JP4554493B2 - Document issuer classification apparatus and program - Google Patents

Description

The present invention relates to a document publisher content RuiSo location and program for classifying issuer of the document, in particular, published document for classifying whether an unknown document issuer corresponds to any of the two or more types on Sha minute RuiSo location and program.

  Currently, various documents are released on the Internet, and the amount of e-mails going back and forth on the Internet has become enormous.

  Recently, a simple HMTL creation system called Weblog has become widespread, making it easier for individuals to disclose information than before. In this situation, the proportion of the total amount of advertisement information issued to a large number of unspecified users with respect to the total circulation amount has increased, and the effort to dispose of useful information for users has increased.

  For example, in the case of e-mail, there is a technique for identifying and discarding e-mails that are not required by the user, such as advertisements, mainly from e-mails received by the user. This technology obtains the words in the mail document, generates a vector representing the characteristics of each mail, and calculates the similarity with the vector representing the characteristics of the known advertising mail, thereby determining whether or not the advertisement is (For example, refer to Patent Document 1).

In addition, some search engines on the Internet have a technology that automatically determines whether or not a page is harmful by using a word that appears in the page, in order to prevent a page harmful to children from being displayed in a search result. Some search engines are incorporated (see Non-Patent Document 1, for example).
JP 2004-048492 A http://partners.search.goo.ne.jp/howto/function/“Block harmful content ”

  For the first time in the above existing technologies, multiple technologies have been devised to determine whether each document is an advertisement for a document group that contains documents containing useful information for users and documents such as advertisements that the user does not want. Has been.

  However, in the above existing technology, for example, a keyword including keywords often used for product advertisements such as “large special price” and “limited time” and a specific product name is given as a representative of the advertisement document, and unknown based on this If it is determined whether or not the document is an advertisement, there is a high possibility that a non-advertisement document including these keywords will be accidentally determined to be an advertisement. For example, there may be a case where a customer who has purchased a certain product determines that a useful document describing usage impressions is an advertising document. Such a problem is unavoidable in a technique in which the determination is made only with the appearance word without considering the context as in the existing technique.

The present invention has been made in view of the above points, and whether a new document is a document that is not desired by the user, for example, an advertisement-like document, especially when it is difficult to distinguish only by a word in the document. and to provide a corresponding document publisher component that can be determined RuiSo location and program.

  FIG. 1 is a diagram for explaining the principle of the present invention.

This onset Ming, a document issuer classification method for classifying the issuer of the document into two or more types,
The document group described by the same issuer is acquired from the learning document storage means that stores the document group for learning the classification rule of the document issuer together with the classification information of each document issuer, and between the titles of two different documents, Calculate one or more similarities between texts and link information in the text, and calculate the average value (average similarity) of similarities for each issuer, which is the characteristics of the issuer A similarity calculation step (step 1) as a quantity;
Using the feature amount of each issuer and the known classification information of the issuer acquired from the learning document storage means, the classification rule for classifying the issuer whose classification result is unknown is learned, and the learning result is stored in the classification rule A classification rule learning step (step 2) to be stored in the means;
A document group described by the same issuer is acquired from a classification target document storage unit storing a document group described by an issuer whose classification result is unknown, and a feature amount is calculated in the same procedure as the similarity calculation step ( Step 3) performs a classification step (step 4) in which the issuer's classification destination is determined based on the classification rule recorded in the classification rule storage means, and the classification result is stored in the classification result storage means.

  FIG. 2 is a principle configuration diagram of the present invention.

The present invention (Claim 1 ) is a document issuer classification apparatus for classifying document issuers into two or more types,
Learning document storage means 101 storing a document group for learning the document issuer classification rules together with the classification information of each document issuer;
A classification target document storage unit 102 storing a document group described by an issuer whose classification result is unknown;
The learning document storage unit 101 obtains a document group described by the same issuer, calculates one or more similarities between titles between two different documents, between texts, and between link information in the text, and A similarity calculation means 104 that calculates an average value (average similarity) of each issuer and uses this as a feature amount of the issuer;
Using the feature quantity of each issuer and the known classification information of the issuer acquired from the learning document storage unit 101, the classification rule for classifying the issuer whose classification result is unknown is learned, and the learning result is classified Classification rule learning means 105 stored in the rule storage means 106,
The document group described by the same issuer is acquired from the classification target document storage unit 102, the feature amount is calculated in the same procedure as the similarity calculation unit, and based on the classification rule recorded in the classification rule storage unit 106 A classification unit 107 that determines a classification destination of the issuer and stores the classification result in the classification result storage unit 108.

Further, according to the present invention (claim 2 ), a specific type of keyword is acquired from each document acquired from the learning document storage unit 101 and the classification target document storage unit 102, and the average number of appearances of the keyword per document is obtained. And a keyword counting means for calculating for each issuer,
The classification rule learning means 105 and the classification means 107 are:
Furthermore, a means for adding the average number of appearances as a feature quantity and learning classification rules and classifying document issuers is included.

Further, the present invention (Claim 3 ) is a keyword counting means,
A learning document storage unit 101 and a unit for extracting a monetary expression in a document acquired from the classification target document storage unit 102 and totaling the average number of appearances of the monetary expression per document are included.

According to the present invention (claim 4 ), learning document storage means storing a document group serving as learning material for classifying document issuers, and a discrimination target document storage storing a document group described by an issuer whose classification result is unknown A computer having means, classification rule storage means, and classification result storage means,
A document issuer classification program that functions as the document issue classification device according to any one of claims 1 to 3 .

As described above, according to the present invention, as a feature quantity for classifying an issuer, a keyword of a specific type such as a similarity between documents described by the same issuer or a monetary expression is used instead of a word appearance frequency. Is used. The present invention does not classify based on the meaning of the document, but performs classification based on characteristics of the document described by the issuer, such that the issuer who accidentally used a specific word such as a product name is an advertisement. The problem of misclassification can be solved. Also, by calculating the feature amount after sorting the document group of the same publisher, you can select and use any of the fixed phrases at random. The feature amount difference between the issuer for advertising purposes and the issuer that describes non-advertisement documents with different contents such as diaries every day can be clarified, and these can be classified correctly.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

  The present invention records a classification rule learned from a learning document whose classification result of the document issuer is known, and determines a classification destination of the issuer using a document described by an issuer whose classification result is unknown To do.

[First Embodiment]
FIG. 3 is a configuration diagram of the document issuer classification apparatus according to the first embodiment of the present invention.

  The document issuer classification apparatus shown in FIG. 1 includes a learning document recording unit 101, a classification target document recording unit 102, an order determination unit 103, a similarity calculation unit 104, a classification rule learning unit 105, a classification rule recording unit 106, and a classification unit 107. The classification result recording unit 108 is configured.

  The learning document recording unit 101 is a recording medium such as a disk device, and records a learning group serving as a learning material for classifying document issuers.

  The classification target document recording unit 102 is a recording medium such as a disk device, and records a document group described by an issuer whose classification result is unknown. A document to be recorded in the learning document recording unit 101 includes an issuer ID, an issuer classification (advertisement / non-advertisement), a document ID, a title, a body text, and the like. A document to be recorded in the classification target document recording unit 102 is an issuer ID. , Document ID, title, text, etc.

  The order determination unit 103 rearranges the document groups recorded in the learning document recording unit 101 when the classification rule is learned and in the classification target document recording unit 102 when the issuer is classified, based on a certain condition.

  The similarity calculation unit 104 calculates the average similarity between adjacent documents in the learning document recording unit 101 or the classification target document recording unit 102 according to the order determined by the order determination unit 103.

  The classification rule learning unit 105 uses the similarity calculated by the similarity calculation unit 104 for the document group recorded in the learning document recording unit 101 to learn a classification rule for classifying an issuer whose classification result is unknown. To do.

  The classification rule recording unit 106 is a recording medium such as a disk device, and records the classification rule obtained as a result of learning in the classification rule learning unit 105.

  The classification unit 107 displays the average similarity for each issuer calculated by the similarity calculation unit 104 for the document group recorded in the classification target document recording unit 102 and the classification rule recorded in the classification rule recording unit 106. Used to classify the issuer.

  The classification result recording unit 108 records the classification result of the classification unit 107.

  FIG. 4 is a flowchart showing an outline operation in the first embodiment of the present invention.

  In the similarity calculation unit 104, a learning document whose classification result of the document issuer is known is read from the learning document recording unit 101, and adjacent documents of the same issuer are compared to link information between titles, between texts, and in the text. Any one or more similarities between them are calculated. By calculating various similarities between adjacent documents across all documents of the same publisher and averaging the similarities, the average similarity between titles, the average similarity between texts, and the average between link information in the text The similarity is calculated and used as the feature amount of the issuer (step 101).

  Subsequently, the classification rule learning unit 105 learns the classification rule using the classification result of each issuer acquired from the learning document recording storage unit 101 and the feature amount obtained in step 101, and the classification rule as a result of the learning is obtained. Is recorded in the classification rule recording unit 106 (step 102).

  On the other hand, the document described by the issuer whose classification result is unknown, obtained from the classification target document recording unit 102, is analyzed, and the feature amount is calculated in the same manner as in step 101 (step 103), and the classification rule storage unit 106 is obtained. The most likely classification destination of the issuer is determined by comparing with the classification rule (step 104).

  Prior to the processing in step 101 described above, the order determination unit 103 rearranges the documents described by the same issuer according to a certain standard. In step 101, the similarity calculation unit 104 rearranges the results. Based on the above, the similarity between two adjacent documents may be calculated. In this case, the document group acquired from the classification target document recording unit 102 is rearranged according to the same standard in the order determination unit 103 before the processing of step 103 is performed.

  Hereinafter, the operation will be described in detail based on the configuration of FIG.

  In the document issuer classification apparatus of the present invention, the classification rule learning unit 105 records in advance a classification rule learned using a document group described by a publisher whose classification result is known in the classification rule recording unit 106, The classification unit 107 determines a classification destination of the issuer using a document group described by an issuer whose classification result of the classification target document recording unit 102 is unknown. For this reason, as a learning document stored in the learning document recording unit 101, a plurality of documents described by the same issuer are prepared for the issuer, and a classification result is given to each issuer in advance. There is a need.

  It is also assumed that there are multiple documents for issuers whose classification results are unknown.

  FIG. 5 is an example of a learning document in the learning document recording unit in one embodiment of the present invention. FIG. 4A shows an example of a document whose classification information is “advertisement”, and FIG. 4B shows an example of a document whose classification information is “non-advertisement”.

  Each document has attributes of an issuer ID representing an issuer, a classification indicating whether the document described by the issuer is an advertisement for the purpose of advertising products, etc., a document ID, a title, and a text. . For example, in the example of the advertisement document in FIG. 5A, the document described by the publisher ID “ABCD” is an advertisement, and the title of the document with the document ID “0001” is “Today's attention product (September 1). ”And has a text that continues with“ Products introduced today… ”. The text includes link information (URL) to a document describing detailed information of the product. Similarly, the example of the non-advertisement document shown in FIG. 5B has attributes such as a classification “non-advertisement”. In the document issuer classification apparatus according to the present invention, since it is determined for each issuer ID whether it is an advertisement or a non-advertisement, all the classifications of the same issuer are assumed to be unified.

  Next, classification rule learning will be described.

  FIG. 6 is a flowchart of the classification rule learning procedure according to the first embodiment of the present invention.

  First, the order determination unit 103 refers to the learning document recording unit 101, rearranges the documents described by the same issuer, and stores the result in the learning document recording unit 101 (step 301). For example, the reordering is performed in a lexicographic arrangement based on the document title. For example, a document having an issuer ID “ABCD” issuer shown in FIG. 5 having a title of the form “Today's attention product (○ month × day)” and a document having a title of “Yesterday order quantity ranking” Suppose that the issuer is a frequent issuer. When these are rearranged into a lexicographical arrangement based on titles, the documents are separated for each title of the same format, and a group of documents titled “Yesterday's order quantity ranking” is aggregated in the first half. A group of documents having a title of “product (○ month × day)” is collected. In other words, the first half and the second half are separated according to the title type. Even in the case of an issuer who uses three or more title types, an order in which titles of the same format continue can be similarly obtained by this rearrangement.

  However, the processing of step 301 is not essential and does not necessarily need to be rearranged.

  When the rearrangement of the document group described by one issuer is completed, the similarity calculation unit 104 calculates the average similarity between adjacent documents (step 302). At this time, the title and the text are acquired from the learning document recording unit 101 one document at a time based on the result rearranged by the order determining unit 103. The similarity between adjacent documents is calculated by comparing the title and body of the document acquired immediately before, the similarity is calculated over all the documents of the issuer “ABCD”, and the similarity is averaged to obtain the average similarity Is calculated.

  As the average similarity calculated here, the similarity of the title, the similarity of the text, and the similarity of the link information are used.

  For the similarity between titles, it is preferable to use the similarity between vectors composed of words included in the title. For example, five words “Today”, “Attention”, “Product”, “September”, and “1 day” are acquired from the title “Today's attention product (September 1)”, and a word vector is constructed. The word vectors of the titles configured in this way are compared between adjacent documents, a value representing the strength of relevance between vectors such as Euclidean distance and cosine similarity is calculated, and the similarity is determined. According to this method, comparing two identical titles of “Yesterday's order quantity ranking” maximizes the similarity, and even when comparing titles in the format of “Today's hot products (○ month x day)” The similarity is high because some are in common. Therefore, the average similarity of titles calculated for the document with the publisher ID “ABCD” is high. On the other hand, the issuer who describes and issues a diary-type document like the document of the issuer ID “EFGH” given as an example of the non-advertisement document in FIG. The average similarity is lower than that of the issuer “ABCD”.

  Similar to the similarity of the title, the similarity of the body should be calculated from a word vector. In this case, too, for example, a document with a title of the form “Today's hot product (○ month × day)” issued by the issuer “ABCD” has expressions such as “order number” and “click here for details” Since the document titled “Yesterday's order quantity ranking” has many fixedly used expressions such as “first place” and “second place”, the average similarity is increased.

  The similarity of the link information is determined by analyzing the URL included in the text of the example shown in FIG. Assuming that the link information included in the example of the advertisement document in FIG. 5A leads to a document describing the detailed information of the product introduced in the text, links in other documents issued by the same publisher The information is also likely to be guided to a detailed product information document. The URL is a character string representing the material of a document on the Internet. For example, there are many common parts between URLs of similar documents that introduce products, and the similarity of the URL itself is high. Since the URL represents the hierarchical position of the document location by the symbol “/”, the similarity between the link information can be obtained by treating a plurality of character strings obtained by dividing the URL with a slash like the above word vector. Can be calculated. On the other hand, in a diary-type document such as a non-advertisement document as shown in FIG. 5B, even if the same issuer is used, there is a high probability that link information will be different if the document is different. Lower.

  As described above, the average similarity calculated for each of the title, text, and link information is stored in the memory (not shown) by the classification rule learning unit 105 as the feature amount of the issuer “ABCD” or “EFGH”. Then, the process is repeated until the calculation of the average similarity for all the issuers is completed (step 303, Yes).

  Subsequently, the classification result is unknown using the various feature amounts of all the issuers calculated up to this point by the classification rule learning unit 105 and the classification (advertisement or non-advertisement) of each issuer recorded in the learning document recording unit 101. The classification rule for classifying the issuer is learned (step 304), and the learning result is recorded in the classification rule recording unit 106 (step 305).

  The method of learning rules using the feature quantities of multiple issuers and the known classification of each issuer is, for example, “Statistics of Pattern Recognition and Learning” Aso et al., Iwanami Shoten, pp20-23, pp107. -108 ", such as decision trees and support vector machines. Each of these learns based on the feature quantity of the publisher whose classification is known, learns and records the classification rule for classification, and the feature quantity calculated for the issuer whose classification is unknown is stored in the classification rule. This is a technique for classifying issuers into the most likely classification destinations.

  Finally, a procedure for executing classification of an issuer whose classification result is unknown will be described.

  FIG. 7 is a flowchart of a classification procedure for an issuer whose classification result is unknown according to the first embodiment of this invention.

  The processing of step 401 and step 402 is exactly the same as that of step 301 and step 302 except that the processing is performed using the document recorded in the classification target document recording unit 102 instead of the learning document recording unit 101. .

  As in FIG. 5, the classification target document stored in the classification target document recording unit 102 has an issuer ID, a document title, and a body text. However, since the advertisement or non-advertisement is unknown, the classification attribute is do not have.

  For the issuer whose classification result is unknown, when the calculation of the average similarity of title, text, and link information is finished, the classification unit 107 refers to the classification rule that is the learning result recorded in the classification rule recording unit 106, Whether the issuer is an advertisement or a non-advertisement is classified with a higher possibility (step 403). The classification result is output to the classification result storage unit 108 as a pair with the issuer ID (step 404).

  This completes the classification process for one issuer, repeats it for all issuers (step 405), and ends the process.

  FIG. 8 shows an example of the classification result in the first exemplary embodiment of the present invention.

  The classification results obtained above are stored in the classification result storage unit 108 as shown in FIG. In this embodiment, since it is determined whether it is an advertisement or a non-advertisement, the issuer “PPPP” is an advertisement, “QQQQ” is a non-advertisement, and so on. Yes.

[Second Embodiment]
FIG. 9 shows the configuration of the document issuer classification apparatus according to the second embodiment of the present invention. The same components as those in FIG.

  In the present embodiment, the same issuer is used as the feature amount calculated in the above-described step 101 and step 103 shown in FIG. 4 with one or more average similarities between titles, between texts, and between link information in texts. The feature amount is obtained by adding the average number of occurrences of a specific type of keyword in the document described by. The keyword may be a monetary expression such as “1,000 yen” or “$ 48”, for example.

  The document issuer classification apparatus shown in the figure is obtained by adding a keyword totaling unit 601 that calculates the average number of appearances of a specific type of keyword in a document to the issuer in addition to the configuration shown in FIG.

  FIG. 10 is a flowchart of the classification rule learning procedure according to the second embodiment of the present invention.

  Based on the order rearranged in the lexicographic arrangement of titles in the order determination unit 103, the similarity calculation unit 104 calculates an average similarity between adjacent documents (steps 701 and 702). This process is the same as that in the first embodiment. In this embodiment, in addition to the average similarity between the title and the text link information, the keyword count unit 601 counts the average number of appearances of the keyword (step 703). The average number of appearances of a keyword is the average number of appearances of a specific type of keyword per document described by a certain publisher.

  This embodiment is also an example in which the issuer is classified into an advertisement and a non-advertisement, and it is preferable to use the entire monetary expression as a keyword. That is, the keyword totaling unit 601 acquires the money amount expression in each document of the learning document recording unit 101 regardless of the amount of money, and totals the number of appearances in the sentence. This is performed for each document, and the average number of appearances of the monetary expression is calculated for each issuer.

  In step 704, when the keyword totaling unit 601 finishes calculating the average similarity for all issuers and the average number of appearances of the keyword (in this embodiment, money expression) (step 704, Yes), the classification rule learning unit 105 A classification rule is learned using both the average similarity and the average number of appearances (step 705), and the rule learning result is recorded in the classification rule recording unit 106 (step 706).

  FIG. 11 is a flowchart of a procedure for classifying issuers whose classification results are unknown according to the second embodiment of this invention.

  Similar to the learning procedure shown in FIG. 10, in addition to the calculation of the average similarity in the similarity calculation unit 104 (steps 801 and 802), the keyword totaling unit 601 totals the average number of appearances of the monetary expression (steps). 803). Next, the classification unit 107 classifies the advertisement or non-advertisement using both the average similarity and the average number of appearances (step 804), and records the result in the classification result recording unit 108 (step 805).

  The operations of the first and second embodiments described above can be constructed as a program and installed in a computer used as a document issuer classification device for execution or distributed via a network. is there.

  The constructed program can be stored in a portable storage medium such as a hard disk device or a flexible disk / CD-ROM, and installed in a computer to be executed.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made within the scope of the claims.

  The present invention is applicable to a technique for classifying issuers who issue documents.

It is a figure for demonstrating the principle of this invention. It is a principle block diagram of this invention. It is a block diagram of the document issuer classification apparatus in the 1st Embodiment of this invention. It is a flowchart of the outline | summary operation | movement in the 1st Embodiment of this invention. It is an example of the learning document of the learning document recording part in the 1st Embodiment of this invention. It is a flowchart of the learning procedure of the classification rule in the 1st Embodiment of this invention. It is a flowchart of the classification | category procedure of the issuer whose classification result is unknown in the 1st Embodiment of this invention. It is an example of the classification result in the 1st Embodiment of this invention. It is a block diagram of the document issuer classification device in the 2nd Embodiment of this invention. It is a flowchart of the learning procedure of the classification rule in the 2nd Embodiment of this invention. It is a flowchart of the classification | category procedure of the issuer whose classification result is unknown in the 2nd Embodiment of this invention.

Explanation of symbols

101 learning document storage unit, learning document recording unit 102 classification target document storage unit, classification target document recording unit 103 order determination unit 104 similarity calculation unit, similarity calculation unit 105 classification rule learning unit, classification rule learning unit 106 classification rule storage Means, classification rule storage unit 107 Classification unit, classification unit 108 Classification result storage unit, classification result storage unit 601 Keyword totaling unit

Claims (4)

A document issuer classification device for classifying document issuers into two or more types,
Learning document storage means for storing a document group for learning the classification rule of the document issuer together with the classification information of each document issuer;
Classification target document storage means storing a document group described by an issuer whose classification result is unknown,
A document group described by the same publisher is acquired from the learning document storage means, and one or more similarities between titles between two different documents, between texts, and between link information in the text are calculated, A similarity calculation means for calculating an average value (average similarity) of each issuer and using this as a feature amount of the issuer;
Using the feature quantity of each issuer and the known classification information of the issuer acquired from the learning document storage means, the classification rule for classifying the issuer whose classification result is unknown is learned, and the learning result is classified A classification rule learning means to be stored in the rule storage means;
A classification rule recorded in the classification rule storage unit by acquiring a document group described by the same issuer from the classification target document storage unit, calculating the feature amount in the same procedure as the similarity calculation unit A classification unit that determines a classification destination of the issuer based on the classification result and stores the classification result in the classification result storage unit;
A document issuer classifying apparatus comprising: Keyword learning means for acquiring a specific type of keyword from each document acquired from the learning document storage means and the classification target document storage means, and calculating an average number of appearances of the keyword per document for each issuer In addition,
The classification rule learning means, and the classification means,
Further, the average number of appearances is added as a feature amount, and includes means for learning classification rules and classifying document publishers.
The document issuer classification apparatus according to claim 1 . The keyword counting means is:
The learning document storage means, and a means for extracting the money amount expression in the document acquired from the classification target document storage means, and totaling the average number of appearances of the money amount expression per document,
The document issuer classification device according to claim 2. Learning document storage means storing a document group as a learning material for classifying document issuers, distinction target document storage means storing a document group described by an issuer whose classification result is unknown, classification rule storage means, classification result A computer having storage means;
A document issuer classification program which functions as the document issue classification apparatus according to claim 1 .

Images