JP5113206B2 - Spam blog determination apparatus and method - Google Patents

Description

  The present invention relates to a spam blog determination apparatus and method.

  Conventionally, there is a blog called “spam blog” that is automatically created and posted for the purpose of guiding to a specific site. Spam blogs are also referred to as “splogs”. This spam blog has various risks in the Internet. For example, for a blog service provider, a spam blog places a high load on a server and a line, so there is a risk that the service will be hindered. In addition, there is a need for companies and the like to know the search rank of their products and information. However, there is a risk that spam blogs can corrupt the search rank. Then, there is a risk that the information desired by the user cannot be reached from the search service due to the spam blog, for example.

  As a countermeasure against such a spam blog, for example, a method of determining a spam blog based on information such as frequent terms of the spam blog and a spam blog pattern is considered (for example, Patent Document 1 (paragraph [0005]). reference). What is described in the corresponding paragraph of Patent Document 1 is a so-called spam filter that uses machine learning to exclude those having specific characteristics.

JP 2006-331297 A

  By the way, some spam blogs include popular terms in the blog. And terms that are popular are characterized by the fact that they are rapidly becoming obsolete, are replaced one after another, and it is difficult to predict how long a term that has been once popular will continue. Even the technique disclosed in Patent Document 1 cannot appropriately cope with such a situation, and an administrator needs to change the program logic each time to determine a spam blog. It was a complicated task for the administrator of the service provider.

  An object of the present invention is to provide a spam blog determination apparatus and method for determining a spam blog by facilitating an operation by an administrator.

  The present inventors have found that an environment in which spam blogs can be determined by registering popular terms and determining spam blogs following the trend duration is completed. It came to do. Specifically, the present invention provides the following.

(1) In response to receiving registration designation of a predetermined keyword, predetermined keyword storage control means for storing the predetermined keyword in a predetermined keyword storage means, and in response to receiving a blog article to be determined, Machine learning means for determining whether or not the blog article is a spam blog using the predetermined keyword stored in the predetermined keyword storage means as a feature, and the blog article to be determined by the machine learning means Among them, a spam determination result output means for outputting a blog article including the predetermined keyword stored in the predetermined keyword storage means and a result of the determination by the machine learning as to whether or not it is a spam blog, In response to receiving the deletion designation of the predetermined keyword, the predetermined keyword storage means stores the predetermined keyword. Information indicating the correctness / incorrectness of the determination result by the machine learning for the blog article from which the determination result by the machine learning is output by the adjustment determination return means for deleting the predetermined keyword and the spam determination result output means In response to accepting certain correctness determination information, the correctness determination information storage means for storing the correctness determination information in association with the blog article, and the correctness determination information stored in the correctness determination information storage means for a predetermined period. An error rate of determination by the machine learning based on each predetermined keyword based on the first trend end determination means for determining that the trend has ended for the predetermined keyword for which the error rate is equal to or greater than a predetermined threshold; A spam blog determination device comprising:

According to such a configuration of the present invention, the spam blog determination apparatus receives, for example, a predetermined keyword registration specification for processing a predetermined keyword that is out of fashion and includes a predetermined keyword. And output the result. Therefore, the spam blog determination apparatus can determine a spam blog including a predetermined keyword only by allowing the administrator to specify and register a predetermined keyword. Further, the spam blog determination device removes the predetermined keyword from the processing target by receiving a deletion designation for excluding the predetermined keyword from the administrator who viewed the determination result output by the spam determination result output unit. . Therefore, the spam blog determination device can appropriately determine a spam blog that includes a predetermined keyword that has become obsolete. Further, according to the present invention, the spam blog determination device accepts correctness / incorrectness determination information, which is information indicating the correctness / incorrectness determination of the result of the machine learning determination for the blog post that outputs the determination result, and associates it with the blog post. An error rate of determination by machine learning is calculated for each predetermined keyword based on the correctness / incorrectness determination information stored in a predetermined period, and when the error rate is equal to or higher than a predetermined threshold, the prevalence of the predetermined keyword is increased. Judge that it is over. Therefore, the administrator who sees the determination result output by the spam determination result output means makes a correct / incorrect determination on the result of the machine learning determination on the blog article and inputs the correct / incorrect determination information, so that the spam blog determination apparatus is predetermined. The error rate of determination by machine learning is calculated for each keyword, and the “temporary” fashion end state of a predetermined keyword can be automatically determined by comparing with a preset threshold value.

( 2 ) In response to receiving registration designation of a predetermined keyword, predetermined keyword storage control means for storing the predetermined keyword in a predetermined keyword storage means, and in response to receiving a determination target blog article, Machine learning means for determining whether or not the blog article is a spam blog using the predetermined keyword stored in the predetermined keyword storage means as a feature, and the blog article to be determined by the machine learning means Among them, a spam determination result output means for outputting a blog article including the predetermined keyword stored in the predetermined keyword storage means and a result of the determination by the machine learning as to whether or not it is a spam blog, Stored in the predetermined keyword storage means in response to receiving the deletion specification of the predetermined keyword And adjusting the return means for deleting the predetermined keyword, among the posts of the determination target containing the predetermined keyword in a predetermined period, the proportion of the posts of the machine learning unit is determined to be spam blog There spam blog determination device provided in the case of equal to or less than a predetermined threshold value, and a second outbreak end determining unit determines that the epidemic has been completed for the predetermined keyword.

  According to such a configuration of the present invention, the spam blog determination device has a predetermined ratio of blog articles that the machine learning means determines to be spam blogs among the blog articles to be determined that include a predetermined keyword in a predetermined period. When the threshold is less than or equal to the threshold, it is determined that the trend has ended for the predetermined keyword. Therefore, the spam blog determination apparatus can determine the “temporary” trend end state of a predetermined keyword according to a predetermined criterion.

(3) the adjusting the return means, the predetermined keyword epidemic is determined to have ended by the first outbreak termination judgment means or the second outbreak end determining unit is deleted from the predetermined keyword storage unit, (1 ) Or the spam blog determination device according to ( 2 ).

  According to such a configuration of the present invention, the spam blog determination device deletes the predetermined keyword determined to be end of the fashion from the predetermined keyword storage unit, so that the predetermined keyword storage unit is automatically maintained. be able to.

( 4 ) A predetermined keyword storage step of storing the predetermined keyword in the predetermined keyword storage means in response to the computer receiving registration designation of the predetermined keyword, and the computer has received the blog article to be determined And a machine learning step for determining whether or not the blog article is a spam blog by using the predetermined keyword stored in the predetermined keyword storage means as a feature, and the computer is a spam blog. Corresponding between the blog article including the predetermined keyword stored in the predetermined keyword storage unit and the result of the determination by the machine learning whether the blog article is a spam blog among the blog articles to be determined whether or not and spam determination result output step of attaching output, the computer, the predetermined key In response to reception of the deletion specified word, the a predetermined keyword adjusted return step to remove the predetermined keyword stored in the storage unit, the computer, the determination by the machine learning in the spam analysis result output step results In response to accepting correct / incorrect determination information, which is information indicating the correct / incorrect determination of the determination result by the machine learning, the correct / incorrect determination that stores the correct / incorrect determination information in association with the blog article. An information storage step; and a computer calculates an error rate of determination by the machine learning for each of the predetermined keywords based on the correctness / incorrectness determination information stored in the correctness / incorrectness determination information storage step in a predetermined period, and the error rate is predetermined First to determine that the trend has ended for the predetermined keyword that is equal to or greater than the threshold of Spam blog determination method, including, and end-of-line determination step.
(5) A predetermined keyword storage step of storing the predetermined keyword in the predetermined keyword storage means in response to the computer receiving registration designation of the predetermined keyword, and the computer has received the blog article to be determined And a machine learning step for determining whether or not the blog article is a spam blog by using the predetermined keyword stored in the predetermined keyword storage means as a feature, and the computer is a spam blog. Corresponding between the blog article including the predetermined keyword stored in the predetermined keyword storage unit and the result of the determination by the machine learning whether the blog article is a spam blog among the blog articles to be determined whether or not And a spam determination result output step for outputting, and the computer outputs the predetermined key An adjustment return step of deleting the predetermined keyword stored in the predetermined keyword storage means in response to accepting a word deletion designation; and a computer including the predetermined keyword in the predetermined period The second trend end determination that determines that the trend has ended for the predetermined keyword when the ratio of the blog articles determined to be spam blogs in the machine learning step is equal to or less than a predetermined threshold. And a spam blog determination method including steps.

  ADVANTAGE OF THE INVENTION According to this invention, the spam blog determination apparatus and method which determine a spam blog easily by the operation | work by an administrator can be provided.

It is a figure which shows the whole structure of the spam blog determination system which concerns on 1st Embodiment, and the function structure of a spam blog determination apparatus. It is a figure which shows the example of the various data memorize | stored in the memory | storage part of the spam blog determination apparatus which concerns on 1st Embodiment. It is a figure for demonstrating the determination result by the machine learning of the spam blog determination apparatus which concerns on 1st Embodiment. It is a flowchart of the predetermined keyword reflection process of the spam blog determination apparatus according to the first embodiment. It is a flowchart of the spam determination process of the spam blog determination apparatus according to the first embodiment. It is a figure which shows the whole structure of the spam blog determination system which concerns on 2nd Embodiment, and the function structure of a spam blog determination apparatus. It is a flowchart of the predetermined keyword deletion process of the spam blog determination apparatus according to the second embodiment. It is a figure which shows the whole structure of the spam blog determination system which concerns on 3rd Embodiment, and the function structure of a spam blog determination apparatus. It is a flowchart of the predetermined keyword deletion process of the spam blog determination apparatus according to the third embodiment.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. This is merely an example, and the technical scope of the present invention is not limited to this.

(First embodiment)
[Overall Configuration of Spam Blog Determination System 100 and Functional Configuration of Spam Blog Determination Device 1]
FIG. 1 is a diagram showing an overall configuration of a spam blog determination system 100 and a functional configuration of a spam blog determination apparatus 1 according to the first embodiment. FIG. 2 is a diagram illustrating examples of various data stored in the storage unit 20 of the spam blog determination apparatus 1 according to the first embodiment. FIG. 3 is a diagram for explaining a determination result by machine learning of the spam blog determination apparatus 1 according to the first embodiment.

  As shown in FIG. 1, the spam blog determination system 100 includes a spam blog determination device 1, a blog server 3, a management terminal 5, and a communication network 9.

  The spam blog determination device 1 stores a predetermined keyword (predetermined keyword) received from the management terminal 5 and uses the stored predetermined keyword as a feature to determine whether the blog article received from the blog server 3 is a spam blog. Is determined by machine learning, and the determination result is output to the management terminal 5. Further, the spam blog determination device 1 is a device that removes the predetermined keyword received from the management terminal 5 from the stored keyword, thereby removing it from the feature used in the determination process by machine learning as to whether it is a spam blog. . The spam blog determination device 1 includes a control unit 10 and a storage unit 20.

  The control unit 10 includes a predetermined keyword receiving unit 11, a predetermined keyword storage control unit 12, a blog article receiving unit 13, a machine learning unit 14, a spam determination result output unit 15, and an adjustment return unit 17.

  The predetermined keyword receiving means 11 is a control unit that receives a predetermined keyword registration request transmitted from the management terminal 5. Predetermined keywords are keywords that are out of fashion and tend to be subject to spam blogs. Keywords that are out of fashion are in contrast to “permanent” and cannot be predicted for a “temporary” duration. A spam blog is, for example, a meaningless blog article that uses a word that is generally used as a topic, and by placing a link to a certain web page, the ranking of that web page is raised. The one used in Therefore, the predetermined keyword accepting unit 11 transmits the word that is commonly used in spam blogs and is a popular topic and is out of fashion as a predetermined keyword. Accept.

  The predetermined keyword storage control unit 12 is a control unit that stores the predetermined keyword received from the management terminal 5 in a predetermined keyword DB 21 (DB: database) (predetermined keyword storage unit).

  The predetermined keyword DB 21 shown as an example in FIG. 2A is a DB that stores the predetermined keyword received from the management terminal 5. The predetermined keyword DB 21 includes items of a serial number 21a, a predetermined keyword 21b, a registration date 21c, and a type 21d.

  The serial number 21 a stores the serial numbers from 1 assigned by the control unit 10 in the order received from the management terminal 5. The predetermined keyword 21 b stores a predetermined keyword received from the management terminal 5. The registration date 21 c stores the date when the predetermined keyword is received from the management terminal 5. Type 21d stores a category of a predetermined keyword. The type stored in the type 21d may be received from the management terminal 5 together with a predetermined keyword.

  Returning to FIG. 1, the blog article receiving means 13 is a control unit that receives a blog article from the blog server 3. Here, the blog article refers to each article constituting the blog. The blog article accepting unit 13 may accept a blog article transmitted from the blog server 3 every time the blog article is updated in the blog server 3. A newly updated blog article may be received from the blog server 3 by requesting transmission of a blog article at a fixed time every day.

  The machine learning unit 14 is a control unit that determines whether or not the blog article received by the blog article receiving unit 13 is a spam blog by using a learning model using, for example, an SVM (Support Vector Machine) engine 22. As the mechanism, the machine learning means 14 learns in advance a blog article of a spam blog and a blog article (a normal blog article) that is not a spam blog, performs a statistical process, The standard data which shows the standard of distinction from the blog article is generated beforehand. The determination using the SVM engine 22 of the machine learning means 14 is a blog that is an unknown case from a training case that belongs to one of two classes (a set of spam blog articles and a set of normal blog articles). This is to determine whether an article belongs to any class. Here, the machine learning unit 14 uses the predetermined keyword stored in the predetermined keyword DB 21 as a feature, and determines whether or not the blog article is a spam blog when the blog article includes the predetermined keyword. To do.

  Here, the learning result 30 in the learning model using the SVM engine 22 will be described with reference to FIG. The machine learning means 14 using the SVM engine 22 is a classifier that estimates a label for a blog article in which a label indicating whether or not a spam blog is unknown. Taking a spam blog as an example, the distance (margin) between the two discriminating surfaces 31 and 32 of a case generated from a blog article of a spam blog and a case generated from a normal blog article is on the feature space. The discriminating surfaces 31 and 32 are calculated so as to be maximum. As described above, the machine learning means 14 learns after converting a spam blog or the like into a case.

  The machine learning means 14 categorizes cases where the labels are unknown, using the cases generated from the spam blog closest to the identification surfaces 31 and 32 and the cases generated from the normal blog articles as support vectors 33 and 34, respectively. To use. The machine learning means 14 performs a statistical process on a case generated from a spam blog and a case generated from a normal blog article as a case. An identification surface, which is identification data for distinguishing the case group generated from the blog article, is generated.

  Then, the machine learning means 14 uses a learning result 30 for a blog article whose label indicating whether or not it is a spam blog, and uses a learning result 30 as a feature of a predetermined keyword as a feature. Classify by use. A feature characterizes input data. As described above, the spam blog determination apparatus 1 can make a determination by machine learning using the SVM engine 22 which is a known model. And the spam blog determination apparatus 1 can respond by using the machine learning means 14 and only maintaining the features without redoing the machine learning itself. Therefore, the spam blog determination apparatus 1 can be used to determine whether or not the spam blog.

  Returning to FIG. 1, the spam determination result output unit 15 is a control unit that outputs a blog article including a predetermined keyword and a determination result thereof to the management terminal 5 as a determination result by the machine learning unit 14.

  The adjustment return unit 17 is a control unit that deletes the predetermined keyword received from the predetermined keyword DB 21 by receiving a deletion request for the predetermined keyword transmitted from the management terminal 5.

  The storage unit 20 includes a predetermined keyword DB 21, an SVM engine 22, and a determination result table 23.

  The predetermined keyword DB 21 and the SVM engine 22 are as described above. The determination result table 23 is a data table storing the blog article including a predetermined keyword and the determination result, which is transmitted by the spam determination result output unit 15 to the management terminal 5 and is a determination result by the machine learning unit 14. is there.

  The determination result table 23 shown as an example in FIG. 2B includes items of a predetermined keyword 23a, a blog article 23b, and a spam blog determination 23c. The predetermined keyword 23a stores a predetermined keyword included in the blog article. The blog article 23b stores a blog ID (ID: identifier) that identifies the blog article. The blog article 23b may store the blog article itself. The spam blog determination 23c stores a code of a determination result of whether or not the machine learning means 14 is a spam blog. Here, when the spam blog determination 23c is “1”, it is determined that the machine learning unit 14 determines that the spam blog is a spam blog, and when the spam blog determination 23c is “0”, the machine learning unit 14 determines spam. This is the case when it is determined that it is not a blog.

  The hardware of the spam blog determination apparatus 1 of the first embodiment may be configured by a general computer. For example, a general computer includes a central processing unit (CPU) as the control unit 10 and a memory (RAM, ROM), a hard disk (HDD), and an optical disk (CD, DVD, etc.) as a storage unit 20 in a network. As a communication device, various wired and wireless LAN devices, as a display device, for example, various displays such as a liquid crystal display and a plasma display, and as an input device, for example, a keyboard and a pointing device (mouse, tracking ball, etc.) are appropriately provided. These are connected by a bus line. In such a general computer, the CPU controls the spam blog determination device 1 in an integrated manner, reads and executes various programs as appropriate, and cooperates with the above-described hardware, thereby providing various functions according to the present invention. Is realized.

  The blog server 3 is a server that stores blog articles, and includes a storage unit that stores blog articles and a control unit that controls the entire blog server 3. The hardware of the blog server 3 may be configured by a general computer.

  The management terminal 5 is, for example, a personal computer (PC) or a mobile terminal such as a mobile phone. The management terminal 5 may be any device as long as it has a communication function and can transmit / receive data to / from the spam blog determination device 1.

  In the first embodiment, the spam blog determination device 1 and the blog server 3 are described as separate devices. However, the spam blog determination device 1 also has the function of the blog server 3, and one unit It may be realized by a computer.

  The communication network 9 is a communication line such as the Internet for performing communication among the spam blog determination apparatus 1, the blog server 3, and the management terminal 5. The communication network 9 may be wired or part or all of it may be wireless.

[Processing of Spam Blog Determination Device 1]
Next, processing in the spam blog determination apparatus 1 will be described. First, reflection of a predetermined keyword will be described. FIG. 4 is a flowchart of the predetermined keyword reflection process of the spam blog determination apparatus 1 according to the first embodiment. This process is executed each time a specified keyword is specified from the management terminal 5.

  S1: The control unit 10 (predetermined keyword accepting means 11) determines whether or not designation of registration of a predetermined keyword has been accepted from the management terminal 5. When the designation of registration of the predetermined keyword is received (S1: YES), the control unit 10 moves the process to S2. On the other hand, when the designation of the registration of the predetermined keyword is not accepted (S1: NO), the control unit 10 moves the process to S3.

  S2: The control unit 10 (predetermined keyword storage control means 12) stores the predetermined keyword received in S1 in the predetermined keyword DB 21.

  S3: The control unit 10 (adjustment return means 17) determines whether or not the designation of deletion of the predetermined keyword has been received from the management terminal 5. When designation of deletion of the predetermined keyword is received (S3: YES), the control unit 10 moves the process to S4. On the other hand, when the designation of deletion of the predetermined keyword is not accepted (S3: NO), the control unit 10 ends this process.

  S4: The control unit 10 (adjustment return means 17) deletes the predetermined keyword received in S3 from the predetermined keyword DB 21. Then, the control part 10 complete | finishes this process.

  As described above, the spam blog determination device 1 receives the registration specification of the predetermined keyword that has become obsolete from the management terminal 5, and as a feature used in the determination process by machine learning of the spam blog including the predetermined keyword, the spam blog determination device 1 stores the specified keyword DB 21. Predetermined keywords can be registered. Moreover, the spam blog determination apparatus 1 can remove the predetermined keyword from the feature used in the determination process by machine learning by accepting the deletion designation of the predetermined keyword from the management terminal 5. Note that the administrator who is the user of the management terminal 5 sees the determination result output by the spam determination result output means 15 described later, and removes the predetermined keyword included in the blog determined as not a spam blog from the feature. Can be considered.

  Next, spam determination for blog articles will be described. This process is executed every time a blog article is received from the blog server 3. FIG. 5 is a flowchart of the spam determination process of the spam blog determination apparatus 1 according to the first embodiment.

  S <b> 11: The control unit 10 (blog article accepting unit 13) accepts a blog article from the blog server 3.

  S12: The control unit 10 (machine learning means 14) performs machine learning processing on the blog article received in S11. The machine learning process is, for example, using a learning model using the SVM engine 22 to determine whether or not the blog article accepted in S11 is a spam blog using the predetermined keyword stored in the predetermined keyword DB 21 as a feature. The process to do.

  S13: The control unit 10 (spam determination result output means 15) outputs the result of the machine learning process (spam determination result) executed in S12 to the management terminal 5. Further, the control unit 10 stores the spam determination result in the determination result table 23. Then, the control part 10 complete | finishes this process.

  Note that the control unit 10 may perform the predetermined keyword reflection process described in FIG. 4 and the spam determination process described in FIG. 5 in parallel.

  As described above, the spam blog determination apparatus 1 receives a blog article from the blog server 3, determines a spam blog including a predetermined keyword, and outputs the result. Therefore, the spam blog determination apparatus 1 can determine a spam blog including a predetermined keyword only by allowing the administrator to specify the predetermined keyword in advance. Moreover, the spam blog determination apparatus 1 can appropriately determine a spam blog including a predetermined keyword that has become obsolete by using the predetermined keyword DB 21 in which a predetermined keyword is registered and deleted according to an instruction from the management terminal 5.

(Second Embodiment)
In the first embodiment, the predetermined keyword received from the management terminal is deleted from the stored keyword, thereby removing it from the feature used in the determination process by machine learning as to whether it is a spam blog. In the second embodiment, a description will be given of a case where a predetermined keyword that satisfies a predetermined condition is automatically deleted, so that it is excluded from the feature used in the determination process by machine learning as to whether or not it is a spam blog. In the following description, parts that perform the same functions as those in the first embodiment described above are given the same reference numerals or the same reference numerals at the end, and redundant descriptions are omitted as appropriate.

[Overall Configuration of Spam Blog Determination System 200 and Functional Configuration of Spam Blog Determination Device 201]
FIG. 6 is a diagram showing the overall configuration of the spam blog determination system 200 and the functional configuration of the spam blog determination apparatus 201 according to the second embodiment.

  The spam blog determination system 200 includes a spam blog determination apparatus 201, a blog server 3, a management terminal 5, and a communication network 9.

  The spam blog determination apparatus 201 includes a control unit 210 and a storage unit 20. In addition to the predetermined keyword receiving unit 11, the predetermined keyword storage control unit 12, the blog article receiving unit 13, the machine learning unit 14, and the spam determination result output unit 15, the control unit 210 includes a second epidemic end determination unit. 216 and adjustment return means 217.

  The second fashion end determination means 216 has a predetermined ratio (for example, one week) of the blog articles determined to be spam blogs by the machine learning means 14 among the blog articles to be determined including the predetermined keyword. This is a control unit that determines that the trend of a predetermined keyword has ended when the threshold value (for example, 20%) or less is reached.

  The adjustment return unit 217 is a control unit that deletes, from the predetermined keyword DB 21, the predetermined keyword determined to have ended by the second trend end determination unit 216.

[Processing of Spam Blog Determination Device 201]
Next, processing in the spam blog determination apparatus 201 will be described. FIG. 7 is a flowchart of the predetermined keyword deletion process of the spam blog determination apparatus 201 according to the second embodiment. The registration of the predetermined keyword is the same as the processing of S1 to S2 (FIG. 4) of the first embodiment, and the spam determination is the same as the processing of S11 to S13 (FIG. 5) of the first embodiment. is there.

  S21: The control unit 210 (second fashion end determination unit 216) determines the ratio of the blog articles determined to be spam blogs by the machine learning unit 14 among the determination target blog articles including the predetermined keyword in a predetermined period. It is determined whether or not it is below a predetermined threshold value. When it is below the predetermined threshold (S21: YES), the control unit 210 moves the process to S22. On the other hand, when it is not below the predetermined threshold value (S21: NO), the control unit 210 ends this process.

  S22: The control unit 210 (adjustment return means 217) deletes the predetermined keyword determined to be equal to or less than the predetermined threshold in S21 from the predetermined keyword DB 21. Thereafter, the control unit 210 ends this process.

  As described above, the spam blog determination apparatus 201 has a ratio of blog articles determined by the machine learning means 14 as spam blogs to be equal to or less than a predetermined threshold among blog articles to be determined that include a predetermined keyword in a predetermined period. In this case, it is determined that the trend has ended for the predetermined keyword. Therefore, the spam blog determination apparatus 201 can determine the “temporary” trend end state of a predetermined keyword based on a predetermined criterion.

  And since the spam blog determination apparatus 201 deletes the predetermined keyword determined that the fashion has ended from the predetermined keyword DB 21, the maintenance of the predetermined keyword DB 21 can be automatically performed.

(Third embodiment)
In the third embodiment, the spam blog determination apparatus accepts a correct / incorrect determination of the result of determination by machine learning from the management terminal, automatically deletes a predetermined keyword satisfying a predetermined condition according to the error rate, What is excluded from the features used in the determination process based on whether or not there is machine learning will be described.

[Overall Configuration of Spam Blog Determination System 300 and Functional Configuration of Spam Blog Determination Device 301]
FIG. 8 is a diagram illustrating the overall configuration of the spam blog determination system 300 and the functional configuration of the spam blog determination apparatus 301 according to the third embodiment.

  The spam blog determination system 300 includes a spam blog determination device 301, a blog server 3, a management terminal 5, and a communication network 9.

  The spam blog determination apparatus 301 includes a control unit 310 and a storage unit 320. The control unit 310 includes the predetermined keyword receiving unit 11, the predetermined keyword storage control unit 12, the blog article receiving unit 13, the machine learning unit 14, the spam determination result output unit 15, and the adjustment return unit 217. Correct / error determination information storage control means 318 and first fashion end determination means 319 are provided.

  The right / wrong judgment information storage control means 318 receives right / wrong judgment information indicating the right / wrong judgment with respect to the result of machine learning determined to be a spam blog or not a spam blog from the management terminal 5, and the right / wrong judgment information DB 324. It is a control part to memorize. The correctness / incorrectness determination information to be accepted is calculated by associating with a blog article including a predetermined keyword whether it is correct for the determination that it is a spam blog and whether it is correct for the determination that it is not a spam blog. There are four judgments. In addition, the correctness determination information received from the management terminal 5 is input by the administrator. And it is not necessary for the administrator to input correct / incorrect determination information for all the determination results output by the spam determination result output means 15, and it may be one sampling for several cases, or for each predetermined keyword. You may narrow down to.

  The first fashion end determination unit 319 calculates an error rate for each predetermined keyword based on the correctness / incorrectness determination information stored in the correctness / incorrectness determination information DB 324 in a predetermined period (for example, one week). The control unit determines that the trend of a predetermined keyword has ended when a threshold (for example, 70%) or more is reached.

  The correctness / incorrectness determination information DB 324 stores the correctness / incorrectness of the result of determination by machine learning received by the correctness / incorrectness determination information storage control unit 318 in association with a blog article including a predetermined keyword.

[Processing of Spam Blog Determination Device 301]
Next, processing in the spam blog determination apparatus 301 will be described. FIG. 9 is a flowchart of predetermined keyword deletion processing of the spam blog determination apparatus 301 according to the third embodiment. The registration of the predetermined keyword is the same as the processing of S1 to S2 (FIG. 4) of the first embodiment, and the spam determination is the same as the processing of S11 to S13 (FIG. 5) of the first embodiment. is there.

S31: The control unit 310 (correction determination information storage control unit 318) accepts correctness determination information from the management terminal 5.
S32: The control unit 310 (correction determination information storage control unit 318) stores the received correctness determination information in the correctness determination information DB 324.
S33: The control unit 310 (first fashion end determination unit 319) calculates an error rate of determination by machine learning for each predetermined keyword based on the correctness determination information stored in the correctness determination information DB 324 in a predetermined period.
S34: The control unit 310 (first fashion end determination means 319) determines whether or not the error rate is equal to or higher than a predetermined threshold value. If it is equal to or greater than the predetermined threshold (S34: YES), the controller 310 moves the process to S35. On the other hand, when it is not equal to or greater than the predetermined threshold (S34: NO), the control unit 310 ends this process.

  S35: The control unit 310 (adjustment returning means 217) deletes the predetermined keyword determined to be equal to or greater than the predetermined threshold in S34 from the predetermined keyword DB 21. Then, the control part 310 complete | finishes this process.

  As described above, the spam blog determination device 301 accepts correct / incorrect determination information on the result of the machine learning determination for the blog article that outputs the determination result among the determination target blog articles including the predetermined keyword, and handles the blog article. In addition, the error rate is calculated for each predetermined keyword based on the correctness / incorrectness determination information stored in the correct / incorrect determination information DB 324 and the error rate is equal to or greater than a predetermined threshold value. It is determined that the process has ended. Therefore, the spam blog determination apparatus 301 can automatically determine the “temporary” trend end state of a predetermined keyword based on a predetermined criterion simply by the administrator inputting correctness of determination by machine learning.

  And since the spam blog determination apparatus 301 deletes the predetermined keyword determined that the fashion has ended from the predetermined keyword DB 21, the maintenance of the predetermined keyword DB 21 can be performed automatically.

(Deformation)
In the first embodiment, deletion of a predetermined keyword from the predetermined keyword DB is performed according to an input from the management terminal. In the second and third embodiments, the spam blog determination device deletes the predetermined keyword from the predetermined keyword DB. Although shown as being performed automatically, it is not limited to this. Any deletion may be performed. By doing so, it is convenient because a predetermined keyword that does not need to be monitored can be automatically deleted and can be deleted by the operation of the administrator.

  As mentioned above, although embodiment of this invention was described, this invention is not restricted to embodiment mentioned above. The effects described in the embodiments of the present invention are only the most preferable effects resulting from the present invention, and the effects of the present invention are limited to those described in the embodiments of the present invention. is not.

1, 201, 301 Spam blog determination device 3 Blog server 5 Management terminal 10, 210, 310 Control unit 11 Predetermined keyword receiving means 12 Predetermined keyword storage control means 13 Blog article receiving means 14 Machine learning means 15 Spam determination result output means 17, 217 Adjustment returning means 20,320 Storage unit 21 Predetermined keyword DB
22 SVM engine 23 Judgment result table 100, 200, 300 Spam blog judgment system 216 Second epidemic end judgment means 318 Correct / incorrect judgment information storage control means 319 First epidemic end judgment means 324 True / false judgment information DB

Claims (5)

Predetermined keyword storage control means for storing the predetermined keyword in the predetermined keyword storage means in response to receiving registration designation of the predetermined keyword;
Machine learning means for determining whether or not the blog article is a spam blog by machine learning using the predetermined keyword stored in the predetermined keyword storage means as a feature in response to receiving the determination target blog article When,
Among the blog articles to be determined by the machine learning means, a blog article including the predetermined keyword stored in the predetermined keyword storage means, and a result of the determination by the machine learning as to whether or not it is a spam blog. A spam determination result output means for outputting in association with each other;
An adjustment return unit for deleting the predetermined keyword stored in the predetermined keyword storage unit in response to receiving the deletion specification of the predetermined keyword;
In response to accepting the correctness / incorrectness determination information, which is information indicating the correctness / incorrectness of the determination result by the machine learning, for the blog article for which the spam determination result output means has output the determination result by the machine learning, Correctness determination information storage means for storing determination information in association with the blog article;
An error rate for determination by the machine learning is calculated for each of the predetermined keywords based on the correctness / incorrectness determination information stored in the correctness / incorrectness determination information storage unit in a predetermined period, and the predetermined error rate is equal to or greater than a predetermined threshold. A first fashion end judging means for judging that the fashion has ended for the keyword;
A spam blog determination device comprising: Predetermined keyword storage control means for storing the predetermined keyword in the predetermined keyword storage means in response to receiving registration designation of the predetermined keyword;
Machine learning means for determining whether or not the blog article is a spam blog by machine learning using the predetermined keyword stored in the predetermined keyword storage means as a feature in response to receiving the determination target blog article When,
Among the blog articles to be determined by the machine learning means, a blog article including the predetermined keyword stored in the predetermined keyword storage means, and a result of the determination by the machine learning as to whether or not it is a spam blog. A spam determination result output means for outputting in association with each other;
An adjustment return unit for deleting the predetermined keyword stored in the predetermined keyword storage unit in response to receiving the deletion specification of the predetermined keyword;
Among the blog articles to be determined that include the predetermined keyword in a predetermined period, the ratio of the blog articles that the machine learning means determines to be spam blogs is equal to or lower than a predetermined threshold value. A second fashion end determination means for determining that the fashion has ended for the keyword;
A spam blog determination device comprising: The adjustment return means deletes the predetermined keyword determined to have ended by the first trend end determination means or the second trend end determination means from the predetermined keyword storage means;
The spam blog determination apparatus according to claim 1 or 2 . A predetermined keyword storage step for storing the predetermined keyword in a predetermined keyword storage means in response to the computer accepting the registration specification of the predetermined keyword ;
When the computer accepts the blog article to be determined, the machine determines whether the blog article is a spam blog by using the predetermined keyword stored in the predetermined keyword storage unit as a feature. Machine learning steps,
Among the blog articles to be determined whether or not the computer is a spam blog, the blog article including the predetermined keyword stored in the predetermined keyword storage unit and the machine learning whether or not the computer is a spam blog A spam determination result output step for outputting the determination result in accordance with
An adjustment return step of deleting the predetermined keyword stored in the predetermined keyword storage means in response to the computer receiving the deletion specification of the predetermined keyword;
In response to accepting correct / incorrect determination information, which is information indicating correct / incorrect determination of the result of determination by machine learning, for the blog article that has output the determination result by machine learning in the spam determination result output step , A correct / incorrect determination information storing step for storing the correct / incorrect determination information in association with the blog article;
The computer calculates an error rate for determination by the machine learning for each of the predetermined keywords based on the accuracy determination information stored in the accuracy determination information storage step in a predetermined period, and the error rate is equal to or greater than a predetermined threshold. A first fashion end determination step for determining that the fashion has ended for the predetermined keyword;
Spam blog determination method including. A predetermined keyword storage step for storing the predetermined keyword in a predetermined keyword storage means in response to the computer accepting the registration specification of the predetermined keyword;
When the computer accepts the blog article to be determined, the machine determines whether the blog article is a spam blog by using the predetermined keyword stored in the predetermined keyword storage unit as a feature. Machine learning steps,
Among the blog articles to be determined whether or not the computer is a spam blog, the blog article including the predetermined keyword stored in the predetermined keyword storage unit and the machine learning whether or not the computer is a spam blog A spam determination result output step for outputting the determination result in accordance with
An adjustment return step of deleting the predetermined keyword stored in the predetermined keyword storage means in response to the computer receiving the deletion specification of the predetermined keyword;
When the ratio of the blog articles determined to be spam blogs in the machine learning step out of the blog articles to be determined that include the predetermined keyword in a predetermined period is equal to or less than a predetermined threshold, A second trend end determination step for determining that the trend has ended for the predetermined keyword;
Spam blog determination method including.

Images