JPH11203310A - Search expression creation method and apparatus - Google Patents

Abstract

(57) [Summary] [Problem] Without requiring a large amount of learning documents,
Provided is a search formula creation device capable of creating a search formula in consideration of a group of documents to which designation information is not added. A document set dividing unit calculates a mutual information amount when dividing a document set output from a keyword extracting unit by regarding an indefinite document group as a virtually unnecessary document group, and calculates a mutual information amount. Is determined as the search keyword. The document set division unit 14 determines when the set of documents is classified into a required document group and an unnecessary document group by the determined search keyword, and when the number of required documents in the document set at that time and in the required document group is a predetermined value for the indefinite number of documents. When the ratio is exceeded, the division of the document set and the determination of the search keyword are stopped. The search expression creating unit 33 creates a search expression by combining the determined search keywords with the logical operators “and” and “or”.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a document database for retrieving specific information from a large amount of stored electronic documents, and a pre-stored example of an electronic document for supporting document creation and idea development. Related to the document search technology applied to various support systems to be used, in particular, using a keyword extracted from an electronic document, a search formula for a searcher to efficiently search for a document of interest is tried. It relates to a method of making mistakes.

[0002]

2. Description of the Related Art A keyword is extracted from a document database in which electronic documents to be searched are stored, and a required search formula is created by trial and error in cooperation with a searcher by a combination of a logical product and a logical sum of the keywords. There is known a search formula creation device.

[0005] FIG. 5 is a functional block diagram of a conventional search formula creating apparatus of this kind. The search formula creation device 50 includes functional blocks of a keyword extraction unit 51, a document set division unit 52, and a search formula creation unit 53, which are formed by a computer device reading and executing a predetermined program. Note that it is assumed that each document is provided with necessary / unnecessary designation information indicating whether the searcher is a necessary document of interest or an unnecessary document of no interest.

[0004] A keyword extracting unit 51 performs a process of extracting a plurality of keywords for each document from a plurality of documents by a known morphological analysis process. In addition, determination information indicating the presence or absence of a keyword in each document and designation information indicating whether the document is a necessary document or an unnecessary document are output as a document set together with a document identifier such as a document name and a document number. Code 31
B illustrates the contents of the document set output from the keyword extraction unit 51.

[0005] The document set division unit 52 divides the document set in stages based on the discrimination information, and determines a plurality of search keywords that are the basis for creating a search formula used for document search. In this case, by dividing the document set by the discrimination information of one (small) keyword as much as possible, it is possible to extract the intention of the searcher who distinguishes the necessary documents from the unnecessary documents. The plurality of search keywords determined by the document set division unit 52 are combined by the logical operators “and”, “or”, and “not” in the search expression creation unit 53, and are output to the subsequent processing as a search expression.

[0006] The document set dividing process in the document set dividing section 52 is performed, for example, by a known MDL (Minimum Description).
Length: minimum description length) based on the principle. This M
The DL principle is a heuristic method that states that "manual (searcher) intentions can be expressed more accurately by distinguishing between more and more unnecessary documents with as few keyword combinations (search formulas) as possible". There is
A strict realization of the MDL principle requires a large amount of processing. Therefore, in order to reduce this processing amount, MDL
Generally, the principle is realized approximately.

A method for approximately realizing the MDL principle is, for example, a known decision tree (a logical expression represented by a tree structure).
This is performed based on “ID3” which is a learning algorithm. For details on “ID3”, see “Knowledge Acquisition and Learning Series 1: Introduction to Knowledge Acquisition” (Michalski, R.A.).
S. Other editions, Kyoritsu Publishing) can be referred to. Hereinafter, an outline of the document set division processing by the decision tree learning algorithm “ID3” will be described with reference to FIG.

First, the document set sent from the keyword extracting unit 51 is set as an initial document set Set ₀ (step S2).
01). Next, the “undivided” flag of the initial document set Set ₀ is turned on (step S202). This is Set
_i (step S203). Next, this document set S
Mutual information I (t _j ) representing the relative relationship between the information amount of the individual document and the information amount of the entire document for each keyword t _j (1 ≦ j ≦ N) included in the required document and unnecessary document in et _i
Is calculated (step S204). Mutual information I (t _j )
Is, specifically, the information amount H about the undivided document set.
Is subtracted from the information amount H (t _j ) about the document set including the keyword t _j and the document set not including the keyword t _{j, and} is represented by the following equation (1).

Equation (1) I (t _j ) = H−H (t _j )

The information amounts H and H (t _j ) are expressed by the following equations (2) and (3), respectively.

However, the parameters in the equations (2) and (3) are as follows. p _i: Set _i need the number of documents in, n _i: Set _i unnecessary number of documents _{in, s i: p i + n} i, p i (t j): necessary number of documents that contain the keyword t _j in Set _{i , n i (t j):} Set i unnecessary number of documents that contain the keyword t _{j in, s i (t j):} p i (t j) + n i (t j), p i not (t j): Set _i need the number of documents that do not contain the keyword t _{j in, n i not (t j)} : Set i unnecessary number of documents that do not contain the keyword t _{j in, s i not (t j)} : p i not (t j) + N _i not (t _j ), h (a, b, c):-{a / c · log ₂ (a / c) + b / c · log ₂ (b / c)}

Next, a keyword t that can maximize the value of the mutual information I (t _k ) from a plurality of keywords t _j
k is selected and set as a search keyword (step S
205). This mutual information I (t _k ) has a positive finite value (>
In the case of 0) (step S206), the search keyword t _k
Set _i 'a document set consisting of number of documents containing the flag of the search keyword set of documents consisting of number of documents that do not contain t _k Set _i "divided as, for each document set divided" undivided " Turn on (step S207)
To S210). i ′, i ″ are already document sets Set _i ′, S
If there is et _i "good at any value. If the mutual information I (t _k) is zero value (= 0), it does not perform the division of document set (step S206). Thereafter, the set Set _i
Is turned off (step S21).
1). If there is a document set whose flag of “undivided” is on, the process returns to step S203 (step S212: Ye
s) The process is repeated until there is no document set for which the “undivided” flag is on. Then, the process ends when the “undivided” flag is turned off for all the document sets (step S212: No).

The processing by the above algorithm "ID3" is performed, for example, by a known algorithm "C4.5".
It is also possible to substitute the above method. For details of this "C4.5", see "C4.5 Programs for Machine Learni".
ng "(Quinlan, JR, Morgan Kaufmann Publishers
Publication) can be referred to.

FIG. 7 is an explanatory diagram showing a process in which the above-mentioned search formula creating apparatus 50 divides one document set into a plurality of document sets and creates a search formula by trial and error. Hereinafter, with reference to FIG. 7, a description will be given of a conventional search formula creation procedure. First, from the initial document set Set ₀ output from the keyword extraction unit 51, a keyword that maximizes the mutual information is determined based on the above-described decision tree learning algorithm “ID3”, and is set as a search keyword. Here, it is assumed that the search keyword kwd3 has been determined. Then, by this search Kwd3, divides the initial document set Set _0, the search keyword Kwd3 in a set Set ₂ required documents and unnecessary documents that do not contain the set Set ₁ the search keyword Kwd3 necessary documents including.

The document set Set ₁ is composed of a search keyword kw
No further division by d3 is possible, while the document set Set ₂ can be further divided. Therefore,
Mutual information in this document set Set ₂ determines a search keyword Kwd2 which maximizes the document set Set ₂ by this search Kwd2, need to include a search keyword Kwd2 a set Set ₃ of unnecessary documents that do not contain a search term Kwd2 Set _{4 of} unnecessary documents and unnecessary documents
And split into

Since the document set Set ₄ can be further divided, the keyword kwd1 that maximizes the mutual information in the document set Set ₄ is determined as a search keyword, and a set Set _{5 of} necessary documents including the search keyword kwd1 is determined. , And a set Set _{6 of} documents that do not include the search keyword kwd1. Document Set Set ₅ and Se
At t ₆ , the division process ends because no further division is possible. The plurality of search keywords kwd1 to kwd3 determined in the division processing are sequentially stored in a storage unit (not shown), and are passed to the search expression creation unit 53 when the division processing ends.

In the search formula creating section 53, the document set dividing section 5
Each search keyword that is the result of 2 is assigned to the logical operator “an
d "," or "and" not "
Create a query. Reference numeral 53B denotes a search formula creation unit 53.
3 illustrates a search expression output from.

[0018]

In the above-mentioned "ID3" algorithm for decision tree learning, a search keyword used for dividing a document set is determined based on required / unnecessary designation information given to a document and a mutual information amount. The search accuracy cannot be improved unless a large number of documents to which the designated information from the searcher is added are used as the learning documents. In other words, to create an accurate search expression,
Large amounts of learning documents are required.

However, the designated information is determined and given by the searcher himself, and the required number of documents for creating an accurate search formula is not always provided. Therefore, a keyword included in a document to which no designation information is added is not reflected in a search formula even if it is an important keyword for a searcher, and an accurate search formula cannot be created. This is due to the fact that as a result of completely dividing the document set based on the decision tree learning algorithm “ID3”, the created search formula corresponds only to the required document specified by the searcher. is there. That is, the group of documents searched by the search formula is a group of learning documents known to the searcher. Even if a search formula corresponding to only the necessary document specified by the searcher is created, the effectiveness is poor. Further, for example, if the searcher attempts to sequentially determine and assign designation information for all documents, a large amount of work is required, and the load on the searcher increases.

According to the present invention, there is provided a computer apparatus for determining a search keyword and creating a search formula in consideration of a document group to which designation information is not added without requiring a large number of learning document groups. An object of the present invention is to provide an improved method for creating a search formula, which can be performed by using the method. Another object of the present invention is to provide a search formula creation device suitable for implementing the above search formula creation method.

[0021]

SUMMARY OF THE INVENTION In order to solve the above-mentioned problems, the present invention provides the following two forms of a retrieval formula creation method. (1) A plurality of words are extracted from a learning document group to which necessary / unnecessary designation information is added in advance for each document and an indefinite document group whose designation information is unknown, and the appearance of each extracted word is determined for each document. And the information amount of the document group including the word and the document group not including the word is converted into a single information in which the mutual information amount obtained by the difference from the total information amount of the learning document group and the indefinite document group is maximized. Determining a word as a search keyword; and suppressing the determination of the search keyword when a document group including and not including the search keyword can be distinguished into a necessary document group and an unnecessary document group. Creating a search formula for use in document search by combining one or more search keywords obtained by a logical formula, and creating a search formula reflecting words of the indefinite document group. .

(2) A plurality of words are extracted from a group of learning documents to which necessary / unnecessary designation information has been added in advance for each document and a group of indefinite documents whose designation information is unknown, and the appearance of each extracted word. And the mutual information obtained from the difference between the information amount of the document group containing the word and the information amount of the document group not including the word from the total information amount of the learning document group and the indefinite document group is maximized. A process of determining a single word as a search keyword, and a group of documents including and not including the search keyword can be distinguished into a necessary document group and an unnecessary document group, and a document group including the search keyword. When the required number of documents exceeds a predetermined ratio with respect to the number of indefinite documents, and a process of combining the determined one or more search keywords by a logical expression and using the determined keyword in a document search. Method characterized by comprising the steps of creating a formula, and creates the reflecting word indefinite documents search expression.

A search formula creation device according to the present invention that solves the above-mentioned other problems includes a predetermined document group, for example, a group of necessary documents of interest to a searcher, a group of unnecessary documents of no interest, and a group of indefinite documents of unknown interest. And a search formula for searching for a specific document from a group of documents to which necessary, unnecessary, or unspecified designation information is added, which is used for determination in determining the search keyword. An apparatus, which includes the following elements.

(1) The document group is subjected to morphological analysis to extract a plurality of words, discrimination information indicating whether or not each of the extracted words is included in the document, a necessary document, an unnecessary document, Alternatively, a document set generation means for generating a document set in which designation information indicating an unfixed document is collected together with identification information of each document. The document set generation means preferably includes a word database constructed with identification information indicating whether or not the extracted individual words appear in the document together with identification information of each document. (2) Based on the mutual information amount obtained as a difference from the total information amount of the document group and the number of documents in which the word appears, the information amount of the document group including and not including the individual word is determined based on the number of documents in which the word appears. Is determined as a search keyword,
Document set dividing means for dividing one document set into a plurality of document sets using the determined search keyword. (3) Search expression creating means for combining the search keywords used for dividing the document set with a logical expression to create the search expression.

The document set dividing means is constituted, for example, as follows. (2-1) The mutual information amount is calculated by virtually considering the indefinite document group as the unnecessary document group based on a predetermined minimum description length principle and increasing the number of learning cases. (2-2) A word having the maximum mutual information is sequentially determined as the search keyword based on a predetermined minimum description length principle, and a document set of the document group is determined using the determined search keyword. Divide into document sets. (2-3) The determination of the search keyword is stopped when at least one of the plurality of divided document sets is classified into a necessary document group or an unnecessary document group. (2-4) When at least one of the divided document sets is classified into a necessary document group or an unnecessary document group, and the required number of documents in the document set related to the required document group is: The determination of the search keyword is stopped when a predetermined ratio to the number of indefinite documents is exceeded.

[0026]

Embodiments of the present invention will be described below in detail. FIG. 1 and FIG. 2 are functional configuration diagrams of a search formula creation device suitable for implementing the above search formula creation method. FIG.
Constituent elements having the same functions as those of the conventional search formula creation device 50 described in the above are denoted by the same reference numerals, and redundant description will be omitted.

The retrieval formula creation device 10 of the present embodiment includes a keyword extraction unit 11 and a document database (hereinafter, referred to as “D” in this specification) formed by a computer device reading and executing a predetermined program.
B), a keyword DB 13, a document set division unit 14, and a search expression creation unit 33.

The above-mentioned program is usually stored in an internal or external storage device of a computer device and read and executed as needed. However, a recording medium distributed in a form separate from the computer device, for example, a CD −
It may be stored in a portable medium such as a ROM or FD, installed in the internal or external storage device at the time of use, and provided for execution at any time.

In the present embodiment, a document group as a learning example (hereinafter referred to as a learning document or a learning document group) is provided with necessary / unnecessary designation information indicating whether it is a necessary document or an unnecessary document by a user or the like in advance. Documents that do not have this required / unnecessary designation information are indefinite documents (non-designated documents)
It is assumed to be determined as Thus, in the present embodiment, there are document groups classified into three types: required, unnecessary, and undefined.

The document DB 12 stores a group of learning documents and a large amount of indefinite documents. The keyword DB 13 stores individual documents for all the keywords included in the learning documents and the indefinite documents in the document DB 12 in advance. Is stored in association with each piece of document identification information such as a document number. Reference numeral 13A is an example of information in the keyword DB 13.

The keyword extracting unit 11 performs a morphological analysis on the input required documents and unnecessary documents, or the documents stored in the document DB 12, and performs a keyword extraction process for each document. Further, it outputs determination information indicating presence / absence of a keyword in each document and designation information of the document as a document set together with document identifiers such as a document name and a document number. Reference numeral 11A exemplifies the contents of a document set output from the keyword extraction unit 11. In this case, the specified information is, for example, necessary / specified by the searcher /
In addition to the unnecessary learning document group, a document set is created by adding an “undefined” tag, and is input to the document set dividing unit 14.

The document set dividing section 14 determines a search keyword used for dividing the document set output from the keyword extracting section 11. To divide a document set, the above-described decision tree learning algorithm “ID3” (hereinafter simply referred to as “I3
D3 ”). In general, most documents in the document DB 12 are a group of documents that are not related to the interest of the searcher, and the ratio of the group of documents necessary for the searcher is considered to be extremely small. That is, the document DB 12
Can be said to be virtually all unnecessary documents. Therefore, the document set dividing unit 14
In the first embodiment, the number of learning cases, that is, the number of learning documents is increased to calculate the mutual information amount by treating the indefinite documents in the document DB 12 as unnecessary documents. In the present embodiment, since the document group is classified into three types as described above, the mutual information I (t _j ) of the document set Set _i represented by the above equation (1) by the decision tree learning algorithm “ID3”.
The information amounts H and H (t _j ) for obtaining are expressed by the following equations (4) and (5), respectively.

[0033]

However, the parameters in the equations (4) and (5) are as follows. p _i: Set _i need the number of documents in, n _i: unnecessary number of documents and the sum of the indefinite number of documents in _{Set i, s i: p i} + n i, p i (t j): keyword in Set _i t the number of necessary documents, including the _{j, n i (t j)} : sum of Set _i unnecessary number of documents that contain the keyword t _j in the indefinite number of _{documents, s i (t j):} p i (t j) + n i (t _{j), p i not (t} j): Set i need the number of documents that do not contain the keyword t _{j in, n i not (t j)} : Set i unnecessary number of documents and indefinite number of documents that do not contain the keyword t _j in S _i not (t _j ): p _i not (t _j ) + n _i not (t _j ), h (a, b, c): − {a / c · log ₂ (a / c) + b / clog ₂ (b / c)｝

Equations (4) and (5) are the same as Equations (2) and (3) above, but are shown because the form of the document set is different from that of the above-described method that does not take into account the indefinite document group. It is a thing. Evaluating the expressions (4) and (5),
The search formula created as a result of completely dividing the document set based on “ID3” corresponds only to the required document group specified by the searcher. That is, the group of documents searched by the search formula is a group of learning documents known to the searcher, and even if a search formula corresponding to only the necessary document specified by the searcher is created, the effectiveness is poor. Therefore, in the present embodiment, the process of dividing the document set is stopped halfway, and an optimal search formula corresponding to an indefinite document group is created. In this example, the stop of the document set division process will be specifically described with reference to the following two methods.

<First Method> At the stage (time point) when the necessary document group and the unnecessary document group are distinguished in the process of the division process, the division of the set is stopped.

<Second Method> The required document group and the unnecessary document group are distinguished from each other, and division of the set is stopped when the required number of documents with respect to the number of indefinite documents in the document set exceeds a predetermined ratio. I do.

The predetermined ratio in the second method is as follows.
For example, as in the case where "the number of required documents is half of the number of indefinite documents, that is, 50% or more", a predetermined ratio reference value is appropriately set in advance using system parameters or the like. The first and second methods can be used independently. In addition, the second method includes the first method, and provides a more narrow stop condition as compared with the first method.

FIG. 3 is a processing procedure diagram showing an outline of the processing in the search expression creating apparatus 1 of the present embodiment when the above-described second method is used. In the keyword extracting unit 11, the document D
An indefinite document group to be searched for in B12,
Then, all keywords are extracted from the learning document group and stored in the keyword DB 13 (step S101). The keyword extraction unit 11 also creates an initial document set to be searched for based on the data in the keyword DB 13 (step S102). Document set division unit 14
Then, the initial document set is divided (step S10).
3).

If it is determined in step S103 that at least one of the two document sets divided by the search keyword is a necessary or unnecessary document group (step S104: Yes), the document set dividing unit 1
In step S105, it is detected whether or not the required number of documents exceeds a predetermined ratio with respect to the number of indefinite documents in the set of distinguished documents.
s), the process of dividing the document set is stopped (step S10)
6). Steps S104 and S105 are processing corresponding to step S212 in the processing procedure of "ID3" shown in FIG. 6, and are provided in the present invention in order to reflect the indefinite document group in the search formula. On the other hand, if at least one of the two document sets divided in step S103 is not distinguished as a necessary or unnecessary document group (step S104: No), or the required number of documents in the document set is indeterminate If the predetermined ratio to the number of documents is not exceeded (step S10
5: No), the process returns to step S103, and the process of dividing the document set is repeated.

The search formula creation unit 53 uses the logical operators “and”, “and”, as the search keywords obtained in the document set division process up to the point where the document set division process is stopped.
Search expression "quer" linked by "or" and "not"
y is created (step S107).

It should be noted that in this processing procedure, step S104
And 105 correspond to the processing procedure of the second method. If the above first method is used alone, step S10
When the distinction between the necessary and unnecessary document groups in 4 is detected, the process of dividing the document set is stopped.

As described above, according to the above-described processing procedure, the result that the determination of the search keyword is suppressed based on the stop condition in steps S104 and S105, that is, the result that the document set is not completely divided is obtained. The indefinite document group is reflected in the search formula.

FIG. 4 is a schematic diagram showing a stop state of the document set dividing process according to the first and second methods. In this figure, the initial document set is divided into a document set including the keyword "kwd1" and a document set not including the keyword. Next, a document set including the keyword “kwd1” is added to the keyword “kwd”.
The document set is divided into a document set containing wd2 "and a document set not containing it. Here, a document set containing the keyword" kwd2 "and a document set not containing it can be distinguished into a required document group and an unnecessary document group. Therefore, at this point, the determination of the search keyword is stopped, which means that the stop condition according to the first method has been met, and the necessary document group and the unnecessary document group distinguished by the keyword “kwd2” are added. Includes an unfixed document group including the keyword “kwd2” and an undefined document group not including the keyword “kwd2”.
The mutual information amount at the time of determining wd2 ″ is calculated by virtually assuming the indefinite document group as an unnecessary document group, but does not regard the indefinite document as an unnecessary document when determining the condition for stopping the division process. .

In the document set including the keyword "kwd2", the determination of the search keyword is stopped when the required number of documents becomes a predetermined ratio or more with respect to the indefinite number of documents. This means that the stop condition in the second method has been met.

As described above, in the retrieval formula creating apparatus 10 of the present embodiment, the unspecified indefinite document group is virtually regarded as an unnecessary document group (only when calculating the mutual information: division processing). When determining the stop condition of, the indefinite document is not regarded as an unnecessary document.) The decision tree learning algorithm “I
Since the mutual information amount is calculated based on “D3” and the document set is divided, compared with the conventional method, the number of learning documents to which the designated information from the searcher is added is small. become.

Further, a search formula reflecting words included in an indefinite document group can be created, and search accuracy and practicality for a searcher can be compared with a search formula according to a conventional method corresponding only to a necessary document group. Can be obtained.

Further, since a small number of learning documents is required as described above, the work of the searcher successively determining and assigning the designation information for all the documents as in the related art is reduced, and the searcher side is reduced. Such a load can be reduced.

[0049]

As is apparent from the above description, according to the present invention, it is possible to create a retrieval formula in consideration of a document group to which no specified information is added without requiring a large amount of learning document group. Has the specific effect of becoming Further, by using the created search formula in the search processing, the search accuracy for the searcher can be maintained at a certain value or more, and a highly practical search result can be obtained.

[Brief description of the drawings]

FIG. 1 is a functional block diagram illustrating an embodiment of a search formula creation device according to an embodiment of the present invention.

FIG. 2 is a functional block diagram illustrating an embodiment of a search formula creation device according to an embodiment of the present invention.

FIG. 3 is a processing procedure diagram in the search expression creating apparatus of the embodiment.

FIG. 4 is a schematic diagram showing a process of dividing a document set according to the embodiment.

FIG. 5 is a functional block diagram of a conventional search expression creation device.

FIG. 6 is an explanatory diagram of a processing procedure in a conventional search expression creation device.

FIG. 7 is a schematic diagram of information obtained in a conventional dividing process.

[Explanation of symbols]

 10, 50 search formula creation device 11, 51 keyword extraction unit 14, 52 document set division unit 12 document DB 13 keyword DB 53 search formula creation unit

Claims (10)

[Claims] 1. A method for extracting a plurality of words from a learning document group to which necessary / unnecessary designation information is added in advance for each document and an indefinite document group whose designation information is unknown, and determining the appearance of each extracted word. The process of detecting for each document, and the information amount of the document group containing the word and the document group not containing the word,
A step of determining, as a search keyword, a single word that maximizes a mutual information amount obtained from a difference from the total information amount of the learning document group and the indefinite document group, and a necessary document group and an unnecessary document depending on whether the search keyword exists. A step of inhibiting the determination of the search keyword when it can be distinguished from a group, and a step of combining the determined one or more search keywords with a logical expression to create a search expression used for document search; A search formula creation method using a computer device, wherein a search formula reflecting words of the indefinite document group is created. 2. A plurality of words are extracted from a learning document group to which necessary / unnecessary designation information is added in advance for each document and an indefinite document group whose designation information is unknown, and the appearance of each extracted word is determined. The process of detecting for each document, and the information amount of the document group containing the word and the document group not containing the word,
A step of determining, as a search keyword, a single word that maximizes the mutual information obtained as a difference from the total information amount of the learning document group and the indefinite document group; a document group including the search keyword and a document group not including the search keyword However, when the required document group and the unnecessary document group can be distinguished from each other, and the required document number of the document group including the search keyword exceeds a predetermined ratio to the indefinite document number, the determination of the search keyword is suppressed. Creating a search formula for use in document search by combining the determined one or more search keywords with a logical formula, and creating a search formula that reflects the words of the indefinite document group A method for creating a retrieval formula using a computer device. 3. The learning method according to claim 3, wherein the indefinite document group is virtually regarded as an unnecessary document group of the learning document group based on a predetermined minimum description length principle, and the number of learning cases is increased and reflected in the calculation of the mutual information amount. The method according to claim 1 or 2, wherein 4. An apparatus for creating a search formula for searching for a specific document from a group of documents, comprising: performing a morphological analysis on the group of documents to extract a plurality of words;
Generates a document set in which identification information indicating whether or not each extracted word is included in a document, and designation information indicating whether the document is a required document, an unnecessary document, or an indefinite document are collected together with identification information of each document. A document set generating means, and a mutual information amount obtained as a difference between the information amount of the document group including and not including the individual word from the total information amount of the document group, and the number of documents in which the word appears. Document set dividing means for determining a single word as a search keyword based on the determined keyword, and dividing one document set into a plurality of document sets using the determined search keyword; and a search used in dividing the document set. A search formula creating means for creating the search formula by combining keywords by a logical formula. 5. The document group includes a required document group that is of interest to the searcher, an unnecessary document group that is not interested in the document group, and an indefinite document group that is of unknown interest, each of which is used for determination when determining the search keyword. 5. The search expression creating apparatus according to claim 4, wherein any of necessary, unnecessary, or indefinite designation information is used. 6. The document set generating means includes a word database constructed by constructing discrimination information indicating whether or not the extracted individual words appear in a document together with identification information of each document. 5. The retrieval formula creation device according to claim 4, wherein: 7. The document set dividing unit calculates the mutual information amount by virtually considering the indefinite document group as the unnecessary document group based on a predetermined minimum description length principle and increasing the number of learning cases. 5. The retrieval formula creation device according to claim 4, wherein the retrieval formula creation device is configured as follows. 8. The document set dividing means sequentially determines a word in which the mutual information amount is maximum as the search keyword based on a predetermined minimum description length principle, and uses the determined search keyword to generate the document group. 5. The retrieval formula creation device according to claim 4, wherein the document set for is divided into a plurality of document sets. 9. The document set dividing means stops the determination of the search keyword when at least one of the plurality of divided document sets is classified into a required document group or an unnecessary document group. 5. The retrieval formula creation device according to claim 4, wherein the retrieval formula creation device is configured. 10. The document set dividing means, when at least one of a plurality of divided document sets is classified into a necessary document group or an unnecessary document group, and a document set related to the necessary document group. 5. The search expression creating apparatus according to claim 4, wherein the determination of the search keyword is stopped when the number of required documents in the document exceeds a predetermined ratio to the number of indefinite documents.