JP2001028006A - Method and device for automatic information filtering - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method and device for automatic information filtering which properly decide whether or not information is improper by easily and accurately setting weight of a word and using the weight of the word. SOLUTION: Improper information which should be stopped from being provided and proper information which need not be stopped from being provided are inputted as learning data to a weighted word list learning part 60, weight of a word is obtained from a linear discrimination function for discriminating between the improper information and proper information in a vector space and stored as a weighted word list in a weighted word list storage part 50, and a word extraction part 3 extracts a word from information from an input part 1; and the weight(w) of this word is obtained from the weighted word list storage part 50 and inputted to an automatic filtering part 30, and the total of weights(w) of those words is calculated, so that the improper information when the total is larger than a threshold or the information when the total is smaller is decided.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention extracts words appearing in information from various types of information including mail provided via the Internet, and determines whether the information is inappropriate based on the words. The present invention relates to an information automatic filtering method and apparatus for judging whether or not to provide inappropriate information.

[0002]

2. Description of the Related Art With the rapid spread of the Internet,
Computers, which were only tools of a limited number of experts, have begun to be introduced to ordinary homes and schools. This has made it possible for many ordinary people who had never even touched a computer to easily access the Internet. Against this background, children's access to harmful information such as pornographic images flooding the Internet has become a serious problem in recent years. To address this issue, the United States proposed a law called the Telecommunications Quality Act, which allowed government agencies to censor information on the Internet, but as a result of a trial, it violated the constitution that guarantees freedom of expression He was ruled and was unable to legislate.

[0003] Recently, a technique called "information filtering" has attracted attention. Information filtering is a technique in which when a user accesses information on the Internet, the harmfulness of the information is checked, and if the information is determined to be harmful, access to the information is blocked by some means.

The methods adopted by currently available harmful information filtering software are roughly classified into the following four.

(1) Filtering based on self-determination (2) Filtering based on determination by a third party (3) Automatic filtering (4) Method using score (score) for words Here, these four methods will be briefly described. First, in the filtering method based on self-determination, a provider of WWW information determines the harmfulness of its own content, and describes the result in an HTML file. The filtering software refers to the described result, and blocks access if it is determined to be harmful. FIG. 4 shows filtering by this method.

[0006] In the filtering based on the self-determination shown in FIG. 4, the World Wide W
PICS (Platform for Int) created by eb Consortium
It uses a criterion for describing Internet content ratings called ernet Content Selection. By using PICS, a content provider can easily depict and disclose information provided by the content provider.

In many cases, when a content provider publishes such an evaluation result, the service of an evaluation organization that outputs the evaluation result by PICS is used. On behalf of such evaluation agencies, Recreational Software Adviso
There are organizations such as the ry Council (RSAC) and SafeSurf, which provide evaluation results based on their own set of standards. The content provider describes the evaluation results from these institutions in the header of the HTML file. FIG. 5 shows a description example of this evaluation result.

[0008] At present, the self-determination is left to the independence of the content provider. Therefore, it can be said that effective harmful information filtering by this method is impossible unless many content providers have a will to receive this determination.

Next, the filtering based on the judgment by the third party will be described. Among companies that create harmful information filtering software, a method of independently determining the harmfulness of a homepage on the WWW and using the result as a criterion for the filtering software is adopted. Generally, the URL of the harmful homepage as a result of this evaluation
The list is built. This list of URLs is distributed to the user together with the filtering software, and serves as a criterion for the filtering software. In many cases, filtering software is designed to periodically download the harmful URL list. FIG. 6 shows a mechanism of harmful information filtering based on a judgment by a third party.

A typical example of software having such a mechanism is CyberPatrol. CyberPatr
ol for 13 genres such as "violence" and "sex"
Each has a harmful URL list, and performs harmful information filtering according to these systems.

The list of harmful URLs used in this method is accessed by a software company on a homepage,
Since it is created and expanded by making a judgment, it is possible to use a different URL from a newly established homepage or a conventional URL.
It is impossible to deal with the homepage moved to the RL. Therefore, at present, it is not possible to cope with filtering for pages that are not evaluated.

Next, automatic filtering will be described. Some harmful information filtering software checks the contents of the accessed homepage to determine harmfulness.

Specifically, harmful information, that is, words that would be included in inappropriate information are registered in advance,
In this method, it is checked whether or not the registered word appears in the information, and when the registered word is included, provision of the information is prevented. For example, when the provision of pornographic information is blocked, if the information includes a character string such as "sex" or "xxx", the provision of the information is blocked. As an application of this method, there is also a method of preventing the provision of information when the ratio of registered words included in the information exceeds a predetermined threshold.

Next, a method of using a score (score) for a word will be described. In this method, a word that is likely to be included in inappropriate information and a score for the word are registered in advance, and whether or not the registered word appears in the information is checked. If the sum is greater than a predetermined threshold, the information is prevented from being provided.

[0015]

The main purpose of automatic information filtering is to increase the rate at which inappropriate information is blocked and reduce the rate at which appropriate information is erroneously blocked. There are advantages and disadvantages, and there is a problem that conventional information automatic filtering cannot provide sufficient filtering performance.

Specifically, in the conventional automatic filtering method, for example, a bad example has been reported in which, for example, a homepage about a British town called "Susex" is blocked. Further, in the conventional method using the score for a word, the setting of the word and the score of the word is ad hoc, and there is no guide as to how the setting is most effective for the user. Therefore, there is a problem in terms of performance such that information that should be prevented from being provided cannot be prevented or information that should not be provided should be prevented.

For example, it is assumed that the word "high school girl" generally appears frequently in pornographic information, and the word "high school girl" and its score are registered as 40. As a result, since the expression “sample image of a high school girl, free” includes “high school girl”, the score of the entire expression is 40. Similarly, for the expression "the bus on which a high school girl got on accident in Hokkaido", the total score of this expression is 40, and the scores of these expressions are the same. For this reason, if the threshold value is set to 20, there is a problem that the latter expression that should not be blocked is blocked, and if the threshold value is set to 50, the former expression that should be blocked is not blocked. There's a problem. To distinguish between these two expressions, you need to use “sample”, “image”
Words such as "free" and words such as "bus", "Hokkaido" and "accident" will also need to be scored, but these words are commonly used words. It is not clear how to set it, and there is a problem that the performance greatly fluctuates due to the setting of the score, and it is not possible to obtain sufficient performance for determining whether the expression is inappropriate.

The present invention has been made in view of the above,
It is an object of the present invention to provide an information automatic filtering method and apparatus for easily and accurately setting the weight of a word and accurately determining whether information is inappropriate by using the weight of the word. Is to do.

[0019]

In order to achieve the above object, the present invention according to claim 1 identifies inappropriate information among information provided via the Internet, and provides the identified inappropriate information. An automatic information filtering method for preventing the word included in the information by automatic learning with learning data of inappropriate information that needs to be provided and appropriate information that does not need to be provided is provided as learning data. Determine the weight of the word used to determine whether it is necessary to prevent the provision of information, and store and manage the weight of the obtained word as a weighted word list corresponding to each word, Information provided via the Internet is input, words included in the information are extracted, and the weight for each of the extracted words is read out from the weighted word list. Calculating the sum of the weights of each word that is summarized in that to determine whether to prevent the providing of the information based on the calculated sum.

According to the first aspect of the present invention, the weight of a word is obtained by automatic learning using, as learning data, inappropriate information that needs to be prevented from being provided and appropriate information that does not need to be prevented from being provided. The weights of the obtained words are stored and managed as a weighted word list corresponding to each word, the words included in the information provided via the Internet are extracted, and the weight for each of the extracted words is calculated. Read from the weighted word list, calculate the sum of the weights of the read words, and determine whether or not to prevent the provision of the information based on the sum. The weight of the word is determined accurately by automatic learning, and whether or not the information is inappropriate information is accurately determined by using the weight of the word accurately determined, and provision of inappropriate information is prevented. It can be.

The present invention described in claim 2 provides the present invention in claim 1.
In the invention described in the above, the processing for obtaining the weight of the word is performed by:
The gist of the present invention is to determine the weight of a word by automatic learning based on a linear discriminant function that can discriminate an inappropriate document from an appropriate document in a vector space.

According to the second aspect of the present invention, the weight of a word is determined by automatic learning based on a linear discriminant function that can discriminate an inappropriate document from an appropriate document in a vector space. Can be set appropriately.

Further, the present invention according to claim 3 is an information automatic filtering device for identifying inappropriate information among information provided via the Internet, and preventing provision of the identified inappropriate information, Whether it is necessary to prevent the provision of information for words contained in the document by automatic learning using learning data of an inappropriate document that needs to be provided and an appropriate document that does not need to be provided. Word weight learning means for calculating the weight of a word used to determine whether the weight of the word is determined, weighted word list storage means for storing and managing the weight of the determined word as a weighted word list corresponding to each word, Input means for inputting information provided via the PC, word extracting means for extracting words included in the input information, and weighting for each of the extracted words. From the weighted word list, calculate the sum of the weights of the read words, and determine whether to prevent provision of the information based on the calculated sum. I do.

According to the third aspect of the present invention, the weight of a word is obtained by automatic learning using, as learning data, an inappropriate document that needs to be prevented from being provided and an appropriate document that does not need to be prevented from being provided. The weights of the obtained words are stored and managed as a weighted word list corresponding to each word, the words included in the information provided via the Internet are extracted, and the weight for each of the extracted words is calculated. Read from the weighted word list, calculate the sum of the weights of the read words, and determine whether or not to prevent the provision of the information based on the sum. The weight of the word is determined accurately by automatic learning, and whether or not the information is inappropriate information is accurately determined by using the weight of the word accurately determined, and provision of inappropriate information is prevented. It can be.

According to a fourth aspect of the present invention, in the third aspect of the present invention, the word weight learning means is based on a linear identification function capable of discriminating the inappropriate document from the appropriate document in a vector space. The gist of the present invention is to have means for calculating the weight of a word by automatic learning.

According to the fourth aspect of the present invention, the weight of a word is determined by automatic learning based on a linear discriminant function that can discriminate an inappropriate document from an appropriate document in a vector space. Can be set appropriately.

[0027]

FIG. 1 is a block diagram of an automatic information filtering apparatus according to an embodiment of the present invention. The information automatic filtering device shown in the figure obtains the weight of a word by automatic learning, determines whether the information is inappropriate by using the weight of the word obtained by the automatic learning, and An input unit 1 for inputting HTML information provided via the Internet, a word extraction unit 3 appearing in information input via the input unit 1, and a need to prevent provision Whether or not it is necessary to prevent the provision of information for words contained in the document by automatically learning the document which is inappropriate information and the document which is appropriate information that does not need to be prevented from being provided as learning data. Weighted word list learning unit 60 for calculating the weight of the word used for determining the word, the weight of the word obtained by the weighted word list learning unit 60 is used as a weighted word list corresponding to each word. The weighted word list storage unit 50 to be managed and input from the input unit 1 based on the words extracted by the word extraction unit 3 and the weights w of the words obtained from the weighted word list storage unit 50 for the words. It comprises an automatic filtering unit 30 for determining whether or not provision of information should be prevented, and an output unit 40 for outputting a determination result obtained by the automatic filtering unit 30.

The automatic information filtering apparatus according to the present embodiment is characterized in that the weights of words are obtained in advance by automatic learning in the weighted word list learning section 60, and the weights of the words obtained by the automatic learning are used. This method of automatically learning the weight of a word is performed by a word weight learning algorithm shown in the flowchart of FIG. That is,
In the learning algorithm shown in FIG.
= {D 1,..., Dn}, the inappropriate information that needs to be prevented from being provided and the appropriate information that does not need to be provided are input to the weighted word list learning unit 60. The word weight is obtained from a linear discriminant function that discriminates information and appropriate information in a vector space. Specifically, the following procedure is performed.

First, an HTML document input from the input unit 1 is represented by a vector space model. That is,
The n words representing all documents are selected, and each document is represented by an n-dimensional vector as in the following expression.

[0030]

(Equation 1) Each element of this vector is obtained by normalizing the frequency of occurrence of each word in the document d. To normalize the appearance frequency of words, a technique called TF * IDF expressed by the following equation is used.

[0031]

(Equation 2) Here, tfdi is the frequency at which word i appears in document d, N is the number of all documents, and dfi is the number of documents in which word i appears.

The automatic filtering is performed by a linear discriminant function represented by the following equation, and the function is used to calculate the sum of the word weights Dis (d).

[0033]

(Equation 3) Here, wi is the weight for each word i, fdi is the value of the above equation (2), and is the fdi value of each word in the document.

From the above equation (3), the sum Dis (d)
Is greater than 0, the document is deemed harmful; if it is less than 0, it is determined to be harmless.

The weight for each word i described above is set so that when the document d is harmful, the total sum Dis (d)> 0, and when it is harmless, the total sum Dis (d) ≦ 0.

Next, the word weight learning algorithm will be described with reference to the flowchart shown in FIG. The learning of the weight of this word is perceptron lear
ning algorithm (PLA) is used.

In FIG. 2, first, various parameters are set (step S51). This parameter includes
A set of weights of each word W = (w1,..., Wn), N pieces of learning data E = {d1,..., Dn}, a constant η, a maximum number of learnings Max, and a number of learnings for repeating the learning processing shown in FIG. m.

Then, among the words representing all the documents, n frequently-selected words are selected (step S5).
2).

Next, a set W of word weights is initialized (step S53). In this initialization, a random number is input as the weight of each word. Then, the sum Dis (d) of the word weights is calculated for all the learning data by the above equation (3) (step S55).

Then, as a result of this calculation, it is checked whether or not the total sum Dis (d) ≦ 0 for all harmless documents d and the total sum Dis (d)> 0 for all harmful documents d. (Step S57) If so, the process ends. If not, the weight change is performed on all the documents d classified in this way as shown in the following steps S61 and S63. The degree S is corrected (step S59).

That is, in step S61, the document di
Is harmful and the total sum Dis (d) ≦ 0, the weight change degree S is corrected so as to increase, and in step S63, the document di is harmless and the sum D
If is (d)> 0, correction is made so as to reduce the weight change degree S.

Then, using the weight change degree S thus corrected, the word weight set W is corrected as in the equation shown in step S65. Then, the learning number m is incremented by +1 (step S67), and it is checked whether the learning number m is smaller than the maximum learning number Max (step S69). If the learning number m is smaller than the maximum learning number Max, the process proceeds to step S55. Returning, the processing from step S55 is repeated until the condition shown in step S57 is satisfied. Then, a set of word weights for the n words is finally obtained.

The weight of each word obtained by the weighted word list learning unit 60 is stored in the weighted word list storage unit 50 as a weighted word list corresponding to each word.
Table 7 shown below is a table showing a weighted word list stored in the weighted word list storage unit 50, and a word weight w is stored for each word.

[0044]

[Table 1] Next, based on the word weights obtained in the weighted word list learning unit 60 and stored in the weighted word list storage unit 50, it is determined whether or not the information provided from the Internet is inappropriate information. Will be described.

In FIG. 1, the information from the Internet input from the input unit 1 is collated by the word extraction unit 3 with the word list stored in the weighted word list storage unit 50, and appears in the input information. Find words and their frequency of appearance. Further, the weight w of the word that appears at the same time is also obtained from the weighted word list storage unit 50, and the appearing word and its frequency and weight are supplied to the automatic filtering unit 30. The automatic filtering unit 30 calculates the sum of the weights w for all the words that appear in the input information from the weight w and the appearance frequency for the input word, compares the sum with a predetermined threshold, and determines that the sum is the threshold. If the sum is larger than the threshold value, the information is determined to be inappropriate. If the sum is smaller than the threshold value, the information is determined to be appropriate information.

A specific description will be given. As shown in Table 1, in the weighted word list learning unit 60, the weight of “image” is 10.9, the weight of “sample” is 18.7, and the weight of “accident” is − 16.6, the weight of "high school girl" is 82.2, and the weight of "bus" is -101.
9, the weight of "Hokkaido" is determined to be -112.5, and the weight of "free" is determined to be -6.3, which are stored in the weighted word list storage unit 50. For the entire expression "the bus on which the bus rides in Hokkaido", the total sum of the weights of the words is calculated by the automatic filtering unit 30, and 82.2-101.9-112.5-1.
6.6 = −148.8. In addition, for the entire expression of “sample image of high school girl, free”, the sum total of each word is calculated by the automatic filtering unit 30, and 82.2 + 18.
7 + 10.9-6.3 = 105.5. If the threshold value is set to 0 in the same manner as in the processing of FIG. 2, the expression “the bus on which a high school girl got on accident in Hokkaido” is below the threshold value. Since the expression "free" exceeds the threshold, it can be correctly determined that the provision of information is blocked.

Next, an automatic filtering device according to another embodiment of the present invention will be described with reference to FIGS. The automatic filtering device shown in FIG.
And a harmful UR used for referring to a harmful URL in the third-party determination filtering unit 23 for the information automatic filtering device 25 that creates a word list by learning described in
An L list table storage unit 17 is added.

The harmful URL list table storage unit 17 stores URLs that provide harmful information as a harmful URL list table, and the third party judgment filtering processing unit 23 receives the URL from the input unit 1. The URL of the HTML document is collated with each URL registered in the harmful URL list table of the harmful URL list table storage unit 17 to determine whether there is a matching URL.

FIG. 6 is a block diagram showing a more detailed configuration of the automatic filtering device shown in FIG. The automatic filtering device shown in FIG. 6 includes an input unit 1, a word extraction unit 3, a weighted word list storage unit 50, and an automatic information filtering device that use the weighted word list created by learning shown in FIG. Filtering unit 30,
In addition to the output unit 40, a filtering unit 15 based on a URL list and a harmful URL list table storage unit 17 corresponding to the third party determination filtering processing unit 23 in FIG.

In the filtering process performed by the automatic filtering device configured as described above, that is, the filtering process performed by the information filtering device using the URL list list and the weighted word list created by learning by the third party judgment filtering processing unit, the Internet 21 is first used. The URL of the HTML document input via the URL is collated with each URL registered in the harmful URL list table of the harmful URL list table storage unit 17 to determine whether there is a matching URL. Then, the URL registered in the harmful URL list table of the harmful URL list table storage unit 17
If the URL matches, the presentation of the information indicated by the URL is blocked.

Filtering unit 1 based on URL list
5 as a result of referring to the harmful URL list table, if there is no URL that matches the URL registered in the harmful URL list table in the harmful URL list table storage unit 17, the weighted word list created by learning is used. The filtering by the used information automatic filtering device 25 is performed as described in FIG.

As described above, in the present embodiment, both filtering based on judgment by a third party and filtering using a weighted word list created by learning are performed, so that harmful information is accurately detected and prevented. be able to.

[0053]

As described above, according to the present invention,
The weight of a word is determined by automatic learning using inappropriate information that needs to be prevented from providing and inappropriate information that does not need to be provided as learning data, and the weight of this word is assigned to each word. Store and manage as a list, extract words included in information provided via the Internet, read the weight for each of the extracted words from the weighted word list, calculate the sum of the weights of each word,
Based on this sum, it is determined whether the provision of information should be stopped or not. Therefore, the weight of words that had to be set as ad hoc in the past is accurately obtained by automatic learning, and the weight of the accurately obtained words is calculated. It is possible to accurately determine with high performance whether or not the information is inappropriate information by using the information, thereby preventing the provision of the inappropriate information.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of an information automatic filtering device according to another embodiment of the present invention.

FIG. 2 is a flowchart showing a procedure for setting word weights used in the flowchart shown in FIG. 1;

FIG. 3 is an explanatory diagram showing a schematic configuration of an automatic filtering device according to another embodiment of the present invention.

FIG. 4 is a diagram for explaining conventional filtering based on self-determination.

5 is a diagram illustrating an example of description of an evaluation result by RSACi and SafeSurf as an example of filtering based on self-determination illustrated in FIG. 4;

FIG. 6 is a diagram for describing conventional harmful information filtering based on determination by a third party.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Input part 3 Word extraction part 30 Automatic filtering part 50 Weighted word list storage part 60 Weighted word list learning part

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] August 18, 1999 (1999.8.1)
8)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0027

[Correction method] Change

[Correction contents]

[0027]

FIG. 1 is a block diagram of an automatic information filtering apparatus according to an embodiment of the present invention. The information automatic filtering device shown in the figure obtains the weight of a word by automatic learning, determines whether the information is inappropriate by using the weight of the word obtained by the automatic learning, and An input unit 1 for inputting HTML information provided via the Internet, a word extracting unit 3 for extracting words appearing in the information input via the input unit 1, and a blocking of the provision. It is necessary to prevent the provision of information for the words included in the document by automatically learning a document that is inappropriate information that requires the information and a document that is appropriate information that does not need to be prevented from being provided as learning data. A weighted word list learning unit 60 for determining the weight of a word used to determine whether or not there is a word; A weighted word list storage unit 50 stored and managed as a list, a word extracted by the word extraction unit 3 and an input from the input unit 1 based on the word weight w obtained from the weighted word list storage unit 50 for the word The automatic filtering unit 30 determines whether or not to provide the provided information, and an output unit 40 that outputs the determination result obtained by the automatic filtering unit 30.

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0047

[Correction method] Change

[Correction contents]

Next, an automatic information filtering apparatus according to another embodiment of the present invention will be described with reference to FIG. The automatic information filtering apparatus shown in FIG. 3 uses a third-party judgment filtering processing unit 23 and a harmful URL in the third-party judgment filtering processing unit 23 for the information automatic filtering apparatus that creates a word list by learning described in FIG. The difference is that a harmful URL list table storage unit 17 used to refer to the URL is added.

[Procedure amendment 3]

[Document name to be amended] Statement

[Correction target item name] 0048

[Correction method] Change

[Correction contents]

In FIG. 3, the harmful URL list table storage unit 17 stores URLs that provide harmful information as a harmful URL list table, and the third party determination filtering processing unit 23 shown in FIG. The URL of the HTML document input via the input unit 1 is collated with each URL registered in the harmful URL list table of the harmful URL list table storage unit 17 to determine whether there is a matching URL. is there.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0049

[Correction method] Change

[Correction contents]

That is, the automatic information filtering apparatus shown in FIG. 3 has an input unit 1, a word extraction unit 3, a weighted word list storage unit 50, and an automatic filtering unit 30 shown in FIG.
3 and a function of performing filtering using a weighted word list created by learning, the third-party determination filtering processing unit 23 shown in FIG.
And a harmful URL list table storage unit 17
It has a function of performing filtering by L.

[Procedure amendment 5]

[Document name to be amended] Statement

[Correction target item name] 0050

[Correction method] Change

[Correction contents]

The thus configured automatic information filtering apparatus, that is, the third-party judgment filtering processing unit 2
In the automatic information filtering process using the harmful URL list stored in the harmful URL list table storage unit 17 and the weighted word list created by learning, first, the HTML document input via the Internet 21 has its URL. Harmful U in harmful URL list table storage unit 17
It is checked against each URL registered in the RL list table, and it is determined whether there is a matching URL. When the URL matches the URL registered in the harmful URL list table of the harmful URL list table storage unit 17, presentation of the information indicated by the URL is prevented.

[Procedure amendment 6]

[Document name to be amended] Statement

[Correction target item name] 0051

[Correction method] Change

[Correction contents]

Filtering unit 1 based on URL list
5 as a result of referring to the harmful URL list table, if there is no URL that matches the URL registered in the harmful URL list table in the harmful URL list table storage unit 17, the weighted word list created by learning is used. Filtering is performed by the used information automatic filtering device 25.

[Procedure amendment 7]

[Document name to be amended] Drawing

[Correction target item name] Fig. 1

[Correction method] Change

[Correction contents]

FIG.

 ──────────────────────────────────────────────────続 き Continuing from the front page (72) Inventor Kazuo Hashimoto 2-1-15 Ohara, Kamifukuoka-shi, Saitama F-term in K.D. Laboratory (reference) 5B075 KK07 KK13 KK33 KK54 KK70 ND03 NR02 NR12 NR12 QM10 UU40

Claims (4)

[Claims] 1. An automatic information filtering method for identifying inappropriate information among information provided via the Internet and preventing provision of the identified inappropriate information, the method comprising: Is used to determine whether it is necessary to prevent the provision of information for words included in the information by automatic learning using appropriate information and appropriate information that does not need to be prevented from being provided as learning data. The weight of the word is obtained, and the weight of the obtained word is stored and managed as a weighted word list corresponding to each word, and information provided through the Internet is inputted, and the words included in this information are extracted. Then, a weight for each of the extracted words is read from the weighted word list, a sum of the weights of the read words is calculated, and a weight is calculated based on the calculated sum. Automatic information filtering method characterized by determining whether to block the provision of information. 2. The processing for obtaining the weight of a word, wherein the weight of the word is obtained by automatic learning based on a linear discriminant function that can discriminate the inappropriate information and the appropriate information in a vector space. The information automatic filtering method according to claim 1. 3. An automatic information filtering apparatus for identifying inappropriate information out of information provided via the Internet and preventing provision of the identified inappropriate information, wherein the information is inappropriate. Is used to determine whether it is necessary to prevent the provision of information for words included in the information by automatic learning using appropriate information and appropriate information that does not need to be prevented from being provided as learning data. Word weight learning means for obtaining word weights, weighted word list storage means for storing and managing the obtained word weights as weighted word lists corresponding to the respective words, and inputting information provided via the Internet Inputting means, word extracting means for extracting words included in the input information, and a weighted word list for weighting each of the extracted words. Et read, calculates the sum of the weights of each word thus read out, information automatic filtering device, characterized in that it comprises a determination means for determining whether to block the provision of the information based on the calculated sum. 4. The method according to claim 1, wherein the word weight learning unit includes a unit that obtains a word weight by automatic learning based on a linear discriminant function that can discriminate the inappropriate information and the appropriate information in a vector space. The information automatic filtering device according to claim 3.