanatory diagram showing a hardware of a client in the information processing system constituting the document classification device according to the first embodiment;

FIG. 4 is a block diagram functionally showing the configuration of the document classification device according to the first embodiment.

FIG. 5 is another block diagram functionally showing the configuration of the document classification device according to the first embodiment.

FIG. 6 is another block diagram functionally showing the configuration of the document classification device according to the first embodiment.

FIG. 7 is an explanatory diagram showing an example of document-word matrix data and a document feature vector of the document classification device according to the first embodiment.

FIG. 8 is a flowchart showing a procedure of a series of processes of the document classification device according to the first embodiment.

FIG. 9 is a flowchart showing another procedure of a series of processes of the document classification device according to the first embodiment.

FIG. 10 is a block diagram functionally showing a configuration of a document classification device according to a second embodiment of the present invention.

FIG. 11 is a flowchart illustrating a procedure of a series of processes of the document classification device according to the second embodiment.

FIG. 12 is a block diagram functionally showing a configuration of a document classification device according to a third embodiment of the present invention.

FIG. 13 is a flowchart showing a procedure of processing contents of a vector correction unit of the document classification device according to the third embodiment.

FIG. 14 is an explanatory diagram showing an example of a procedure for deleting a feature dimension from a document feature vector of the document classification device according to the third embodiment.

FIG. 15 is a flowchart illustrating a procedure of a series of processes of the document classification device according to the third embodiment.

FIG. 16 is a block diagram functionally showing a configuration of a document classification device according to a fourth embodiment of the present invention.

FIG. 17 is a flowchart illustrating a procedure of a series of processes of the document classification device according to the fourth embodiment.

FIG. 18 is a block diagram functionally showing a configuration of a document classification device according to a fifth embodiment of the present invention.

FIG. 19 is a flowchart showing a partial procedure of a series of processes of the document classification device according to the fifth embodiment.

FIG. 20 is a block diagram functionally showing a configuration of a document classification device according to a sixth embodiment of the present invention.

FIG. 21 is a flowchart showing a partial procedure of a series of processes of the document classification device according to the sixth embodiment.

FIG. 22 is an explanatory diagram illustrating an example of a process of obtaining an initial cluster barycenter of the document classification device according to the sixth embodiment.

[Explanation of symbols]

 101 server 102 client 103 network 201 CPU 204 I / F 206 disk device 301 CPU 306 hard disk 308 display 309 I / F 311 keyboard 312 mouse 313 scanner 401 input unit 402 analysis unit 403 vector generation unit 404 conversion function calculation unit 405 vector conversion unit 406 Classification unit 407 Classification result storage unit 421 Inner product calculation unit 431 Inter-document similarity information setting unit 1001 Vector storage unit 1002 Conversion function storage unit 1201 Vector correction unit 1601 Conversion function correction unit 1801 Conversion function correction instruction unit 2001 Initial center of gravity specification unit 2002 Initial center of gravity registration section

Continued on the front page (72) Inventor Kazutoshi Takeya 1-3-6 Nakamagome, Ota-ku, Tokyo Inside Ricoh Company (72) Inventor Akiko Nakajima 1-3-6 Nakamagome, Ota-ku, Tokyo Ricoh Company (72) Inventor Tetsuro Nagatsuka 1-3-6 Nakamagome, Ota-ku, Tokyo Ricoh Co., Ltd. (72) Inventor Makoto Yamazaki 1-3-6 Nakamagome, Ota-ku, Tokyo Ricoh Co., Ltd. (72 ) Inventor Katsuhiko Fujita 1-3-6 Nakamagome, Ota-ku, Tokyo Inside Ricoh Co., Ltd.

Claims (33)

[Claims] An input unit for inputting document data; an analysis unit for analyzing the document data input by the input unit to obtain analysis information; and an analysis unit for analyzing the document data based on the analysis information obtained by the analysis unit. A vector generation unit for generating a document feature vector; and a conversion function calculation for calculating an expression space conversion function for projecting the document feature vector generated by the vector generation unit onto a space reflecting similarity between the document feature vectors. Means, vector conversion means for converting the document feature vector generated by the vector generation means using the expression space conversion function calculated by the conversion function calculation means, between the document feature vectors converted by the vector conversion means Classifying means for classifying documents based on the similarity of the documents; And a classification result storing means for storing a result of the class. 2. An inner product calculating means for calculating an inner product between the document feature vectors generated by the vector generating means, wherein the conversion function calculating means expresses the inner product calculated by the inner product calculating means. The document classification device according to claim 1, wherein a space conversion function is calculated. 3. An inter-document similarity information setting unit for setting inter-document similarity information of document data such as a creator and a creation date of a document input by the input unit, wherein the conversion function calculating unit includes 3. The document classification apparatus according to claim 2, wherein the expression space conversion function is calculated using the inner product calculated by the output unit and the inter-document similarity information set by the inter-document similarity information setting unit. 4. A vector storage unit for storing a document feature vector generated by the vector generation unit, and a conversion function storage unit for storing an expression space conversion function calculated by the conversion function calculation unit. The document classification device according to any one of claims 1 to 3, wherein: 5. The method according to claim 1, further comprising, before changing the document feature vector by the vector conversion means, using a rule constituted by the characteristics of the words extracted by the analysis means. The document classification apparatus according to any one of claims 1 to 4, further comprising a vector correction unit configured to correct the document feature vector by manipulating a feature dimension of the document. 6. When a feature dimension is changed by correcting a document feature vector in the vector correction means, the vector feature means can appropriately convert the document feature vector by the changed feature dimension. 6. The document classification device according to claim 5, further comprising a conversion function correction unit that corrects the expression space conversion function calculated by the conversion function calculation unit. 7. A conversion function correction instructing unit for instructing an operation on a feature dimension of the expression space conversion function, and the instruction content on the operation of the feature dimension instructed by the transformation function correction instructing unit. The document classification device according to claim 1, further comprising: a conversion function correction unit configured to correct a representation space conversion function. 8. The method according to claim 1, wherein the instruction content related to the operation of the feature dimension specified by the conversion function correction instructing means is to operate the feature dimension of the expression space conversion function using arbitrary document vector data. The document classification apparatus according to claim 7, which performs the following. 9. The method according to claim 1, wherein the instruction content related to the operation of the feature dimension specified by the conversion function correction instructing means is to operate the feature dimension of the expression space conversion function using a document feature vector. Item 8. The document classification device according to Item 7. 10. The instruction content regarding the operation of the characteristic dimension instructed by the transformation function modification instructing means operates the characteristic dimension of the expression space transformation function using the analysis information obtained by the analyzing means. The document classification device according to claim 7, wherein: 11. An instruction content relating to an operation of a feature dimension instructed by said transformation function modification instructing means, wherein said instruction content is operated by using a classification result stored by said classification result storage means. The document classification device according to claim 7, wherein 12. An initial center of gravity specifying means for specifying an initial cluster center of gravity, and an initial center of gravity registering means for registering an initial cluster center of gravity specified by the initial center of gravity specifying means. The document classification apparatus according to any one of claims 1 to 11, wherein the documents are classified according to the initial cluster centroid registered by (1). 13. The document classification apparatus according to claim 12, wherein any document vector data is specified as an initial cluster centroid specified by said initial centroid specifying means. 14. The document classification device according to claim 12, wherein a document feature vector is specified as an initial cluster centroid specified by said initial centroid specifying means. 15. The document classification apparatus according to claim 12, wherein the analysis information obtained by the analysis unit is specified as the initial cluster center of gravity specified by the initial center of gravity specification unit. 16. The document classification apparatus according to claim 12, wherein the classification result stored by said classification result storage means is specified as an initial cluster centroid specified by said initial centroid specification means. 17. A first step of inputting document data; a second step of analyzing the document data input in the first step to obtain analysis information; and A third step of generating a document feature vector for the document data; and calculating a representation space conversion function for projecting the document feature vector generated in the third step into a space reflecting similarity between the document feature vectors. A fourth step of converting the document feature vector generated in the third step using the expression space conversion function calculated in the fourth step; and a document feature converted in the fifth step. A sixth step of classifying the document based on the similarity between the vectors, and a seventh step of storing a result of the document classification classified by the sixth step classification means. Document classification method and butterflies. 18. An eighth step of calculating an inner product between the document feature vectors generated in the third step, and the fourth step uses the inner product calculated in the eighth step to represent a representation space conversion function. The document classification method according to claim 17, wherein is calculated. 19. A ninth step of setting inter-document similarity information of document data such as a creator and a creation date of the document input in the first step, wherein the fourth step is calculated by the eighth step. 19. The document classification method according to claim 18, wherein the expression space conversion function is calculated using the obtained inner product and the similarity information between documents set in the ninth step. 20. The method according to claim 20, further comprising: a tenth step of storing the document feature vector generated in the third step, and an eleventh step of storing the expression space conversion function calculated in the fourth step. The document classification method according to any one of claims 17 to 19, wherein: 21. Further, before changing the document feature vector in the fifth step, the document feature vector and / or the document feature are changed using a rule constituted by the characteristics of the words extracted in the second step. 21. The document classification method according to claim 17, further comprising a twelfth step of modifying the document feature vector by manipulating a feature dimension constituting the vector. 22. When the feature dimension is changed by correcting the document feature vector in the twelfth step, the document feature vector can be appropriately converted in the fifth step by the changed feature dimension. 22. The document classification method according to claim 21, further comprising a thirteenth step of correcting the expression space conversion function calculated in the fourth step. 23. A fourteenth step of giving an instruction on operation of a feature dimension of the expression space conversion function, and the expression space conversion function based on the instruction content on the operation of feature dimension instructed in the fourteenth step. The document classification method according to any one of claims 17 to 21, further comprising: a fifteenth step of correcting. 24. The method according to claim 15, wherein the instruction content relating to the operation of the characteristic dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using arbitrary document vector data. Item 23. The document classification method according to Item 23. 25. The method according to claim 23, wherein the instruction content related to the operation of the feature dimension specified in the fifteenth step is to operate the feature dimension of the expression space conversion function using a document feature vector. Document classification method described in. 26. The instruction content relating to the operation of the feature dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using the analysis information obtained in the second step. 24. The method of claim 23, wherein
Document classification method described in. 27. The instruction content relating to the operation of the feature dimension specified in the fifteenth step is to operate the characteristic dimension of the expression space conversion function using the classification result stored in the seventh step. 3. The method according to claim 2, wherein
3. The document classification method according to 3. 28. A sixteenth step of specifying an initial cluster centroid, and a seventeenth step of registering the initial cluster centroid specified in the sixteenth step, wherein the sixth step is registered in the seventeenth step. 28. The document classification method according to claim 17, wherein the documents are classified according to the initial cluster centroid. 29. The document classification method according to claim 28, wherein any document vector data is specified as the initial cluster centroid specified in the sixteenth step. 30. The document classification method according to claim 28, wherein a document feature vector is specified as an initial cluster centroid specified in the sixteenth step. 31. The document classification method according to claim 28, wherein the analysis information obtained in the second step is specified as an initial cluster barycenter specified in the sixteenth step. 32. The method according to claim 28, wherein the classification result stored in the seventh step is specified as the initial cluster centroid specified in the sixteenth step. 33. A computer-readable recording medium on which a program for causing a computer to execute the method according to claim 17 is recorded.