JP3603392B2 - Document classification support method and apparatus - Google Patents

Description

[0001]
[Industrial applications]
The present invention relates to a document classification method and apparatus for classifying digitized documents including text information into categories, and more particularly to a document classification support method and apparatus for efficiently performing a task of checking a classification result by a computer by a user. About.
[0002]
[Prior art]
With the socialization of information and the development of information infrastructure, a large amount of information has become flooded, and it has become essential to extract necessary information efficiently. One of the solutions is to classify documents into appropriate categories in advance, and the development of a computer-based automatic classification technique has been required.
[0003]
Examples of automatic classification technology of digitized text documents include Proceedings of Second Annual Conference on Innovative (1990), IPSJ Research Reports NL-98-11, and Info-Tech '94 Lecture Papers pp. 138-pp. 146. These determine the category based on the appearance tendency of the keyword in the text document.
[0004]
[Problems to be solved by the invention]
The above-mentioned technique classifies texts automatically by a computer, and does not refer to a method of determining a classification result in cooperation with a user in the literature. In addition, the classification accuracy by the above technique has not reached the same level as humans.
[0005]
However, in a situation where the same level of classification accuracy as that of a human is required, the user needs to check the classification result of the computer. Therefore, if the computer and the user share the roles and perform the classification work cooperatively, it leads to cost reduction. In other words, how to determine the category to be classified efficiently and with a small work load based on the classification result of the computer is an issue.
[0006]
In particular, when the number of documents to be classified is large, it is an issue how to reduce the work time required for each document and how to reduce the work load. In addition, when the number of categories is relatively large, when the categories are complicated and their identification is extremely difficult, when the classification result output by the computer is determined to be correct, or when the classification result is incorrect In addition, it is very difficult to find a true category from scratch. Therefore, how to perform these operations efficiently is an issue.
[0007]
Therefore, it is an object of the present invention to efficiently perform the operation of determining whether the classification result is correct and the operation of finding a true category when the classification result is incorrect.
[0008]
Also, when a large number of documents are sequentially classified, the order often does not depend on the contents of the documents. In this case, the content described changes greatly every time the document changes, so that the user who checks needs to switch his head to the content every time the content changes. For this reason, there is a problem that the efficiency of the check is low and the work load increases.
[0009]
Therefore, another object of the present invention is to reduce the work load due to frequent and significant changes in the contents of documents to be classified, and to improve the efficiency of the classification work.
[0010]
[Means for Solving the Problems]
In the present invention, log data relating to why a document to be classified is classified into the category is presented to the user via the category estimation result output means, and the presented log data is presented to the user via the user input means. The above problem is solved by causing the user to make a correction, re-estimating the category based on the corrected log data, and presenting the re-estimated category to the user via the estimation result output means.
[0011]
Further, in the present invention, similar document recognition means for estimating a category to be classified for a plurality of documents by a text analysis means and a category estimation means, and recognizing a document set in which the estimated categories are similar or identical to each other. And sequentially presenting the category estimated by the category estimating means to the user via the estimation result output means for the similar document, and determining the category to be classified to the user via the user input means for the presented document. Solves the above problem.
[0012]
[Action]
Documents having similar contents are put together based on the estimated category, and the documents are sequentially presented to the user to prompt the user to check. Therefore, the number of switching of the user's head due to a large change in the content is reduced, and the work load is reduced. In addition, since documents having similar contents continue, it is easy to utilize tips, know-how, lessons learned, data, and the like when checking the previous document in checking the next document, and the checking work time is reduced.
[0013]
【Example】
Embodiments of the present invention will be described below in detail with reference to the drawings.
In this embodiment, newspaper articles are classified into a certain category and stored in a document database. Newspaper article data stored in the database for each category can be searched by using a known search system.
[0014]
FIG. 1 is a diagram illustrating an outline of the present embodiment.
First, a document to be classified is input by a document input 1. The document data may be obtained from the outside via a network, may be obtained via a medium such as a floppy disk, or may be input by handwriting such as a voice recognition device, an image recognition device (including character recognition), and a pen. It may be obtained via a device or the like. Further, the document data may be collectively acquired periodically, or the distributed document data may be acquired irregularly and sequentially. The acquired document data is temporarily stored in the document file 10.
[0015]
Next, the document data is analyzed in accordance with the designation of the document data to be classified and the instruction to execute the category estimation from the user. It is determined whether or not there is a document that has not been estimated (1a). If not, the process proceeds to step 3a.
[0016]
If there is, first, in the text analysis 2, a keyword that characterizes the content is automatically extracted from the text by natural language processing. That is, the text is divided into words with reference to the word dictionary 11 storing the words and their part of speech / utilization information, and the words whose nouns are nouns are set as keywords, and stored in the keyword table 12 together with the appearance frequency of each keyword.
[0017]
Next, in the category estimation 3, referring to the classification knowledge 13 defining and storing the keywords characterizing each category in advance and the category definition table 14 defining the category system, the keywords in the keyword table 12 extracted from the text A search is performed to determine whether the category is included. If the category is included, a score is assigned to the category. Then, it is estimated that the category with the highest score is the category to which the text should be classified. The estimation result is stored in the estimation category table 15. Further, data such as keyword information used in estimating the category and score information of the category are stored in the log data 17.
[0018]
Next, the estimation result is output so that the user can check the estimation result. At this time, the user is instructed whether or not to display the estimation result for each document whose content is similar (3a). If the estimation result is not displayed for each document whose content is similar, the estimation result is displayed in the order of the document ID. indicate.
[0019]
In the case of displaying each document having similar contents, similar document recognition 4 recognizes similar documents from the category estimation result of each document stored in the estimated category table 15 and divides the result into a similar document. Stored in table 16.
[0020]
Next, if there is a document whose category has not been determined by the user (4a), the category estimation results are sequentially presented to the user (5), and the results are checked and the category to be classified is determined (5a). At this time, the analysis data stored in the log data 17 is also presented to the user.
[0021]
The user checks whether the presented category is correct. If it is correct, the category is determined and the document is registered in the document database 18.
If not, you have to find the right category. At that time, if the user instructs to re-estimate the category to be classified, first, the user corrects the presented log data (6), and re-estimates the category based on the corrected data. (7) The new estimation result is presented to the user together with the new analysis data. When the user determines that the category is correct, the category is determined (8) and registered in the document database (9). If the correct category cannot be found even after re-estimating the category several times, the user manually determines the category.
[0022]
When the category is determined, the process proceeds to the next document check (9a), and the category estimation result and log data of the document are output.
[0023]
FIG. 2 is a diagram illustrating an outline of hardware of the present embodiment. It comprises a keyboard 20, a mouse 25 for inputting operation instructions and data from a user, an output monitor 30 for outputting results, a processing device 40 for executing various processes, and a storage device 50 for storing files and programs. In addition, it is connected to a computer network 90 for acquiring document data, and can acquire documents via the network.
[0024]
The storage device 50 includes a working area 61 for temporarily storing data, a document file storage area 62 for temporarily storing acquired document data, a word dictionary storage area 63, a keyword table storage area 64, a classification knowledge storage area 65, and a category definition. It includes a table storage area 66, an estimated category table storage area 67, a similar document table storage area 68, a log data storage area 69, and a document database storage area 70. What is stored in the storage area other than the working area 61 is a file in a data format.
[0025]
Further, the storage device 50 includes a text analysis processing unit storage area 71, a category estimation processing unit storage area 72, a similar document recognition processing unit storage area 73, a category estimation result display unit storage area 74, a log data correction unit storage area 75, a category It also includes a re-estimation processing unit storage area 76, a category determination processing unit storage area 77, and a document database registration processing unit storage area 78. Stored in these storage areas are executable load module files.
[0026]
Note that the numbers in parentheses shown in FIG. 2 indicate the correspondence with the respective parts shown in FIG.
[0027]
FIG. 3 is a diagram illustrating an example of text information included in a document.
The document data handled in this embodiment is a newspaper article, but other types of document data such as electronic news, e-mail, scientific and technical papers, patent specifications, claims / questions / opinions, minutes of meetings, etc. It may be something. In this embodiment, it is assumed that the document data includes text information, and that such information is stored in a file in a text code format. However, a still image, a moving image, audio information, and the like may be linked.
[0028]
FIG. 4 is a diagram illustrating an example of the word dictionary 11 referred to in the text analysis 2. The word dictionary has, in addition to the heading 201, word attribute information such as a part of speech 202, a utilization type 203, and a utilization line 204.
[0029]
FIG. 5 is a diagram illustrating an example of a word segmentation result in text analysis 2. In the text analysis 2, first, the text shown in FIG. 3 is divided into each sentence with reference to the word dictionary 11 in FIG. 4, and as shown in FIG. Is extracted. Since a specific method of realizing word division is already known, as shown in, for example, the 44th National Convention of Information Processing Society of Japan, (3) 3-181, detailed description is omitted here.
[0030]
FIG. 6 is a diagram illustrating an example of the keyword table 12 that stores keywords extracted from text. In the text analysis 2, after the text is divided into words, words whose nouns are nouns are extracted and used as keywords, and the appearance frequency of each keyword in the text is calculated and used as the weight of the keywords. Of course, a part of speech other than a noun may be used as a keyword, and in addition to using the appearance frequency as a weight, weighting may be performed in consideration of the appearance position of the keyword, the relationship with words before and after the keyword, and the like. The keyword table 12 includes a document ID 221 for identifying a document, a keyword heading 222, and a weight 223.
[0031]
FIG. 7 is a diagram illustrating an example of the category definition table 14 in which a category system is defined. In this embodiment, as a category for classifying newspaper articles, a category having two hierarchies of a large category 231 and a small category 232 is defined. One or more small categories 232 belong to each of the large categories 231 and have a tree structure. There can be any number of categories.
[0032]
FIG. 8 is a diagram illustrating an example of the classification knowledge 13. In this embodiment, a method of estimating a category to be classified based on the presence or absence of a keyword is used. Therefore, the classification knowledge 13 is a set of keywords characterizing the category. That is, the classification knowledge 13 includes a large category 241, a small category 242, a keyword 243 characterizing the category, and a weight 244 depending on the importance of the keyword. The weight 244 has a larger value as the keyword is an important keyword characterizing the category. The classification knowledge 13 is stored in the storage device 50 in advance. The classification knowledge may be created manually, or may be created by preparing a text whose category is already determined for each category and automatically extracting a keyword for each category.
[0033]
FIG. 9 is a diagram illustrating a processing procedure of the third category estimation. First, a table storing scores of each category is initialized to 0 (step 501).
[0034]
Next, the following processing is performed for all keywords of the document stored in the keyword table 12 (step 502). It is determined whether or not there is a category in the classification knowledge 13 including the keyword (step 503). For the existing category, the weight Wi of the keyword of the document (corresponding to 223 in FIG. 6) and the category The product of the weights Wj (corresponding to 244 in FIG. 8) of the keyword is calculated and added as the score of the category (step 504).
[0035]
When the above processing is performed for all keywords, the score of each category is determined, and the deviation value of the score of each category is calculated from these scores (step 505). Further, the categories are sorted in descending order of the deviation value (step 506). Then, the document ID, the category, and the value of the deviation value are paired and stored in the estimated category table 15 in the descending order of the deviation value (step 507). In this embodiment, the top three categories are stored. Of course, the top n categories may be stored, or a lower limit may be set for the value of the deviation value, and the categories above the lower limit may be stored. Finally, the log data 17 stores the document ID, the keyword extracted from the document, and when each keyword is included in the keyword of each category, the values of the weight Wi, the weight Wj, and the product thereof in step 504 are stored ( Step 508).
[0036]
In the present embodiment, a category system of two layers (large category and small category) is used. However, in category estimation 3, the estimation is performed on the small category, and the estimation of the large category is uniquely determined once the small category is determined. ,not going. As another estimation method, first, a category may be estimated for a large category, and a small category may be estimated in a form limited to a large category ranked higher. In this case, the classification knowledge 13 defining the keyword characterizing the large category and its weight is required. It may be newly created manually, or may be easily created by summarizing the classification knowledge about the small categories for each large category.
[0037]
FIG. 10 is a diagram illustrating an example of the estimated category table 15. The estimated category table 15 includes a document ID 251, an estimated category rank 252, an estimated large category candidate 253, an estimated small category candidate 254, and a deviation value 255 of the category.
[0038]
FIG. 11 is a diagram showing a processing procedure of similar document recognition 4. First, the similar document table 16 is initialized (step 521). Next, the following processing is performed for all categories (step 522). With reference to the estimated category table 15, the document ID of the document estimated to be classified into the first place in the category is extracted from the documents whose category is estimated (step 523).
[0039]
Next, the extracted document IDs are grouped for each category estimated to be classified into the second place, and stored in the similar document table 16 in association with the category (step 524).
[0040]
FIG. 12 is a diagram illustrating an example of the similar document table 16. As illustrated in FIG. 11, in the present embodiment, the category estimated as the first place and the category estimated as the second place are grouped for each same document and stored in the similar document table 16. That is, the similar document table 16 includes the category 261 estimated as the first place, the category 262 estimated as the second place, and the document ID 263 having these as the estimation result.
[0041]
FIG. 13 is a diagram illustrating an example of the category estimation result display. Here, the document designation button 401 is for designating the range of the document to be processed, and designates the directory where the document exists. The classification button 402 executes the text analysis 2 and the category estimation 3 on the specified document, and obtains the estimation result and log data. The re-classification button 403 executes re-estimation of the category based on the data corrected by the user, and outputs a re-estimation result. The narrowing down classification button 404 allows the user to specify a category in an upper hierarchy, executes category estimation limited to lower categories belonging to the category, and outputs an estimation result, as described later. The category list button 405 displays the contents of the category definition table 14. The classification knowledge reference button 406 displays the keywords and their weights stored in the classification knowledge 13 for each category. An end button 407 ends the system.
[0042]
An area 411 displays the contents of the text, and also displays the ID of the document text. An area 412 displays the keywords extracted from the text and the weights (appearance frequencies) in pairs in descending order of the weight.
[0043]
413 displays, for each category, which keyword among the 412 keywords is included, and how large the score is. The category is specified by specifying one of the 414 categories that are the classification results. In 413 of FIG. 13, for example, the keyword “yen” is included in the keyword of the small category “international economy”, the weight Wi of the keyword extracted from the text is 4, and the “international economy” of the classification knowledge 13 is The weight “Wj” of the keyword “circle” in the small category “8” is 8, and as a result, 4 × 8 = 32 points are given.
[0044]
An area 414 displays the estimated large category, small category, and their deviation values. An area 415 displays the category determined by the user.
A button 416 displays the category estimation result, log data, and confirmed category of the document checked immediately before the currently checked document. Since the data on these checked documents is stored in the estimated category table and the log data, it can be easily realized by displaying those data.
[0045]
A button 417 is used to determine the category of the currently checked document, and to instruct to check the next document. At this time, the category described in 415 is determined as a category to be classified, and the document is registered in the document database 18 together with the category information.
[0046]
FIG. 14 is a diagram showing another example of the category estimation result display. Reference numeral 421 denotes a list of classification knowledge, which is displayed with reference to the classification knowledge 13 when the classification knowledge reference button 406 is pressed. Reference numeral 422 indicates the category definition table 14 when the category list button 405 is pressed. A text 423 describes the range of the category, and is displayed when any one of the categories is selected in the category list 422.
[0047]
FIG. 15 is a diagram illustrating an example of the screen after the log data has been corrected by the user. Regarding 411 and 412, the user can modify the data displayed via the keyboard 20 and the mouse 25. In FIG. 15, a correction has been made for 412. As for keywords, it is possible to delete displayed keywords, add new keywords, and correct displayed weights. In contrast to FIG. 13 which is the screen before correction, in FIG. 15, the weights of keywords such as “yen”, “foreign exchange market”, and “rapid rise” are corrected. Non-keywords have been deleted.
[0048]
FIG. 16 is a diagram illustrating an example of the category re-estimation result. As a result of correcting the keywords and their weights, the category "exchange" that did not appear as the previous estimation result newly appeared as the first place as the classification result 414. Categories that have newly appeared in this way are distinguished from other categories by adding an asterisk. Of course, the method of distinction may be other than the addition of the star.
[0049]
FIG. 17 is a diagram illustrating a processing procedure of the category re-estimation 7. First, a table storing scores of each category is initialized to 0 (step 541).
[0050]
Next, the document ID, the corrected text, the corrected keyword, and its weight are read from the output screen and stored in the working area 16 (step 542). Next, it is determined whether the text information has been corrected (step 543). If the text information is corrected, the keywords extracted therefrom and their weights change greatly, and it is necessary to start over from the text analysis 2. On the other hand, when the text information has not been corrected, the keyword information read from the display screen can be used, so that the processing may be performed from the category estimation 3. Whether or not the text information has been corrected can be determined by providing a text correction flag and turning it on and off.
[0051]
If the text information is corrected in step 543, text analysis 2 is executed to extract keywords and weights from the corrected text and store the result in the working area 61 (step 544).
[0052]
Next, the following processing is performed for all keywords stored in the working area 61 (step 545). It is determined whether or not a category in the classification knowledge including the keyword exists (step 546). For the existing category, the weight Wi of the keyword of the document (corresponding to 223 in FIG. 6) and the The product of the weights Wj (equivalent to 244 in FIG. 8) of the keyword is calculated and added as the score of the category (step 547).
[0053]
Since the score of each category is determined at the point of time when all the keywords are performed, a deviation value of the score of each category is calculated from these scores (step 548). Further, the categories are sorted in descending order of the deviation value (step 549). Then, the document ID, the category, and the value of the deviation value are paired and stored in the estimated category table 15 in the descending order of the deviation value (step 550).
[0054]
FIG. 18 is a diagram illustrating an example of the log data 17. The log data 17 stores and holds the document ID, keywords extracted from the text and their weights, the breakdown of the scores for each category, and data related to the determined categories until the system is terminated. Therefore, when checking the category estimation result of a certain document, it is possible to refer to the data of the document which has been checked so far.
[0055]
FIG. 19 is a diagram illustrating an example of the category determination 8. The user determines the category with reference to the classification result 414. In the present embodiment, in the classification result 414, by double-clicking the category to be determined with the mouse, the selected category is displayed as the determined category 415.
[0056]
As described above, according to the present embodiment, when a document is to be classified, a computer-assisted category candidate is estimated, the result is displayed, and the user checks the result, thereby realizing a man-machine sharing type document classification support system. it can. In addition, when displaying the classification results, the results are presented sequentially in a grouped manner for each estimated category, so that the user can efficiently check. In addition, even if the presented result is incorrect, the data can be corrected and re-classified to improve the accuracy of classifying the data into a correct category. The rate of performing large tasks can be reduced as much as possible.
[0057]
Next, a modified example of the present embodiment will be described.
In the similar document recognition 4, in the present embodiment, the recognition is performed based on the top two estimated categories. However, instead of the estimated category, the recognition may be performed using a keyword having a high weight extracted from the text.
[0058]
FIG. 20 is a diagram showing the processing method. First, the similar document table 16 is initialized (step 561). Next, the following process is executed while there is a document that has not been recognized as a similar document (step 562). For a certain document that has not been certified, n (m> = n) or more keywords out of the m weight keywords extracted from the document are included in the m weight keywords. The extracted document is extracted and stored in a similar document table together with a set identifier for identifying a similar document set (step 563). In FIG. 11, the name of the category is used as the equivalent of the set identifier, but here, the set identifier is defined as a substitute for the name. This may be in any form as long as a similar document set can be identified.
[0059]
After step 563, the documents stored in the similar document table 16 are excluded from the processing target of step 562 (step 564).
By the above processing, when presenting the result of the category estimation to the user, it is possible to present each similar document based on how much a keyword having a high weight is shared.
[0060]
Next, an extended example of the present embodiment will be described.
When the category is composed of a plurality of layers as in the present embodiment, the upper category is presented to the user to be specified, and the category is estimated only for the lower category belonging to the specified higher category, whereby the classification accuracy is improved. Improvement can be expected. This is particularly effective when the number of lower categories is enormous.
[0061]
FIG. 21 is a diagram showing an example of a screen for specifying a large category. The designation of the large category is performed by displaying the designation screen 424 when the narrow-down classification button 404 is pressed. The designation of the large category may be plural. The display order of the large categories on the designation screen 424 is basically the order defined in the category definition table 14, but in the category estimation 3, the large category is first estimated, and the result is used by using the result. When the method of estimating the small category is adopted, the estimation result of the large category of the document is stored and held in the log data 17 so that the document can be displayed based on the order of the estimation result of the large category. It is.
[0062]
After the large category is designated on the designation screen 424, the category re-estimation 7 limited to the designated large category is executed by pressing the re-classification button 403. When storing the estimation result in the estimation category table 15 in step 550 of the processing procedure of the category re-estimation 7 shown in FIG. 17, when the large category of the estimated category is included in the large category specified by the user. Only by storing, it is possible to narrow down by upper category. In the result display of FIG. 13, if the user narrows down the large category to “economy”, the category “politics: parliament” in the second place in the classification result 414 is removed.
[0063]
As described above, when the upper category is relatively small and the user can easily determine the category, the correct category can be obtained by narrowing down the upper category and estimating the category.
[0064]
【The invention's effect】
When the user checks the automatic classification result of the document, the result classified by the computer is sequentially presented to the user for each similar document, and the check is prompted, so that the tips, know-how, lessons, and It becomes easy to utilize data and the like for the next document check, and the check work time is reduced.
[0065]
Even if the result of the automatic classification is incorrect, the user can correct the log data output together with the result of the automatic classification and re-estimate the result, so that the correct classification result can be derived. In the case where is incorrect, it is possible to reduce the burdensome work of the user re-sorting from the beginning.
[Brief description of the drawings]
FIG. 1 is a diagram showing an outline of the present embodiment.
FIG. 2 is a diagram illustrating an outline of hardware of the embodiment.
FIG. 3 is a diagram illustrating an example of a text included in a document.
FIG. 4 is a diagram illustrating an example of a word dictionary.
FIG. 5 is a diagram illustrating an example of a word segmentation result in text analysis.
FIG. 6 is a diagram illustrating an example of a keyword table.
FIG. 7 is a diagram illustrating an example of a category definition table.
FIG. 8 is a diagram illustrating an example of classification knowledge.
FIG. 9 is a diagram showing a processing procedure of category estimation.
FIG. 10 is a diagram illustrating an example of an estimated category table.
FIG. 11 is a diagram showing a processing procedure for similar document recognition.
FIG. 12 illustrates an example of a similar document table.
FIG. 13 is a diagram showing an example of a category estimation result display.
FIG. 14 is a diagram showing another example of a category estimation result display.
FIG. 15 is a diagram showing an example of a screen after correction by a user.
FIG. 16 is a diagram illustrating an example of a category re-estimation result.
FIG. 17 is a diagram showing a procedure for re-estimating a category.
FIG. 18 is a diagram illustrating an example of log data.
FIG. 19 is a diagram showing an example of category determination.
FIG. 20 is a diagram illustrating another processing procedure of similar document recognition.
FIG. 21 is a diagram illustrating an example of narrowing down upper categories.
[Explanation of symbols]
1: Document input, 2: Text analysis, 3: Category estimation, 4: Similar document recognition,
5: Category estimation result display, 6: Log data correction, 7: Category re-estimation,
8: Category confirmation, 9: Document database registration, 10: Document file,
11: word dictionary, 12: keyword table, 13: classification knowledge,
14: Category definition table, 15: Estimated category table,
16: similar document table, 17: log data, 18: document database

Claims (4)

An input device, an output device, and a processing device having a storage device, a document classification support device for classifying a plurality of documents into categories,
The storage device stores a plurality of documents, a dictionary defining keywords, a category, a keyword characterizing each category, and classification knowledge data defining a first weight indicating the importance of the keyword for each category. ,
The analysis unit of the processing device refers to the dictionary of the storage device and extracts a keyword included in each document,
The estimating means of the processing device calculates a second weight of each keyword based on the frequency of appearance of each keyword included in each document, refers to the classification knowledge data, and defines the second weight by the classification knowledge data. A search is made for a category including the keyword extracted by the analysis means from among the categories, and each of the keywords in the category is determined based on the first weight and the second weight of the keyword in the searched category. Calculate the score of, calculate the deviation value of the category by adding the score of the keyword for each category, sort the category for each document according to the calculated deviation value of the category,
The certifying unit of the processing device certifies, as a similar document, a document having a common category having a high ranking by sorting among the plurality of documents,
The document classification support device, wherein the output device outputs an estimation result of the estimation unit for each of the similar documents recognized by the recognition unit. It said determining means of the processing device, wherein among the estimation result of the category estimation unit, document classification assisting apparatus according to claim 1, characterized in that to determine the user-specified category as the category of the document. In a document classification support device for classifying a plurality of documents into categories,
A storage device for storing a plurality of documents, a dictionary defining keywords, a category, a keyword characterizing each category, and classification knowledge data defining a first weight indicating importance of the keyword for each category;
Analysis means for extracting a keyword included in each document by referring to the dictionary of the storage device;
The second weight of each keyword is calculated based on the appearance frequency of each keyword included in each document, and the analysis unit refers to the classification knowledge data and selects the category defined by the classification knowledge data. Searching for a category including the extracted keyword; calculating a score of each keyword in the category based on the first weight and the second weight of each keyword in the searched category; Estimating means for calculating the deviation value of the category by adding the score of the keyword for each document, and sorting the category for each document according to the calculated deviation value of the category;
A certifying unit for certifying, as a similar document, a document having a common category having a high ranking by sorting among the plurality of documents;
A document classification support device, comprising: output means for outputting an estimation result of the estimation means for each similar document recognized by the authentication means. 4. The document classification support apparatus according to claim 3 , further comprising a determination unit that determines a category specified by a user among the categories of the estimation result of the estimation unit as a category of the document.

Images