JP3472032B2 - Information filter device and information filter method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention determines, on behalf of a user, whether an acquired document is worth reading, and presents to the user only a document that is judged to be worth reading. And an information filtering method.

[0002]

2. Description of the Related Art In recent years, with the increase in capacity and price of information devices and information storage media, an enormous amount of computerized information has been distributed on various media. As the amount of information increases in this way, it is impossible for the user to read all the information, and it becomes necessary to select information that is worth reading. However, it is difficult to select only necessary information from a large amount of information with general users' limited ability and time.

Therefore, various information filters have been proposed which evaluate the contents of a newly arrived document in advance on behalf of the user and present only to the user what is considered to be worth reading. Such a conventional information filter selects or evaluates a document by using a keyword related to the content of the document, and (1) a rule such as a search formula is registered in advance, (2) a neural network Those using the net are known.

However, in the conventional rule description type information filter, it is necessary for the user side to prepare an explicit description of a rule such as a search formula using a keyword. Here, it is inconvenient if the user's interest or the like changes, unless the user redefines the rule each time. Further, in order to describe a rule so as to sufficiently satisfy the user's request, some experience and trial and error are required, and it is not a simple thing to describe an appropriate rule.

Further, in the conventional information filter using the neural network, the neural network learns the user's preference and selects the information based on the learned information, so that the user has no merit of having to write rules and experience. On the other hand, on the other hand, on the neural network side, a huge amount of calculation is required to learn the preference of the user, and if the amount of calculation is suppressed, a sufficient learning effect cannot be obtained.

As described above, the conventional information filter that selects the information required by the user from a large amount of information and presents it to the user is a rule description type information filter. , It is necessary for the user to explicitly describe the rule, and when the interest of the user changes, the user needs to redefine the rule again, which is inconvenient in use.

On the other hand, the conventional information filter using the neural network requires a huge amount of calculation to learn the preference of the user, and if the amount of calculation is suppressed, a sufficient learning effect cannot be obtained. , I can not select the necessary information properly.

[0008]

As described above, in the conventional information filter in which the information required by the user is selected from a large amount of information and presented to the user, the rule description is used. In the case of the type information filter, the user has to explicitly describe the rule, and when the user's interest changes, the user has to redefine the rule again, which is inconvenient.

On the other hand, in the conventional information filter using the neural network, a huge amount of calculation is required to learn the preference of the user, and if the amount of calculation is suppressed, a sufficient learning effect can be obtained. However, there was a problem that necessary information could not be selected accurately.

Therefore, any type of information filter is not easy to use, requires a great improvement, and is not practical. Therefore, the development of a more practical information filter is desired.

Therefore, the object of the present invention is to
When selecting the information required by the user from a large amount of information, it is possible to reflect changes in the user's preference without unnecessary effort, and the estimation accuracy is good,
Another object of the present invention is to provide a highly practical information filter device and information filter method that can perform inference / learning with a small amount of calculation and at high speed.

[0012]

According to the present invention, in selecting information required by a user from a large amount of information,
It is intended to provide a highly practical information filter device, and particularly to provide the following information filter.

(1) It should be possible to reflect a change in user's preference without unnecessary trouble. (2) The estimation accuracy is good. (3) Inference / learning can be performed quickly with a small amount of calculation. Then, in order to achieve the above object, the present invention is as follows. That is, multiple keywords are extracted.
The keyword extracting means to be issued and the plurality of keywords
Keyword that has one or more of these keywords as elements
A combination of codes is generated and the key is added in the document to be learned.
Whether each keyword included in the word combination appears
If it appears, the first identifier is
If not, combine the second identifier with the keyword
, The first identifier is counted, and the counting result is an even number.
If it is a number, the third identifier is used. If it is an odd number, the fourth identifier is used.
Spectrum is assigned by assigning bespoke to the keyword combination
Generated and associated with the spectrum and the document
Based on the evaluation value
The coefficient is calculated and the coefficient is paired with the keyword combination.
Learning means for learning by responding and storing and reasoning
Among the target documents, there is a combination of the above keywords
Based on the coefficient and the spectrum for
Inference means for inferring the importance of a document to be inferred is provided.

[0014]

[0015]

Further, preferably, the learning means, if a new keyword other than the keyword extracted by the keyword extracting means is extracted, the learning in the past is performed.
It is characterized by further comprising keyword adding means for adding the new keyword to the learning result by the means .

[0017] In a preferred embodiment, the learning means, the
At the time of learning, learning is performed only for the range in which the keyword extracted from the document influences, and the inference means includes the keyword extracted by the extraction means in the keyword extracted from the document to be inferred. not the advance holds the importance of the document in the case of, when inferring the importance of the inputted document, <br/> keywords extracted by the extraction unit to the keywords extracted from the document When the keyword is present, a value for varying the importance retained by the keyword is obtained, and the importance retained is corrected based on this value, whereby the inputted importance is calculated. The feature is that it is desired.

[0018]

[0019] Preferably, the extracted keywords in all keywords in extracted from the document, when the percentage of the total keywords in that the extracted is less than the predetermined value, the inference characterized by being configured to be displayed to the user without the inference by means.

According to the present invention, a keyword is extracted from an input document, and information on the relationship between a predetermined combination of keywords stored in the storage means and an evaluation value relating to the importance of the document is extracted from the input document. Based on a predetermined keyword (for example, a keyword registered in the table) among the keywords, the degree of importance of the input document is inferred. Then, the document determined to be important by the inference means is presented to the user.

[0021] The inference is performed based on the spectral theory (fast spectral theory). Further, in the present invention,
An information filter that infers the importance of the document based on a keyword extracted from the input document and displays the document when the document is determined to be important. When learning the information on the relationship between the combination of the obtained keywords and the evaluation value regarding the importance of the document, a predetermined keyword (for example, a keyword registered in the table) among the keywords extracted from the document to be learned, Using the obtained evaluation value, the relationship between the combination of keywords and the evaluation value is learned. This learning,
It is performed based on spectrum theory.

According to the present invention, since inference / learning is performed only on the basis of the relationship between the combination of keywords extracted from the document to be evaluated and the evaluation value regarding the importance of the document determined by this combination of keywords, it is preferable. It is possible to obtain an information filter device capable of performing determination / learning at high speed while maintaining the estimation accuracy and reducing the calculation amount. Therefore, according to the present invention, the user does not need to read documents that are obviously not of interest.

Since the present invention can be distributed by being stored in a portable storage medium as application software that can be read and executed by a computer,
The following forms are also included in the scope of the invention.

[1] A computer-readable storage medium which is used with an information display device such as a computer and stores a processing program for displaying a predetermined document by filtering information from an input document in the medium. A program code means for extracting a keyword from an input document, a program code means for storing a relationship between a predetermined combination of keywords and an evaluation value given to the document, the stored relationship and the input And a program code means for inferring the importance of the input document based on a predetermined keyword among the keywords extracted from the read document, and a readable storage storing a processing program in a medium. Medium.

[2] Program code means for judging whether or not the importance obtained by the inference satisfies a predetermined condition, and when it is judged by the above judgment that the predetermined condition is satisfied, The readable storage medium according to the above [1], wherein the processing program further comprises: a program code unit for displaying predetermined information regarding the input document to the user.

[3] Program code means for inputting an evaluation value for the user based on the displayed predetermined information, and a keyword extracted from the document,
[2] The processing program stored in a medium, further comprising a program code means for learning information on a relationship between the combination of the keywords and the evaluation value based on the input evaluation value. A readable storage medium as described.

[4] The learning program code means is a medium for a processing program further including program code means for adding the new keyword when a new keyword other than the predetermined keyword is extracted. The readable storage medium according to [3], which is stored in the storage medium.

[5] The program code means for learning includes program code means for learning a range in which a keyword extracted from the document influences when learning the relationship between the combination of the keywords and the evaluation value, The program code means for inferring holds the importance of the document in advance when the keyword extracted from the document does not include a predetermined keyword, and when inferring the importance of the input document, When a keyword extracted from the document includes a predetermined keyword, a value for varying the importance held by the keyword is obtained, and the held importance is corrected based on this value. By storing the processing program including the program code means for determining the inputted importance level in the medium. A readable storage medium according to [3].

[0029]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Specific examples of the present invention will be described below with reference to the drawings.

(First Specific Example) FIG. 1 is a diagram showing a configuration of an information filter device according to a first specific example of the present invention.
The information filter device according to the first specific example of the present invention includes a document input unit 1, a document storage unit 2, a keyword extraction unit 3, a data storage unit 4, an inference unit 5, a display unit 6, and an evaluation data input unit 7.
And a learning unit 8.

The document input unit 1 is for inputting electronic data of a document (hereinafter, simply referred to as a document) from the outside. The document input unit 1 has a network connection device and a wireless reception device according to the transmission form of the document. , A magnetic disk / tape reader, a CD-ROM reader, etc. can be used.

The document storage unit 2 is for temporarily storing a document input from the outside, and includes a magnetic disk device,
A desired device such as a magnetic tape device, an optical disk device, or a semiconductor memory can be used.

The keyword extracting unit 3 extracts a predetermined keyword from the new document temporarily stored in the document storage unit 2. When extracting keywords from a document, the keywords to be extracted are determined as follows.

Since there are a form in which the keyword is added to the document in advance and a form in which it is not added, the following method may be applied to the extraction of the keyword.

When no keyword is added to the document in advance, (1) publicly known keyword extracting means (for example, Iwao Ishikawa et al .: "Subject index creation support system based on document analysis process", IPSJ Journal Vol. .132, 1991
The keyword is extracted from the document using the keyword extracting means disclosed in (1).

In the case of an English sentence, the stem of the word is extracted.

When a keyword is added to a document in advance, the following means are used.

[1] Using the keyword extracting means of the above (1), [2] extracting the keyword added to the document.
Then, the extracted one is used as a keyword.

[3] Both of the above [1] and [2] are used together.

By doing so, the required keyword is determined, and the keyword corresponding to the keyword is searched and extracted from the document.

The data storage unit 4 stores and holds data, and details are shown in FIG. 2 which is used to create a coefficient (αs described below) used for inference based on spectrum theory described later, an input vector and a case vector. The keyword table as shown and / or the learning result of the learning unit 8 and the like are stored.

The keyword table as shown in FIG.
For example, keywords such as "word processor", "dictionary",
Words (keywords) such as "induction", "learning", "information", "filter", "optics" ... Are registered together with keyword numbers. Specifically, for example, the word "word processor" is registered with the keyword number "1", the word "dictionary" is registered with the keyword number "2", and the word "induction" is " 3 "
It is registered with the keyword number.

The inference unit 5 extracts the keyword group extracted by the keyword extracting unit 3 from the input document into the keyword extracting unit 3
The keyword number of the keyword received from the keyword extraction unit 3 is obtained using the keyword table stored in the data storage unit 4 and the input vector is generated, and then the Inference based on spectrum theory is performed using the coefficient αs of. The output in this inference is information indicating whether the document should be presented to the user. For example, "1" is output when it is determined that the document should be presented to the user, and "-1" is output otherwise.

The display section 6 has a control function section 6a and an output section 6b. The control function unit 6a is for presenting to the user the contents of the document determined to be presented to the user as a result of the above inference by the inference unit 5.
The document instructed from is read from the document storage unit 2 and output is controlled. The output unit 6b displays or prints out the document based on the output control by the control function unit 6a, and may be a display device, a printer device, a voice output device, or the like.

The evaluation data input unit 7 inputs the evaluation data for the presented document. The evaluation data is, for example, information indicating whether the document is worth reading. As the evaluation data to be actually input, if the document is worth reading, the evaluation value “1” may be input, and if not, the evaluation value “−1” may be input. Form (for example, "O", "x" or "true,"
false ”etc.) key input or selection input,
The evaluation data input unit 7 may convert the evaluation values into “1” and “−1”.

The learning unit 8 learns the coefficient αs based on the evaluation value (“+1” or “−1”) obtained from the input to the evaluation data input unit 7.

The general operation of the information filter device of this example will be described. The operation of the information filter device of this specific example is roughly divided into two processes: an inference process including an actual information filtering process and a selective presentation operation of the result, and a learning process for obtaining a good inference result. .

In the process of inference, the document input unit 1
The document newly input from is temporarily stored in the document storage unit 2. Then, the keyword extraction unit 3 extracts keywords from this document. The inference unit 5 receives the keyword group extracted with respect to the document, obtains the keyword number of the appearing keyword by using the keyword table as shown in FIG. 2, generates the input vector, and calculates the current coefficient αs. We use it to make inferences based on spectrum theory.

Then, the inference unit 5 instructs the display unit 6 to display the document when it is determined that the document should be presented to the user as a result of the inference based on the spectrum theory. When this instruction is received from the inference unit 5, the display unit 6 displays the document storage unit 2
The content of the document determined to be presented, which is stored in, is read out, output to the output unit 6b, and presented to the user.
The user will read the document output from the output unit 6b.

The learning process proceeds as follows.

The user inputs the evaluation for the document presented on the display unit 6 from the evaluation data input unit 7 by the inference process. The inputted evaluation data is given to the learning unit 8, and the learning unit 8 learns the coefficient αs based on the given evaluation value based on the spectrum theory.

This learning process is performed at the time of learning associated with the process of initial learning and inference based on teaching data given in advance. That is, in order to obtain a good prediction result, teaching data is given and the coefficient αs described later is calculated. In addition, when the user gives an evaluation to the document taught in the inference process, the coefficient α
s is calculated, and learning will proceed.

Inference / learning based on the spectrum theory performed by the inference unit 5 and the learning unit 8 will be described. Here, the spectrum theory is described, for example, in "Const by Nathan Linial et al.
antDepth Circuits, Fourier Transform, and Learnabi
lty ", Jouranl of the Association for Computing Mac
hinery, Vol. 40, No. 3, July 1993, pp. 607-620).

In the spectrum theory, a pair of an input (problem) and an output (correct answer) is given to learn the parameters of the evaluation function, and when a new problem that has not been input so far is given, The correct answer is inferred using the parameters at that time. Of course, the correct answer can be obtained even when a problem that has been input is given.

When such a spectrum theory is applied to an information filter device, a set of keywords or a combination of its elements is input (problem), and an evaluation value corresponding to this is output (correct answer) and extracted from a new document. The combination of the specified keywords is given as an input, and the evaluation value corresponding to this is inferred.

That is, in the inference and learning of the present invention, a plurality of cases whose contents and evaluations are known in advance are given, and what kind of input (output) (evaluation value) is output is learned in advance. This is to predict what kind of evaluation value an input answer will have after learning. The present invention uses spectrum theory for such inference and learning.

The spectrum theory will be described more specifically. For example, learning is performed by giving a case including the following input vector (input) and output (evaluation value) as a learning example.

Input (0,1,1,1,1) Output “+1” Input (1,0,1,1,1) Output “−1” Input (1,1,0,1,1) Output “ +1 ”input (1,1,1,0,1) output“ −1 ”input (1,1,1,1,0) output“ +1 ”That is, the spectrum theory method inputs a predetermined input vector. In this case, what kind of output (evaluation value) is obtained according to the input is learned, and after the learning is completed, for example, the input vector becomes “input (1, 1,
It is a method of performing inference that predicts whether the answer in the case of "1, 1, 1)" is "+1" or "-1".

In order to apply such a method of spectrum theory to an information filter, in the inference unit 5 in the system of the present invention, a keyword (extracted keyword) extracted by the keyword extracting unit 3 from the document to be inferred is extracted. ) Are arranged in a uniform order. The unified order means, for example, the order of arrangement of the keywords in the keyword table, but the order may be appropriately edited and changed by the user. Then, the extracted keywords are checked in order by referring to a keyword table as shown in FIG. 2 to see if the keyword corresponds to a word (keyword) registered as an object of interest.
An input vector is created by substituting "1" if it corresponds to the registered one and "0" if not. Specifically, the extracted keyword groups are input one by one, and if there is a match with the registered keyword by comparing with the registered keyword, “1” is set, and if there is no matching keyword, “0” is set. Generate a vector.

In learning, in this specific example, the output (evaluation value) is "1" for a document of which the user is interested, and the output (evaluation value) is for a document of which the user is not interested. ) Is set to “0”, and by using this, a case by an input / output set (a set of an input vector and an evaluation value) is configured, and some cases are given as learning documents to learn the parameters used for inference. To do.

For example, the keywords shown in FIG. 2 are registered in the system in advance as the keywords of the field in which the user is interested, and ““ information ”,
The five keywords "filter", "learning", "user", and "algorithm" are "" ... "," ... ",
"...", "information", "filter", "learning", "...",
"User", "...", "...", "...", "Algorithm", "...", "...", "...""(However," ... "is a term not registered as a keyword in the keyword table. However, in this case, it is generated by the inference unit 5 for this document in this case. The input vector is (0,0,0,1,1,1,0,1,0,0,0,1,
0,0,0).

When learning, the user reads the presented document, judges whether or not it is worth reading, and the user makes an evaluation to that effect through the evaluation data input unit 7. input. From the evaluation data input unit 7, an evaluation value “+1” (worthy to read) and an evaluation value “−1” corresponding to the evaluation.
(Not worth reading) is given to the learning unit 8, so that the learning unit 8 receives the given evaluation values “+1” and “−”.
The coefficient αs is calculated based on 1 ″, and the learning is performed based on the coefficient αs.

In this way, the user inputs the evaluation as to whether or not it is worth reading as a result of reading the displayed document by the data input unit 7, so that the learning unit 8 responds to the evaluation from the data input unit 7. Evaluation value (“+1” or “−”
The coefficient αs is learned based on 1 ″).

Next, the learning method and the inference method based on the spectrum theory will be concretely shown.

First, the elements used in learning and inference are defined and explained. The elements used here include X,
There are Xi, S, s, χs (Xi), αs, f (αi).

Of these, the entire set of input vectors "X" represents the entire input case. For example, suppose the following input vectors come in the order of arrangement.

(0,1,1,1,1) (1,0,1,1,1) (1,1,0,1,1) (1,1,1,0,1) (1, 1, 1, 1, 0) In this example, the input vector entire set X indicating the entire input case is X = ((0,1,1,1,1), (1,0,
(1,1,1), (1,1,0,1,1), (1, l,
It can be written as (1,0,1), (1,1,1,1,0)).

The input vector element "Xi" represents the input vector of the i-th input case. For example, in the above case, the first input case is “(0,1,1,1,1)”,
The second input case is "(1,0,1,1,1)", the third input case is "(1,1,0,1,1)", and so on.

The attribute set "S" represents the entire combination of attributes s.

In this specific example, the keyword attribute set s is a combination of keyword numbers. For example, when there are 1 to 3 attributes s (when there are 3 types of keyword numbers 1 to 3), except for the combinations that are substantially the same, the entire combination S of the keyword attribute set s Is S = ((1), (2),
(3), (1,2), (1,3), (2,3), (1,
2, 3)). However, when the combination of the keyword attribute set s is cut off by a predetermined order, S is the entire combination of attributes within the predetermined order. The order referred to here indicates the number of attributes included in the keyword attribute set s representing one value in S.

That is, if s = (1,3), the degree is “2”, if s = (2), the degree is “1”, and if s = (1,2,3). , The order is “3”.
Therefore, for example, S in the case of cutting off with the order “2” is S = ((1), (2), (3), (1,
2), (1, 3), (2, 3)).

The element "s" represents one value in the attribute set S of the attribute combination. For example, it indicates an attribute surrounded by "()" such as (1, 2, 3) or (3) or (1, 2).

The element “χs (Xi)” is “−1” if the number of the elements corresponding to the keyword attribute set s whose value is “1” is an odd number in the input vector Xi of the i-th case, If an even number (even number includes 0), it represents a function that returns "1" as a coefficient.

For example, the input vector of Xi which is the i-th input case is Xi = (1,0,0), and s = (1,3).
Then, the first value in the input vector Xi is "1", 3
Since the second value is “0”, the number of “1” s is one, which is an odd number, and in this case, “−1”.
Therefore, when expressed as a function χs (Xi), χs (Xi) = χ 1,3 (1,0,0) = − 1. Also, Xi = (1,0,0), s = (2,3)
Then, the second value in Xi is "0", and the third value is "0", so the number of "1" is 0, which is an even number. , 1 is returned, so that the function χs (Xi) is χs (Xi) = χ2,3 (1,0,0) = 1.

These "-1" and "1" are the coefficient α described below.

In other words, if the content of α1 is "-1", the i-th input case Xi = (1,0,
0) and s = (1) indicate that the first keyword in the keyword arrangement order in the input vector Xi has an odd number of registered keywords, and if it is “1”, it is If it is shown that the number is 0 and the content of α2 is "-1", the i-th input case X
i = (1,0,0), s = (2) is the input vector Xi
To the second keyword in the keyword order in
It indicates that the number of occurrences of the registered keyword was an odd number, if it was "1", it was 0, and if the content of α1,2 was "-1", then i The second input case is Xi = (1,0,0), s = (1,
2) indicates that the first and second keywords in the keyword arrangement order in the input vector Xi had an odd number of registered keywords, and if it was “1”, it was 0. If there is an even number, including the case, and the content of α1,2,3 is “-1,”
The i-th input case Xi = (1,0,0), s =
(1, 2, 3) is the first, second, and third keywords in the keyword arrangement order in the input vector Xi,
It indicates that the number of appearances of the registered keyword is an odd number, and if it is “1”, it indicates that there is an even number, including the case where it is 0.

Αs is obtained by arranging the number of appearances of registered keywords for each attribute combination for each input vector in the case of learning as an odd number display and an even number display.

The element "f (Ai)" represents the evaluation for the input case Xi.

[0079]

[Equation 1]

Sign (x): sign if x ≧ 0
If (x) = 1 and x <0, it represents a function such that sign (x) =-1.

Learning based on the spectrum theory is performed by the following mathematical expression (1).

[0082]

[Equation 2]

Find α for all s. m is the total number of cases. Inference based on the spectrum theory is performed by the following mathematical expression (2).

[0084]

[Equation 3]

By using the equation (2) for the given problem case x, the predicted value of f (x) is obtained. As long as the formula (2) is used for inference, since the positive and negative values of the right side are problems, the same result can be obtained without dividing the numerator by m in the formula (1). The equation (3) may be used. In this specific example, the explanation is given using the equation (3).

[0086]

[Equation 4]

By the way, if the above inference / learning is performed by using all the combinations of the attributes s, higher prediction accuracy can be obtained, but the number of attributes (the number of keywords) increases. As a result, the calculation amount increases with the power order. According to the above-mentioned document by "Nathan Linial" and the like, rough learning is completed even if learning is completed at a constant degree, and it is considered that there is not much difference in prediction accuracy.

Therefore, by limiting the order of S used in the learning and inference (prediction) with the same value, it is possible to limit the overall order and reduce the amount of calculation.

The above is the explanation of the inference / learning using the spectrum theory.

Next, the function of the inference unit 5 will be described with reference to FIG. FIG. 3 is a flowchart showing the flow of processing in the inference unit 5.

When the inference unit 5 is activated, the keyword extraction unit 3 reads a list of keywords appearing in the document (step S101). That is, when the document is input, the keyword extracting unit 3 extracts the keywords appearing in the document, forms a list, and holds the list, so the inference unit 5 reads this.

When the keyword list is read,
Next, the keyword number of each of the read keywords is obtained by referring to the keyword table stored in the data storage unit 4 as shown in FIG.
An input vector is generated (step S102).

At this time, a keyword that cannot be found even by referring to the keyword table, that is, a keyword not registered in the keyword table is ignored.

Next, it is checked whether or not there is a combination of the keywords (S described above) (step S103). As a result, if the combination of the keywords (S described above) still exists, then Is generated (step S104), a predicted value is calculated for the generated combination (step S105), and step 103 is performed.
Return to.

If the result of determination in step S103 is that there are no more combinations of keywords,
The calculation of the predicted value ends. Here, the sign of Expression (2)
The value to be assigned to the function, that is, h (x) of the following expression (4) is obtained.

[0096]

[Equation 5]

The inference output (predicted value) of the evaluation with respect to the input document is substituted into sign (x) to obtain h (x).
Is less than “0” (step S107),
The predicted value is "-1", and in this case, the system ends the presentation of the input document that is the evaluation target to the user without executing the evaluation. However, when h (x) is greater than or equal to "0" (step S106), the predicted value is "-1".
At this time, the inference unit 5 instructs the display unit 6 to present the document (step S107) and ends the process. Then, the display unit 6 that has received this presentation instruction reads out the document to be evaluated from the document storage unit 2 and outputs it for display or printing.

The operation of the learning section 8 will be described with reference to FIG. FIG. 4 is a flowchart showing the flow of processing by the learning unit 8. The learning unit 8 is activated, for example, by an input operation from the evaluation data input unit 7 by the user.

When the learning unit 8 is activated, it reads the list of keywords appearing in the document from the keyword extracting unit 3 (step S201). The keyword number of the read keyword is obtained with reference to the keyword table stored in the data storage unit 4 as shown in FIG. 2 to generate a case vector (step S202). At this time, the keywords not registered in the keyword table are ignored.

Next, the learning unit 8 reads the evaluation given by the user by operating the evaluation data input unit 7 (step S
203). This evaluation is a frank judgment evaluation of whether or not the user reads the document instructed to display by the inference unit 5 and is useful or interesting to himself.

When this evaluation is read, it is next checked whether or not there is a combination of keywords (S mentioned above) (step S204). As a result, when there is a combination of keywords (S described above), one combination of keywords is generated (step S205) and a coefficient (α described above) is calculated for the generated combination. (Step S206), the process returns to step S204.
Note that in step S206, f (X) is the evaluation value read in step S203.

Next, in step S204, the presence or absence of a combination of keywords (S described above) is checked again. As a result, if the combination of keywords (S described above) still exists, the next combination is generated (step S205), and the coefficient (α described above) is calculated for the generated combination (step S205). S206) and returns to step S204.

Such processing is repeated as long as there is a combination of keywords, but if the result of determination in step S204 is that there is no more combination of keywords, calculation of the coefficient ends and the calculated coefficient α is obtained. Is stored in the data storage unit 4 (step S207), and the process ends.

In the above operation, the number of keywords registered in the keyword table is "6", and the degree of the coefficient α to be handled is "2".
Up to this point, this specific example will be described more specifically by showing an example of performing filtering.

First, the initial setting of the coefficient α by learning will be described. It is assumed that the registered keywords are as shown in FIG. In the example of FIG. 5, the keyword number 1
"Keyword * 1" as the number, "keyword * 3" as the keyword number 2 and "keywor" as the keyword number 3
d * 3 ”is“ keyword * 4 ”as keyword number 4
But, as keyword number 5, "keyword * 5"
Is registered.

Here, if it is assumed that (keyword * 1, keyword * 3) is a set of required keywords and (keyword * 3, keyword * 4) is a set of unnecessary keywords, then the set of these keywords is Is used to virtually generate a document group.
The generation of a virtual document group is configured such that one document includes three keywords out of the above six keywords, and if different combinations of these are considered as different types of documents, for example, each document is (Keyword * 1, keywor
It is generated as a variation having three keywords in which either d * 3) or (keyword * 3, keyword * 4) is the main subject, and another one keyword is added to this.

As a result, (keyword * 1, keyword * 3, keyword * 3) (keyword * 1, keyword * 3, keyword * 4) (keyword * 1, keyword * 3, keyword * 5) (keyword * 3, keyword * 3, keyword * 4) (keyword * 3, keyword * 4, keyword * 5) (keyword * 3, keyword * 4, keyword * 6) 6 types of virtual documents will be obtained.

All of these virtual documents are documents that include three different combinations of keywords as described above. Among these, a document requiring a virtual document in which a required set of keywords appears Assuming that the virtual document in which the necessary set of keywords appears is an unnecessary document, each is scored.

Furthermore, as noise, "keyword * 5",
There is a virtual document with "keyword * 6". here,
Train the following documents.

"Required Document" (keyword * 1, keyword * 3, keyword * 3) Evaluation value ... + 1 (keyword * 1, keyword * 3, keyword * 4) Evaluation value ... + 1 (keyword * 1, keyword * 3) , Keyword * 5) Evaluation value… + 1 “Unnecessary documents” (keyword * 3, keyword * 3, keyword * 4) Evaluation value… −1 (keyword * 3, keyword * 4, keyword * 5) Evaluation value… − 1 (keyword * 3, keyword * 4, keyword * 6) Evaluation value -1 "Noise" (keyword * 1, keyword * 5, keyword * 6) Evaluation value +1 (keyword * 3, keyword * 5, keyword * 6) Evaluation value ...- 1 As a result of learning the above document group, the value of each α is shown in FIG.
Suppose it becomes like (a). However, in FIG. 6A, αi is the value of α related to keyword * i, and α
i and j mean the value of α related to the combination of keyword * i and keyword * j.

Next, the inference process will be described.

(Keyword * 1, keyword * 3, keyword * 6)
As a keyword, "keyword * 1", "keyword *
3), “...”, “...”, “...”, “keyword * 6” The document which takes the arrangement (When this document extracts the keyword number and makes it an input vector, (1, 1, 0, 0, 0, 1)) is processed up to the degree 2. Then, the following combinations are sequentially generated for each rotation in the processing loop of steps S103 to S105 in FIG.・ ・ Keyword * 1 only, 2nd time ・ ・ ・ keyword * 3 only, 3rd time ・ ・ ・ keyword * 3 only, 4th time ・ ・ ・ keyword * 4 only, 5th time ・ ・ ・ keyword * 5 only, 6th time ・・ ・ Keyword * 6 only, 7th time ・ ・ ・ keyword * 1 and keyword * 3, 8th time ・ ・ ・ keyword * 1 and keyword * 3, 9th time ・ ・ ・ keyword * 1 and keyword * 4, 10th time ・ ・・ Keyword * 1 and keyword * 5, 11th time ・ ・ ・ keyword * 1 and keyword * 6, 12th time ・ ・ ・ keyword * 3 and keyword * 3, 13th time ・ ・ ・ keyword * 3 and keyword * 4, 14th time ... keyword * 3 and keyword * 5, 15th time ... keyword * 3 and keyword * 6, 16th time ... keyword * 3 and keyword * 4, 17th time ... keyword * 3 and keyword * 5, 18th time ... keyword * 3 and keyword * 6, 19th time ... keyword * 4 and keyword * 5, 20th time ... keyword * 4 and keyword * 6, 21st time・ ・ Each combination of keyword * 5 and keyword * 6.

Using this combination, first step S
When passing through 105, since s = 1,
From (a), α1 = −8, and the function χs (Xi)
## EQU1 ## Since .chi.1 (1,1,0,0,0,1) =-1, the following is obtained.

Further, when passing through step S105 for the seventh time, s = 1, 2, and therefore α from FIG. 6 (a).
1,2 = 0, and χ 1,2 (1,1,0,0,0,1) = 1, so that predicted value = predicted value (= 0) + (0 × 1) = 0 Become.

Finally, when passing through step S105, since s = 5, 6, it follows from FIG. 6 (a) that α5,6 =
2 and χ 5,6 (1,1,0,0,0,1) = − 1, so predicted value = predicted value (= 0) + (2 × (−1)) = − 2 Becomes

Here, when these predicted values are summed up, the predicted value finally becomes "12". In this example, since there is no combination, the process proceeds from step S103 of FIG. 3 to step S106, where the final predicted value “1” is obtained.
2 "is compared with a threshold value and the presentation is judged according to the magnitude. When the threshold value is" 0 ", the above-mentioned predicted value" 1 "
Since 2 ″ is larger than the threshold value, it is determined to present it to the user. Then, according to this determination, the output unit 6b displays the input document that is the subject of the evaluation. Information about how much the difference from the threshold value is may be presented to the user.

Next, the learning process will be described. Figure 4
The keywords read in step S201 in step S201 are (keyword * 1, keyword * 3, keyword * 6
). The above "keyword * 1", "keyword * 3",
For a document having an array of "...", "...", "...", "keyword * 6", when the keyword number is extracted in step S202, an input vector of (1, 1, 0, 0, 0, 1) is obtained. can get.

If this document is a document necessary for the user, the evaluation given by the user is "O", "good", "1", or the like. Therefore, the evaluation value read in step S203 is " 1 ”.
The combinations generated in step S205 are similar to those generated in prediction. First, the attribute s in the combination when passing through step S206 is s
= 1 and χ 1 (1,1,0,0,0,1) = − 1, the predicted value is α1 = −8 + (1 × (−1)) = − 9. Further, s in the combination when passing through the step 206 for the seventh time is s = 1,2, and χ1,2 (1,1,0,0,0,1) = 1, so α1,2 = 0 + (1 × 1) = 1.

Finally, s in the combination when passing through step 206 is s = “5,6”, and χ5,6 (1,1,0,0,0,1) = − 1. Therefore, α5,6 = 2 + (1 × (−1)) = 1.

[0120] s = "5,6" to the combination against When the process is finished of in the combination present decision at the next step S204, the other combination is no longer present. Therefore, the process proceeds to step S207, saves each α,
finish.

As a result of such learning, the value of each α is shown in FIG.
It becomes like (b).

The function of each element is as described above.
Therefore, in the present system, if a new document is input from the document input unit 1, the keyword extraction unit 3 extracts the keyword from the document, and the inference unit 5 stores the extracted keyword in the data storage unit 4. Based on the spectrum theory, it was evaluated whether or not the document was worthy of reading by comparing it with the keywords of the keyword table, which is the group of words in the field of interest of the user himself, and was evaluated as worthy of reading. By giving a display instruction to the display unit 6 for a document, the display unit 6 can output the document and present it to the user. Therefore, with this system, from many documents,
It is possible to automatically select documents that are of interest to users.

As described above, in this specific example, the keywords are extracted from the input document, and the extracted keywords are registered keywords (keywords registered in the keyword table, which are keywords of the user's field of interest). It is converted to an input vector that reflects the presence or absence of matching, and inference based on spectrum theory is performed from this input vector, it is determined whether it is worthy of reading, and if it is determined that it is worthy of reading, the document is presented. Moreover, since the result of the user's evaluation of the presented document is learned and reflected in the inference, the estimation accuracy is good, and the inference is based on the spectrum theory. Obtaining an information filter that has features such as a smaller amount of calculation compared to when it is used, and faster determination / learning. It is possible.

In the above-described specific example, the presence or absence of registered keywords is mainly evaluated, but in this case, unregistered keywords are ignored. Therefore, an example of dealing with this will be described as a second specific example.

(Second Specific Example) This specific example can deal with the case where a new keyword not registered in the keyword table of the first specific example appears in the document to be judged.

The information filter of this specific example is constructed as shown in FIG. 7, and is basically the same as that of the first specific example (FIG. 1), and the processing flow of the inference unit 5 is also the first. Since this is the same as the first specific example (FIG. 3), duplicate description here will be omitted, and the points different from the first specific example will be mainly described.

In this specific example, in addition to the function of the learning unit 8 in the first specific example, a learning unit 8a having a function of learning a new keyword is used instead of the previous learning unit 8. The difference is that

Hereinafter, the function of the learning unit 8a that enables learning of a new keyword will be described with reference to FIGS. 7 and 8.

FIG. 8 is a flow chart showing the flow of processing by the learning section 8a.

When the learning section 8a is activated, a list of keywords appearing in the document is read from the keyword extracting section 3 (step S301). Next, the learning unit 8a checks whether or not there is a new keyword not registered in the keyword table in the read keyword group (step S302). That is, the keyword table as shown in FIG. 2 stored in the data storage unit 4 is referred to and it is checked whether or not there is a new keyword that is not in the keyword table. As a result, if a new keyword exists, the new keyword is additionally registered at the end of the keyword table (step S303).
Then, the addition of the new keyword increases the number of elements of S, which is the entire combination of the attributes s, so that necessary αs is added (step S304).

The above processing is repeated for each new keyword in the keyword group extracted by the keyword extracting unit 3. Then, when there are no more new keywords (step S302), the keyword number of the read keyword is obtained from the keyword table and a case vector is generated (step S305).

Next, the evaluation value from the evaluation data input section 7 operated by the user is waited for and read (step S30).
6). When the evaluation value for the presented document is read, next, the sum of all evaluation values up to the present is set to α0, and this is stored as one element of α (step S307).

Next, it is checked whether or not there is a combination of keywords (S described above) (step S30).
8) As a result, if it still exists, the next combination is generated (step S309), the coefficient (α described above) is calculated for the generated combination (step S310), and the process returns to step S308. . Note that f (x) in step S310 is the evaluation value read in step S306.

If the result of determination in step S308 is that there are no more combinations of keywords, the coefficient α is saved in the data storage unit 4 (step S311), and the processing ends.

The method of adding αs in step S304 will be specifically described. For example, as shown in FIG.
Consider a case where there is a keyword table in which individual keywords have already been registered, and a new keyword “database” is newly registered in this table. In this case, the keyword "database" is newly added to the keyword table.
It will be additionally registered as the sixth keyword.

At this point, αs having s = (16) as an element does not exist as data. However, since the word “database” is a newly registered keyword, it can be seen that it has not been present in the filtered text up to now.

That is, the number of appearances is “0” even if it is registered in the keyword table. Therefore, since the learning formula is α (16) = α (16) + f (x) χ (16) (x), χ (16) from the start of learning to the present is always “1” (even number of appearances; It should have returned 0 even numbers), and α (16) is Σf (x).

Therefore, it is given by α (16) = α0. That is, the sum of all evaluation values up to now is set to α0, and this is saved as one element of α, and for newly registered “keyword * 16”, α0 should be used as α (16). become.

Further, α (a, 16) is the same as α (a) because the keyword of keyword number 16 (that is, the phrase “database”) does not appear. Here, a is a keyword number from 1 to 15.

Similarly, α of degree N is α of degree (N-1).
Can be obtained from

As described above, even when a new keyword not registered in the keyword table appears, the sum of all evaluation values up to the present is set to α0, which is stored as one element of α and newly registered. For the keyword, the information filter that can cope with new registration can be obtained by using α0 as α of the keyword and diverting from other keywords of degree 1 as α of the other degree.

A modification of the second specific example will be described.

In the second specific example, the inference unit 5 part can be realized by the same processing contents as in the first specific example (FIG. 3). However, the processing flow of FIG.
You may do like this. That is, among the keywords extracted from the document, the number of types of keywords corresponding to the keyword (registered keyword) registered in the keyword table is R, the number of types of all the extracted keywords is A, and the values “0” to “”. When the threshold value set to a desired value between 1 ”is CO, in step S108, R / A <CO
In case of, the instruction is given to display the document without performing the inference in steps S103 to S105.

In this way, the document is unconditionally presented according to the ratio of the number of types of keywords corresponding to the registered keywords included in the keywords extracted from the document to be evaluated and the number of types of keywords extracted from the document. When processing is possible and the number of registered keywords in the number of types of extracted keywords is less than the set value, by presenting that document, it prevents the occurrence of harmful effects such as oversight of documents in new fields and documents of new technology Can be achieved.

Of course, of the extracted keywords that do not exist in the keyword table, as described in the second specific example (FIG. 8), they are registered in the keyword table, the necessary α is added, and α is set to It is also possible to learn and reflect it in subsequent inference.

(Third Concrete Example) A third concrete example will be described. In this example, the inference and learning based on the above-mentioned spectrum theory are speeded up respectively. The spectrum theory described below is hereinafter referred to as "fast spectrum theory".

The configuration of the information filter according to this example is shown in FIG.
As shown in FIG. 0, the inference unit 5 is replaced with the inference unit 5A and the learning unit 8 is replaced with the learning unit 8A, but the functions of other elements including these are basically the first specific example (FIG. 1). The description of the same parts is omitted here, and the points different from the first specific example will be mainly described.

As a characteristic of the problem vector (that is, the input vector to be evaluated generated from the document to be evaluated) input to the information filter, most of the attribute values are "0". That is, the number of keywords that appear in one document is usually 20 to 30, and it has been known by experiments that many of the words and phrases that make up the document are not registered keywords.

On the other hand, the total number of attributes, that is, the number of keywords registered in the information filter reaches several thousand to ten thousand. Focusing on this, most of χs (x) is “1” in the formula (2) which is the calculation formula of the predicted value.
It can be seen that the value is.

Therefore, the predicted value f (0) when all the attribute values are "0" is obtained in advance, and from here,
The amount of calculation can be reduced by correcting only the part that is affected by the attribute value that is not "0".

For example, in a system in which 1000 keywords are registered, when a document including 10 keywords is filtered up to degree 2, in the method of the first specific example, 1000 + 1000 × 1000 times αsχs.
It can be seen that if (x) is required to be calculated in this specific example, 10 + 10 × 990 calculations will be required, and the amount of calculation will be significantly reduced. Therefore, the amount of calculation is reduced, resulting in high-speed processing.

In order to realize this, the inference unit 5 is replaced with the inference unit 5A as follows in this specific example.

The function of the inference unit 5A will be described below with reference to FIG.

FIG. 11 is a flow chart showing the flow of processing by the inference unit 5A. When the inference unit 5A is activated, the keyword extraction unit 3 reads a list of keywords appearing in the document (step S401). The keyword number of the read keyword is obtained by referring to the keyword table stored in the data storage unit 4 as shown in FIG. 2 (step S402). At that time, if the target keyword for which the keyword number is to be obtained is not registered in the keyword table, it is ignored.

Next, f (0) is first substituted into the predicted value (step S403). Then, it is checked whether or not there is another combination of keywords (step S4).
04) As a result, if there is still a combination of the keywords (s described above), an odd number is selected from the keywords that have appeared in the document, and the remaining keywords are selected to generate a combination (step). S405). The resulting combination is f (0) and f (x)
Is a different part, so that this is corrected by performing an operation of predicted value = predicted value-2α0-2αs (step S40).
6).

Here, α0 is the total sum of the evaluation values up to the present, and αs stores the difference from the total sum unlike the first specific example.

When the correction process in step S406 is completed, the process returns to step S404. Then, if the result of determination in step S404 is that a combination of keywords still exists, the above processing is repeated, but if it does not exist anymore, the calculation of the predicted value ends, and the document is presented to the user according to the result of the prediction. It is determined whether or not to present (step S407). This determination is determined by whether or not the prediction result is “less than 0”.

That is, when the result of the prediction is "less than 0", the inference unit 5A does not issue a presentation instruction, so the system ends without presenting the document to the user. However, when the result of the prediction is “0 or more” (step S407), the inference unit 5A issues a presentation instruction to the display unit 6 (step S408) and ends the process.

In response to this instruction, the display unit 6 displays the document storage unit 2
The document targeted for the evaluation is read from the output unit 6b
Will be output and the document will be presented.

The operation of the learning section 8A will be described with reference to FIG. FIG. 12 is a flowchart showing the flow of processing by the learning unit 8A. When the learning unit 8A is activated,
A list of keywords appearing in the document is read from the keyword extracting unit 3 (step S501). The keyword number of the read keyword is obtained with reference to the keyword table stored in the data storage unit 4 as shown in FIG. 2 (step S502). At this time, if the keyword is not registered in the keyword table, it is ignored.

Next, an evaluation value is added to α0 (step S5
05), and it is checked whether or not there is a combination of keywords (step S506). As a result of checking in this step S506, a combination of keywords (S described above)
If is still present, an odd number is selected from the keywords that have appeared in the document, and the rest are selected from the registered keywords that have not appeared in the document, and a combination is generated (step S
507). The resulting combination is then χs
Since (x) is a portion of "-1", αs is corrected (step S508).

Further, f (0) is also corrected accordingly (f (0) = f (0) -2f (X)) (step S50).
9). Then, the process returns to step S506 to check whether there is a combination of keywords. As a result, if there is a further combination, the processing from step S507 onward is repeated. If the result of step S506 is that there is no combination, calculation of the coefficient ends, and α is stored in the data storage unit 4 (step S510). , F (0) are also stored in the data storage unit 4 (step S511), and the process ends.

As described above, in the present specific example, the keywords extracted from the document are collated with the registered keywords, and if the keywords are the registered keywords, the single keyword of the keyword and the combination of the keywords according to the order are examined, and the predicted values of the keywords are calculated. Then, the presentation or non-presentation of the document is decided based on the predicted value.

As a result, since the processing contents are simplified, it is possible to speed up the inference and learning based on the spectrum theory in this specific example.

(Fourth Concrete Example) A fourth concrete example will be described. The above-described first to third specific examples are all examples of a system configuration of sequential processing in which inference / display is performed for one document. In the fourth specific example, an example will be described in which a plurality of documents are collectively inferred and displayed.

The configuration of the information filter according to this example is shown in FIG.
As shown in Fig. 3, for multiple documents,
The reasoning unit 5B that is capable of reasoning is used. The reasoning unit 5B and other functional elements are basically the same as those in the above-described specific example (FIG. 1). Description is omitted, and different parts will be described.

In this specific example, the inference unit 5B has a content in which the inference processing has been slightly modified from that described with reference to FIGS. Here, as the function of the inference unit 5B, FIG.
FIG. 14 shows an example that is realized by slightly modifying the above-described one.

In this specific example, only the inference process is first executed for each of the plurality of documents, and then the documents with the highest evaluation are displayed in order. In other words, this specific example is an example in which a plurality of documents are evaluated and then the one with the highest need is selected and presented.

Therefore, the documents to be evaluated are input for a plurality of documents, and the keyword extracting unit 3 extracts the keywords appearing in each document for each document.

When the inference unit B5 is activated, a list of keywords appearing in the first document is read from the keyword extracting unit 3 (step S101). When the list of keywords is read, the keyword number of each of the read next keywords is stored in the data storage unit 4.
The input vector is generated by referring to the keyword table as shown in FIG.
102). At this time, if a keyword that cannot be found by reference, that is, a keyword that is not registered in the keyword table, is ignored.

Next, it is checked whether or not there is a combination of keywords (S above) (step S103). As a result, if a combination of keywords (S above) still exists, the following A combination is generated (step S104), a predicted value is calculated for the generated combination (step S105), and the process returns to step 103.

If the result of determination in step S103 is that there are no more combinations of keywords, the calculation of the predicted value ends. Here, the value to be assigned to the sign function of Expression (2) and h (x) of Expression (4) are obtained.

Then, the process proceeds to step S111, and the inference is collectively performed for each document by the processing loop of step S111 and thereafter. Here, for the items determined to be displayed by the formula (2), h of the formula (4)
(X) is stored as an index indicating the degree of necessity of the document.

Such processing is performed on each of the second document, the third document, etc., and for those judged to be displayed by the formula (2), h (x) of the formula (4) is It is saved as an index showing the degree of necessity of the document.

After such processing is completed, step S11
Moving to 2, the documents determined to be displayed are sorted in descending order of the index h (x) indicating the degree of necessity. Then, the process proceeds to step S113, the display unit 6 is instructed to display the target documents in the sorted order, and the documents are displayed in the sorted order.

As a result, each of a plurality of documents is first evaluated, and the plurality of documents are stored with an index indicating the degree of necessity of the document. It can be displayed by giving a display instruction.

Therefore, when a plurality of documents are input with high frequency, it is possible to collect and evaluate the documents in advance, and then select and present the ones with high necessity. Regardless of each terminal user in the network where many mails jump in, regardless of whether selection is necessary, those with high urgency, high importance, or those that are evaluated as highly interesting , It becomes convenient to be able to read in order from the highest index.

(Modification 1 of First to Fourth Concrete Examples) First to First
In the fourth specific example, the sign function for obtaining the inference result is sign (x): sign (x) = if x ≧ 0.
If 1 and x <0, the function was such that sign (x) =-1. Now consider the following function.

Sign ′ (x): if x ≧ c, then sign (x) = 1; if x <c, then sign (x) = − 1 In this function, the reasoning threshold c = 0 is The above is sign (x). Here, the threshold value c of the inference may be set arbitrarily. The display condition becomes more severe as the value of c increases toward the positive side, and the display condition becomes looser as the value of c increases toward the negative side.

(Variation 2 of First to Fourth Concrete Examples) As another example of displaying a document determined to be important, the inference unit 5 should present the document to the user here. When it is determined that the contents are stored in the document storage unit 2, the display unit 6 first reads out the table of contents or the summary of the document and controls the display unit 6b to display the contents. When the user who sees it decides to read the text and gives the instruction from the input operation unit (not shown), the display unit 6 reads the text of the document from the document in the document storage unit 2 in response to the display request. The output unit 6b is provided with a function for controlling the display.

In the system having such a configuration, the inference unit 5
When it is determined that the document should be presented to the user, the display unit 6 first reads out the table of contents or summary of the document in the document storage unit 2 and outputs (displays) the document. Control to do.

If the user who sees the output contents determines that he / she wants to read the text and gives the instruction from the input operation unit (not shown), the display unit 6 displays the document storage unit 2 in response to the display request. The body of the document is read from the document and output (display).

When displaying a document determined to be important in this way, first, a table of contents or a summary of the document is displayed, and when the user decides that he wants to read the text, he or she responds to the display request from the user. By displaying the text of the document, the user can select and read only the more interesting document from the documents selected by the information filter for the user.

In the inference, the first step is to use a set of information (document name, document code, etc.) for specifying the document and the keyword group input from the document input unit 1 of FIG. Inference may be performed with respect to the second stage, and only in the second stage, the body of the document may be input to extract the keyword and the inference may be performed.

In the above specific examples, although the keyword itself is not mentioned, if the document is classified in advance in addition to the words that often represent a normal document (for example, IPC in patent documents). For classification (international patent classification, etc.), this field, author name, author affiliation, etc. can also be used as part of keywords. Further, the present invention is not limited to the above-mentioned specific examples, and various modifications can be carried out without departing from the scope of the invention.

As described above, it is possible to obtain an information filter device that collectively infers and displays a plurality of documents.

(Fifth Specific Example) In the above specific example, the number of keywords has been increasing, but the interest of the user often changes with the passage of time, and in that case, it is not necessary. Keywords occur. Then, if the unnecessary keywords are left as they are, not only the processing will be wasted, but also an error will occur in the evaluation of the document required by the user. Therefore, since the removal of unnecessary keywords is important, a method of removing the unnecessary keywords will be described as a fifth specific example.

Here, unnecessary keyword detection is performed, and an unnecessary keyword detecting section 9 having a keyword deleting function for deleting the unnecessary keyword from the past learning result of the relation when the unnecessary keyword is detected is provided. When the unnecessary keyword detection unit 9 detects an unnecessary keyword, the unnecessary keyword in the data storage unit 4 is deleted from the past learning result of the relationship in the data storage unit 4.

That is, in the fifth specific example, FIG.
As shown in FIG. 10, the unnecessary keyword detection processing unit 9 having the unnecessary keyword detection function and the function of updating the contents of the keyword table is further provided in any of the configurations of FIGS. 01, 07, 10 and 13 to detect the unnecessary keywords. Then, based on this, the contents of the keyword table in the data storage unit 4 are updated.

In order to enable this updating, the keyword table shall be as shown in FIG. 16 in which a list for entering the registration date and time and the number of times of use is added to the keyword table of FIG. , 8A has a function of entering the registration time at the time of registration in the keyword table and updating and recording the cumulative number of times of use of the corresponding keyword for each use of the keyword. Although it is a little different from the specific example, the other points follow the operations of the above specific examples, so the description of the operations of the same parts will be omitted and the following description will be made only for the different parts.

The feature of this specific example resides in the function of the unnecessary keyword detection processing section 9. Among the functions of the unnecessary keyword detection processing section 9, the unnecessary keyword detection function is
It is activated by an instruction from the user, at a certain time interval, or when the filtering speed becomes slower than a certain value.

When the unnecessary keyword detection processing unit 9 is activated, it refers to the keyword table in FIG. 16 to search for keywords that have been registered for a certain period of time and are not used often. For the search, FIG. 16 has a configuration in which a list for entering the registration time and the number of times of use is added to the keyword table of FIG.

In this example, number 1 (keyword number is 1)
The word "word processor" is registered on "November 10, 1994", and it indicates that it has been used "53" times, and the second is "dictionary" (keyword number is 2).
The word is registered on "November 11, 1994" and indicates that it has been used "21" times.
The number 3 (keyword number 3) "induction" was registered on "December 10, 1994" and indicates that it was used "9" times, and the number 4 (keyword number 4 The term “learning” in) is registered on “December 10, 1994” and indicates that the number of times of use is “6”, and so on.

The general operation of the apparatus of the present specific example configured as described above will be described with reference to FIG. FIG. 17 is a flowchart showing the operation of the fourth specific example.

It is confirmed whether the keyword to be deleted still exists, and if it exists, the next keyword is obtained (step 601). If a certain number of days have not passed since the obtained keyword was registered (step 602), the process returns to step 601 to search for the next keyword. If a certain number of days have passed since the registration (step 602) and the number of times the keyword has been used is a constant B or more, the process returns to step 601 without deleting (step 603). If the number of times the keyword is used is less than the constant B (step 603), and if the number of times the keyword is used is less than the constant C that is smaller than the constant B (step 604), the keyword is deleted. (Step 606). If the number of times the keyword has been used is equal to or greater than the constant C (step 604), the keyword is deleted (step 606). If the number of times the keyword is used is the constant C or more (step 6)
04), if the absolute value of the primary value α of the learning coefficient of the keyword is equal to or less than the constant D (step 605), the keyword is deleted (step 606). When the absolute value of the primary value α of the learning coefficient of the keyword is not equal to or less than the constant D (step 605), the keyword is not deleted and the next keyword is searched (step 601).

The keyword deletion process will be described below using a specific example.

When registering a keyword, the data learning unit 8 records the registration date and time, and each time it is used, the number of times the keyword in the keyword table is used is increased (incremented). Then, the unnecessary keyword detecting unit 9 refers to the registration time and the number of times of use of the keyword table, and deletes the items that have been used for a certain period of time since the initial registration and whose use frequency is less than the certain number of times, and delete the corresponding items. .

For example, if it was registered three months or more ago,
A process of deleting a keyword that has been used 10 times or less up to now is performed. Therefore, if today's date is June 1, 1995, in this case, in the keyword table of FIG. 16, keywords 1 to 7 (keywords from "word processor" to "optical") are periodically changed. It becomes a candidate for deletion.

Next, looking at the number of times of use, the keyword of keyword number 3 (“induction”) and the keyword number 4
The keyword ("learning") and the keyword of keyword number 7 ("optical") are to be deleted. Here, if a keyword whose number of uses is 5 or less is deleted, in the case of the example of FIG. 16, the keyword of keyword number 7 (“optical”) is deleted. It should be noted that a keyword which is used more than 6 times and less than 10 times is determined with reference to the primary value of the coefficient α, whether or not to delete.

FIG. 18 shows an example of the coefficient α. The primary value of the coefficient α is directly expressed by each keyword.
It is an indicator of how much it contributes to usefulness. Therefore, when the primary value of α is close to “0”, it means that it does not contribute much and can be deleted. When the threshold value of α is “10”, αs ≦ 10 is a deletion target.

When α (3) corresponding to the keyword of keyword number 3 and α (4) corresponding to the keyword of keyword number 4 are α (3) = 2 and α (4) = 100, , Keyword number 3 (“induction”)
Is determined as the deletion target, but the keyword of keyword number 4 is not set as the deletion target.

When the keyword to be deleted is determined, α
Remove the item related to the keyword to be deleted from the value of. That is, when deleting the keywords of keyword number 3 and keyword number 7, “α3”, “α1,”
3 ”,“ α2,3 ”,“ α3,4 ”,“ α3,5 ”...
... "α7", "α1,7", "α2,7", "α3"
7 ”,“ α4, 7 ”... Are deleted because both of them are elements in which either the keyword of keyword number 3 or the keyword of keyword number 7 is involved.

Next, the keyword of keyword number 3 and the keyword of keyword number 7 are deleted from the keyword table. Next, the keyword number 3 and the keyword number 7 are deleted from the keyword table, and since the table is in the worm-eaten state, the numbers are re-arranged from 1 in order.

Then, the number of times of use of all the keywords currently registered in the keyword table is reduced at a constant rate. For example, here, it is halved.

As described above, in this specific example, the time of initial registration in the keyword table and the frequency of use are managed for each keyword, and an unnecessary keyword detection processing unit is provided to set a fixed period from the initial registration. Has passed,
In addition, since we have made it possible to delete keywords that are rarely used, it is possible to delete unnecessary keywords in a way that reflects the learning result, and it is possible to always manage the keywords that target the user's latest interest. Then, it becomes possible to accurately select the document of interest. Note that the specific numerical values used here, such as threshold values and coefficients, are changed according to the required characteristics of the system, and the numerical values themselves have no meaning.

(Sixth Concrete Example) When the range of interest of the user handled by one information filter device is wide,
There is a risk that the number of required keywords will increase explosively.
Then, as the number of keywords increases, the amount of calculation increases exponentially, so it is advantageous to divide the keywords and perform filtering by theme with a plurality of information filters. In such a case, a method of dividing and using the keyword will be described.

In this specific example, in addition to the configuration of the fifth specific example, the keyword division processing unit 1 for further dividing keywords is used.
The configuration as shown in FIG. The keyword division processing unit 10 has a function of dividing a past learning result of the relationship when the number of predetermined keywords exceeds a certain size.

The keyword division processing section 10 is activated by an instruction from the user or when the filtering speed falls below a certain level. When the keyword division processing unit 10 is activated, it starts keyword classification with reference to the secondary α value.

The value of the secondary α (for example, α (1,2)) is two keywords (α (1,2) indicated by the value,
It shows how much each of the keywords of keyword number 1 and the keyword of keyword number 2 contributes to the effectiveness of the entire document. Unless two keywords appear in the same document, the value is "0". is there.

Therefore, if the value of α (1,2) is close to “0”, the keyword of keyword number 1 and the keyword of keyword number 2 did not appear at the same time, or
It means that it does not contribute much to the judgment of the validity of the document. When a certain threshold value, for example, 5 is set as the threshold value, α
If (i, j) ≧ 5, it is considered that the two keywords i and j (keywords of the keyword numbers i and j) have a link (connection).

[0211] Therefore, when a line is drawn between two keywords having a link, for example, as shown in Fig. 1
The keyword of 8 is expressed as shown in FIG.

In such a case, the keyword of keyword number 1, the keyword of keyword number 2 and the keyword of keyword number 5 are set as one group, and the keyword of keyword number 3, the keyword of keyword number 4 and the keyword of keyword number 6 are separated. The keywords are divided by a method of dividing them into groups.

The above operation will be specifically described with reference to the flowchart shown in FIG.

When the keyword division processing unit 10 is activated, it is determined whether or not division is still required (step 70).
1). The determination is made based on whether the ratio of the sizes of the divided keyword groups is below a certain level. Since the division is not performed immediately after being activated, the size is Mi0 and the division is still necessary. A starting point is randomly selected from the keywords (step 702). Check if there is a keyword that can be reached by following the link from the selected starting point.
The link between two keywords (a, b) uses the absolute value of two-dimensional α (ab). If the absolute value is larger than a predetermined value, it is determined that there is a link (step 703). If there is still a keyword that can be traced, this keyword is marked (step 704).
When there are no more links to follow (step 703), the marked keywords are taken out from the old keyword list. Returning to step 701, it is checked whether further division is necessary.

More specifically, the procedure of division processing will be described. First, starting with the keyword with keyword number 1, search for linked keywords in order. For example, keyword number 1
Since the keyword No. and the keyword No. 2 have a link, the keyword No. 2 belongs to the same group. The keywords of keyword numbers 3, 4, 5, and 6 do not have a direct link with the keyword of keyword number 1.

Next, when a keyword having a link with the keyword of the keyword number 2 newly entered in the same group as the keyword of the keyword number 1 is searched for, the keyword of the keyword number 5 is found. Therefore, keyword number 5
Put the keywords in the same group.

Next, a keyword having a link with the keyword of the keyword number 5 is searched for, but since there is no new keyword to be added to the same group, the processing ends here. Here, as in the case of this example, it is good if two sets with the same number of keywords are created, but if the number of keywords included in the extracted group is less than a certain ratio, Randomly select a keyword that is not included in this group, search for a new group starting from that keyword, and add it to the already extracted group.

In this way, the process is repeated until a fixed rate is reached. If the keywords are divided into two groups, the keyword table and the value of α included in the data storage unit 6 are copied, and the keywords not included in each group are deleted by the deletion method described in the sixth specific example. To do. In this way, the learning result can be divided into two. By this keyword division, it is possible to suppress an increase in the number of keywords and reduce the calculation amount.

(Seventh Concrete Example) In this concrete example, the inference unit 5 is used.
A method of improving the data learning unit 8 to obtain a high-order α with a small storage capacity will be described.

The nth-order value of α is valid only when n keywords appear at the same time. However, of the thousands of keywords handled by the information filter, it is extremely unlikely that n keywords will appear at the same time. In particular, the higher the value of n, the more the combination expands, but the actual effective α is small.

Therefore, only valid α is stored, and the other ones are obtained from lower α.

For example, α (a1, a2, a3, ...,
an-1, an), the keywords "a1", "a2", "a3", ..., "An-1", appearing here are obtained.
If all of the "an" have never appeared at the same time, α (a1, a2, a3, ..., An-1, an) is represented by α up to the n-1th degree. It is known.

For example, when obtaining a cubic α, α (a1
, A2, a3), the keywords "a1", "a2",
Assuming that "a3" never appeared at the same time, α (a1, a2, a3) is χ in the first specific example.
It is the same value as counting the number of times that an odd number of keywords among a1, a2, and a3 appear at the same time.

Therefore, if each value of a1, a2, and a3 is represented by a vector having a value of 1/0 ((1,1,1) when all keywords appear), "α ( a1
, A2, a3) "is" (1,0,0) "or" (0,1,0) "or" (0,0,1) "or" (1,1,1) " You are counting the number of times.

On the other hand, "α (a1, a2)" is "(1,
0,0) ”,“ (0,1,0) ”,“ (1,0,
1) ”and“ (0,1,1) ”are counted as“ α
(A1, a3) "is" (1,0,0) "," (1,
"1,0)", "(0,0,1)", "(0,1,1)"
The number of occurrences of "α (a2, a3)" is "(1,
1,0) ”,“ (0,1,0) ”,“ (1,0,
1) ”and“ (0,0,1) ”are counted as“ α
(A1) "is" (1,0,0) "," (1,0,
1) ”,“ (1,1,0) ”,“ (1,1,1) ”is the number of times α (a2) is“ (0,1,0) ”,
"(0,1,1)", "(1,1,0)", "(1,
1, 1) "and the number of occurrences of" α (a3) "
Is "(0,0,1)", "(1,0,1)",
The numbers of occurrences of "(1,1,0)" and "(1,1,1)" are counted.

Therefore, α (a1) + α (a2) + α (a3) -α (a1, a2
) -Α (a2, a3) -α (a1-a3) = g (0,
0,1) + g (0,1,0) + g (1,0,0) + g
(1,1,1) -4g (1,1,1) = α (a1, a2
, A3) -4g (1,1,1) = α (a1, a2, a3
However, g (a, b, c) (a, b, c is 1 or 0) is a function, and the function g (a, b, c) here is (a, b
represents the number of times b, c) has occurred.

That is, when g (1,1,1) = 0, it can be seen that α (a1, a2, a3) can be expressed by primary and secondary α. Using this mechanism, the combination of keywords and the value of α are held only when three keywords appear at the same time, and in other cases, α is obtained by the above method.

In this way, it is not necessary to store an enormous number of values of α, and it is possible to perform highly accurate prediction with a small storage capacity.

The present invention is not only applied as a system for performing information filter processing and a method for information filter processing, but also provided as a program package and provided as an application package to a computer system such as a personal computer or a workstation. Information filtering processing can be performed, and an embodiment in which a distribution form as a computer program package is adopted can be realized.

[0230]

As described above in detail, according to the present invention,
Since the inference / learning is performed only based on the relationship between the combination of keywords extracted from the document and the evaluation value regarding the importance of the document, the calculation amount is reduced while maintaining good estimation accuracy.
It is possible to obtain an information filter that can perform determination / learning at high speed.

Therefore, according to the present invention, the user does not need to read a document that is obviously not of interest.

[Brief description of drawings]

FIG. 1 is a diagram for explaining the present invention and is a block diagram showing a configuration of an information filter according to first to third specific examples of the present invention.

FIG. 2 is a diagram for explaining the present invention, showing an example of a keyword table used in the present invention.

FIG. 3 is a diagram for explaining the present invention and is a flowchart showing a flow of inference in the first and second specific examples of the present invention.

FIG. 4 is a diagram for explaining the present invention and is a flowchart showing the flow of learning in the first specific example of the present invention.

FIG. 5 is a diagram for explaining the present invention, showing another example of the keyword table used in the present invention.

FIG. 6 is a diagram for explaining the present invention, showing the values before and after learning of each coefficient α used in the present invention.

FIG. 7 is a diagram for explaining the present invention and is a block diagram showing a configuration example of an information filter in a second specific example of the present invention.

FIG. 8 is a diagram for explaining the present invention and is a flowchart showing the flow of learning in the second example of the present invention.

FIG. 9 is a diagram for explaining the present invention and is a flowchart showing a flow of inference in a modified example of the second specific example of the present invention.

FIG. 10 is a diagram for explaining the present invention and is a block diagram showing a configuration of an information filter in a third example of the present invention.

FIG. 11 is a diagram for explaining the present invention and is a flowchart showing a flow of inference in the third example of the present invention.

FIG. 12 is a diagram for explaining the present invention and is a flowchart showing the flow of learning in the third example of the present invention.

FIG. 13 is a diagram for explaining the present invention and is a block diagram showing a configuration of an information filter according to a fourth embodiment of the present invention.

FIG. 14 is a diagram for explaining the present invention and is a flowchart showing the flow of inference in the fourth example of the present invention.

FIG. 15 is a diagram for explaining the present invention, which is a block diagram showing a configuration of an information filter according to a fifth example of the present invention.

FIG. 16 is a diagram for explaining the present invention, showing an example of a keyword table used in a fifth example of the present invention.

FIG. 17 is a diagram for explaining the present invention and is a flowchart showing the flow of a deletion process in the fifth example of the present invention.

FIG. 18 is a diagram for explaining the present invention, showing an example of each coefficient α used in the fifth specific example of the present invention.

FIG. 19 is a diagram for explaining the present invention and a diagram for explaining a sixth specific example of the present invention.

FIG. 20 is a diagram for explaining the present invention and is a block diagram for explaining a sixth specific example of the present invention.

FIG. 21 is a diagram for explaining the present invention and is a flowchart showing the flow of a division process according to the sixth example of the present invention.

[Explanation of symbols]

1 ... Document input section 2 ... Document storage 3 ... Keyword extractor 4 ... Data storage unit 5, 5A, 5B ... Inference section 6 ... Display 6a ... Control function unit 6b ... Output section 7 ... Evaluation data input section 8, 8a, 8A ... Learning section 9 ... Unnecessary keyword detection processing unit.

Continuation of the front page (56) References JP-A-5-204975 (JP, A) Nobuhiro Shimogori Outside, concept of user model: Collection of user information by information filter, 49th IPSJ (1994) (Latter term) National Convention (5), September 28, 1994, pp. 5-103 to 5-104 Chie Morita, Development of inference tool KINO, Research Report of Information Processing Society of Japan (95-AI-98), 1995 January 18, Vol. 95, No. 4, pages 29-38 Peter W. Foltz, Sus an T .; Dumais, Personalyzed and Analyzes, Communication of the ACM, December 1992, Vol. 35, No. 12, P.I. 51-60 Masahiro Morita, Current Status and Prospects of Information Filtering Technology, IEICE Technical Report (AI93-24), July 22, 1993, Vol. 93, No. 153, pp. 49-56 Sanjiv K. Bhatia, User Profiles for Information Retrieval, Right Notes in Artificial Intelligence, USA, October 1991 Paul E. et al. Blacace, Personal Information Intake Filtering, Bellcore Information on Filtering Works hop, October 1991, URL, http: // www. bacrace. net / Resources / ifilte rl. html (58) Fields investigated (Int.Cl. ⁷ , DB name) G06F 17/30 JISST file (JOIS) WPI (DIALOG)

Claims (9)

(57) [Claims] 1. A keyword extraction for extracting a plurality of keywords
Output means and one or more keywords of the plurality of keywords
Generate a combination of keywords that have the words
In the document,
It is determined whether a keyword appears, and if it appears,
Is the first identifier, and if not, the second identifier
The first identifier assigned to the combination of the keywords
Is counted, and if the counting result is an even number, the third identifier is
If it is an odd number, the fourth identifier is combined with the keyword.
To generate a spectrum, and
The key based on the evaluation value associated with the document
Calculate a coefficient for a combination of words,
Learning by storing in association with keyword combinations
Of the learning means to learn and the document to be inferred, the combination of the keywords is
Based on the coefficient and the spectrum for a certain
Inference means that infers the importance of the document to be inferred.
Information filter instrumentation, characterized by comprising; and a step
Place 2. The learning means, when a new keyword other than the keyword extracted by the keyword extracting means is extracted, adds the new keyword to the past learning result by the learning means. The information filter device according to claim 1 , further comprising keyword adding means. 3. The learning means learns only a range in which a keyword extracted from the document influences at the time of the learning, and the inference means selects an inference target.
To the extraction means to extract keywords from the document to be
When the importance of the document is held in advance when the extracted keyword does not exist and the importance of the input document is inferred, the keyword extracted from the document is extracted by the extracting means. If there is a stored keyword, a value for varying the importance held by the keyword is obtained, and the stored importance is corrected based on this value, thereby The information filter device according to claim 1 , wherein the information filter device is for determining the degree of importance. 4. When the ratio of the extracted keywords in all the keywords extracted from the document to all the extracted keywords is less than a predetermined value, the inference is performed by the inference means. claim 1 also characterized by being configured to be displayed to the user without the
Is an information filter device according to any one of 2 . 5. The learning means has a function of detecting an unnecessary keyword in the learning result and deleting the unnecessary keyword, and when the unnecessary keyword is detected, a past keyword is detected.
Claims, characterized in that further comprising an unnecessary keyword detection means for deleting the unnecessary keywords from the learning result
Item 1. The information filter device according to item 1 . When 6. The number of keywords extracted by the extraction unit is equal to or greater than a predetermined size, wherein, characterized in that it comprises a keyword dividing means for dividing the learning result of the past of the relationship further Item 1. The information filter device according to item 1 . 7. A plurality of keywords are extracted, and one or more keywords of the plurality of keywords are extracted.
Generate a combination of keywords that contains the words and include the keyword combination in the document to be learned.
It is determined whether each keyword
If the first identifier is not present, the second identifier is
An identifier is assigned to the combination of the keywords, the first identifier is counted, and when the counting result is an even number,
The third identifier, and if odd, the fourth identifier,
A spectrum is generated by assigning it to a combination of words, and the spectrum and the evaluation value associated with the document.
Calculate the coefficient for the combination of the keywords based on
Then, associate the above-mentioned coefficient with the combination of the above-mentioned keywords.
Of the documents that are learned by remembering and inferred, the combination of the keywords is
Based on the coefficient and the spectrum for a certain
To infer the importance of the document to be inferred.
Characterized information filtering method. 8. The learning is extracted by the extracting means.
And when a new keyword other than the keyword is extracted, information filtering method according to claim 7, further comprising the step of adding the new keywords. Wherein said learning comprises learning about extracted keyword affects the range of the document, the
The inference is performed by preliminarily holding the importance of the document when the keyword extracted from the document to be inferred does not include the keyword extracted by the extracting unit ,
When inferring the importance of the input document, the keyword extracted from the document is extracted by the extracting unit.
When there is a stored keyword, a value that changes the degree of importance held by the keyword is obtained, and the stored degree of importance is corrected based on this value 8. The information filtering method according to claim 7 , further comprising a process of obtaining a degree.