JP6235443B2 - Extraction apparatus, extraction method and extraction program - Google Patents

Description

  The present invention relates to an extraction apparatus, an extraction method, and an extraction program.

  Enormous amounts of information are disclosed on the Internet, and users often use search sites when searching for information they want to know. In such a search site, advertisement distribution called search matching advertisement or search-linked advertisement may be performed.

  In this type of advertisement distribution, a technique is known in which an advertiser designates one or more advertisement-related keywords when submitting an advertisement. In this technique, when an advertisement-related keyword is input as a search keyword from a user, an advertisement related to the advertisement-related keyword is displayed adjacent to a search result corresponding to the search keyword.

JP 2010-211777 A

  However, with the above-described conventional technology, it is not always possible to extract keywords that enhance the appeal effect of content. Specifically, the above-described conventional technology does not disclose a method in which an advertiser appropriately specifies an advertisement-related keyword for an advertisement that is a submitted content. In search-linked advertisement distribution, appropriate advertisements cannot be distributed unless advertisement-related keywords are appropriately specified. Thereby, there exists a possibility that the appeal effect of the advertisement distributed may not fully be exhibited.

  In the above, an advertisement is taken as an example of the content, and the point that the appeal effect cannot always be improved has been described. However, it is not always possible to increase the appeal effect for content other than advertising purposes. In other words, the present invention is not limited to the example of search-linked advertisement delivery as described above. In the technical field of delivering content related to a keyword, appropriate content cannot be delivered unless the keyword is appropriately specified. , It is not always possible to increase the appeal of the content.

  The present application has been made in view of the above, and an object thereof is to provide an extraction device, an extraction method, and an extraction program capable of extracting a keyword that enhances the appeal effect of content.

  The extraction apparatus according to the present application is an extraction model that calculates an evaluation value as a keyword to be linked to content, and extracts the keyword based on the calculated evaluation value, and uses the different methods to calculate the evaluation value. For each of the extraction models to be calculated, a calculation unit that calculates an index value for each of the extraction models based on the degree to which a keyword that meets a predetermined condition is extracted by the extraction model, and an index calculated by the calculation unit An extraction unit for extracting a keyword associated with the content or a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by a value, To do.

  According to one aspect of the embodiment, it is possible to extract a keyword that enhances the appeal effect of the content.

FIG. 1 is a diagram illustrating an example of a web page on which advertisement content according to the embodiment is displayed. FIG. 2 is a diagram illustrating an example of the extraction process according to the embodiment. FIG. 3 is a diagram illustrating a configuration example of the advertising device according to the embodiment. FIG. 4 is a diagram illustrating an example of the first table according to the embodiment. FIG. 5 is a diagram illustrating an example of the second table A according to the embodiment. FIG. 6 is a diagram illustrating an example of the second table B according to the embodiment. FIG. 7 is a diagram illustrating an example of a third table according to the embodiment. FIG. 8 is a diagram illustrating an example of the advertisement information storage unit according to the embodiment. FIG. 9 is a conceptual diagram illustrating an example of extraction processing by the first extraction model according to the embodiment. FIG. 10 is a conceptual diagram illustrating an example of calculation processing using the second extraction model according to the embodiment. FIG. 11 is a conceptual diagram illustrating an example of extraction processing by the second extraction model according to the embodiment. FIG. 12 is a conceptual diagram illustrating an example of extraction processing by the third extraction model according to the embodiment. FIG. 13 is a conceptual diagram illustrating an example of calculation processing by the calculation unit according to the embodiment. FIG. 14 is a flowchart illustrating a calculation processing procedure performed by the advertising device according to the embodiment. FIG. 15 is a flowchart illustrating an extraction processing procedure performed by the advertising device according to the embodiment. FIG. 16 is a diagram illustrating an example of calculation processing by the advertising device according to the modification. FIG. 17 is a hardware configuration diagram illustrating an example of a computer that realizes the function of the advertising device.

  Hereinafter, a mode for carrying out an extraction apparatus, an extraction method, and an extraction program according to the present application (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. Note that the extraction apparatus, the extraction method, and the extraction program according to the present application are not limited to the embodiment. Moreover, in the following embodiment, the same code | symbol is attached | subjected to the same site | part and the overlapping description is abbreviate | omitted.

[1. (Extraction process)
Below, an example of the extraction process which concerns on embodiment is demonstrated using FIG.1 and FIG.2. 1 and 2 show an example in which distribution processing and extraction processing are performed by the advertising device 100 corresponding to the extraction device according to the present application. First, an example of advertisement distribution according to the embodiment will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of a web page on which advertisement content according to the embodiment is displayed.

  FIG. 1 shows a search result page 50 displayed on the user terminal 10 as a web page on which advertisement content according to the embodiment is displayed. The search result page 50 is a web page on which information related to a web page searched based on a search query input from a user is displayed.

  As shown in FIG. 1, the search result page 50 includes a search query input unit 52. A user who uses the search result page 50 inputs a search query to the search query input unit 52 via the user terminal 10. FIG. 1 shows an example in which a search result is displayed on the search result page 50 based on a search query “English conversation classroom” input to the search query input unit 52 by the user. That is, the search result page 50 displays information such as a link of a web page related to “English conversation class” as a search result.

  Further, on the search result page 50, advertisement content is displayed separately from the information of the web page that is the search result. For example, the advertising content is displayed in the advertising space 54 or the advertising space 56 included in the search result page 50. The advertising content includes text that advertises the business of the advertiser. The advertisement content is linked to a web page provided by the advertiser. A web page that is a link destination of such advertising content is called a landing page (hereinafter sometimes referred to as “LP”). For example, the advertiser can display the advertising content to encourage the user to move to LP, browse the web page provided by the advertiser, or purchase the product on the web page. .

  In such an advertisement distribution service linked with a search service, an advertiser bids for a predetermined keyword when or after the advertisement content is submitted. Then, when the user inputs a predetermined keyword as a search query, the advertiser acquires the right to distribute the advertising content linked to the LP to the user. In other words, the advertiser can associate the keyword with the advertisement content by designating the keyword to bid in the search service. Since the link destination of the advertising content is LP, the advertiser can also link the keyword and LP by specifying the keyword to bid. As described above, when the user inputs a keyword that is the same as or similar to the keyword associated with the advertisement content as a search query, the advertiser displays the advertisement content on the search result page 50 in conjunction with the search service. be able to. In other words, the advertiser bids for a keyword for extracting advertisement content. Thereby, since the advertiser can distribute the advertising content at the timing when the desire for specific information is generated in the user, it is possible to distribute the advertising content that can be expected to have a higher advertising effect.

  Here, the flow of the advertisement content distribution process linked to the search service will be described. The distribution process of the advertising content linked to the search service is performed by the advertising device 100, the search server 30, and the user terminal 10 via a network N (not shown) (for example, the Internet).

  The user terminal 10 is an information processing device used by a user. The user terminal 10 is, for example, a mobile terminal such as a smartphone, a tablet terminal, or a PDA (Personal Digital Assistant), a desktop PC (Personal Computer), a notebook PC, or the like.

  The search server 30 is a server device that provides a search service. For example, the search server 30 receives an input of a search query from a user at a predetermined search site. Then, the search server 30 provides the user with a search result based on the search query.

  The advertising device 100 is a server device that distributes advertisements in conjunction with the search server 30. The advertising device 100 accepts bids for keywords as well as submission of advertising content from advertisers. Then, the advertising device 100 distributes the advertising content to the user who has input a keyword (hereinafter, sometimes referred to as “bid keyword”) for which the advertiser has made a bid in the search service according to the procedure described below.

  First, the user accesses a search site provided by the search server 30 via the user terminal 10. Then, the user inputs a search query at the search site. Thereby, the user terminal 10 transmits a search query to the search server 30 (step S01). The search server 30 provides the user with a search result page 50, which is a web page that displays search results corresponding to the search query (step S02).

  In addition, the search server 30 that has received the search query transmits information related to the search query (hereinafter sometimes referred to as “query information”) to the advertising device 100 (step S03). Thereby, advertising device 100 acquires information about a search query inputted by a user. Then, when the search query is a bid keyword, the advertising device 100 selects an advertisement content associated with the bid keyword as a distribution candidate (step S04).

  As described above, the search result page 50 includes the advertising space 54 and the advertising space 56. Therefore, when displaying the search result page 50, the user terminal 10 requests the advertisement apparatus 100 to distribute advertisements in order to acquire the advertisement content displayed in the advertisement frame 54 and the advertisement frame 56 (step S05). Upon receiving the advertisement distribution request, the advertising device 100 distributes the advertisement content associated with the bid keyword to the user terminal 10 (step S06). As a result, for example, the advertising content provided by the advertiser who has bid on the keyword “English conversation classroom” is displayed in the advertising space 54. In addition, the advertising content provided by other advertisers who have bid on the same keyword is displayed in the advertising space 56. As described above, the advertising device 100 can distribute the advertising content to the user in conjunction with the search service.

  In addition, although the search server 30 showed the example which transmits query information to the advertising device 100 in step S03, query information may be transmitted to the advertising device 100 from the user terminal 10. FIG. For example, the query information may be included in a script for an advertisement distribution request on the search result page 50. In this case, the user terminal 10 transmits the query information to the advertising device 100 together with the advertisement distribution request. Then, the advertising device 100 selects the advertising content associated with the query, and distributes the selected advertising content to the user terminal 10.

  By the way, as shown in the example of FIG. 1, in order to display the advertising content on the search result page 50, the advertiser bids on a predetermined keyword. And it is desirable for an advertiser to designate a plurality of appropriate bid keywords in order to distribute advertisement contents to users who have various interests. However, it is not easy for an advertiser to bid on a plurality of appropriate keywords. For example, an advertiser may use a word extracted from an LP provided by the advertiser (that is, a word included in the LP. Hereinafter, it may be referred to as “LP word”), or a product or service provided by the advertiser. It is generally considered to bid for names. In this case, since the advertiser can distribute the advertising content only to the range of the user assumed by the advertiser, it is difficult to efficiently expand the range of the user to be distributed as the advertising content.

  In view of the above circumstances, it is useful to present an appropriate bid keyword to an advertiser in a search-linked advertisement distribution service. Also, the manager of the advertisement distribution service can promote bidding by presenting an appropriate bid keyword to the advertiser, thereby activating the service.

  Therefore, the advertising device 100 according to the embodiment extracts bid keywords suitable for the advertiser in the search-linked advertisement distribution service as shown in FIG. Hereinafter, the extraction process according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram illustrating an example of the extraction process according to the embodiment. In FIG. 2, an example of the extraction process will be described by taking the extraction system 1 as an example.

  As shown in FIG. 2, the extraction system 1 includes an advertiser terminal 20 and an advertising device 100. The advertising device 100 is communicably connected to the advertiser terminal 20 via a network N (for example, the Internet) not shown. Although not shown in FIG. 2, the extraction system 1 may include the user terminal 10 and the search server 30 described above. In this case, the advertising device 100 is connected to the user terminal 10 and the search server 30 via the network N so as to be communicable.

  The advertiser terminal 20 is an information processing device used by the advertiser. The advertiser terminal 20 is, for example, a mobile terminal such as a smartphone, a tablet terminal, or a PDA, a desktop PC, a notebook PC, or the like.

  The advertising device 100 is a server device that extracts an appropriate keyword as a bid keyword for an advertiser to bid by an extraction process described below.

  The advertising device 100 accepts submission of advertising content from a plurality of advertisers in the provided advertisement distribution service. In this case, the advertising device 100 accepts a bid for the keyword designated by the advertiser as well as the submission of the advertising content. Then, the advertising device 100 holds information regarding the accepted bid keyword. In addition, the advertising device 100 acquires, for example, information related to the search query transmitted from the user from the search server 30 and holds the acquired information. Then, the advertising device 100 executes a process of extracting an appropriate bid keyword for an advertiser who uses the advertisement distribution service based on the stored information.

  The advertising device 100 extracts keywords using a plurality of different methods. Here, a series of arithmetic processing used for the keyword extraction method executed by the advertising device 100 is referred to as an “extraction model”. In other words, the advertising device 100 extracts appropriate keywords by performing arithmetic processing on the information held using each extraction model. As will be described in detail later, the advertising device 100 according to the embodiment has three different extraction models. The first extraction model is a technique for extracting an appropriate bid keyword based on information about the LP provided by the advertiser and feature information generated from the LP. The second extraction model is a technique for extracting an appropriate bid keyword based on the co-occurrence of bid keywords actually submitted by the advertiser. The third extraction model is a technique for extracting an appropriate bid keyword based on the tendency of the search query transmitted from the user terminal 10 to the search server 30.

  Each extraction model calculates a predetermined evaluation value for each keyword based on input information, and extracts a keyword that is assumed to be appropriate as a bid keyword based on the calculated evaluation value. Note that the advertising device 100 normalizes the evaluation value calculated by each extraction model by a known method. Thereby, the advertising device 100 can handle and compare evaluation values calculated by different extraction models equally.

  Here, since the advertising device 100 extracts a keyword using a plurality of extraction models, the advertising device 100 learns the extraction model in advance with data for learning (hereinafter may be referred to as “test data”). The plurality of extraction models are weighted (step S11). For example, the advertising device 100 causes each extraction model to learn information related to advertisements received from a plurality of advertisers and information such as a search query transmitted from a user in an advertisement distribution service to be provided. Specifically, the advertising device 100 inputs test data in which a part of information of a bid keyword currently designated by a predetermined advertiser is hidden into an extraction model, and executes an extraction process. Then, the advertising device 100 weights each extraction model depending on whether or not the extraction model can accurately extract the hidden bid keyword. In other words, the advertising device 100 performs learning using a hidden bid keyword as a positive example for the extraction model, and calculates a predetermined index value indicating the weight of each extraction model based on the degree of extraction of the positive example. .

  As described above, the advertising device 100 can execute keyword extraction processing appropriately using a plurality of extraction models by assigning weights to the plurality of extraction models in advance.

  Subsequently, the advertising device 100 accepts submission of advertising content via the advertiser terminal 20 from the advertiser CL10 that is newly using the search-linked advertisement distribution service (step S12). In this case, the advertising device 100 may accept a bid for a keyword for displaying the advertising content on the search result page 50 from the advertiser CL10 together with the submission of the advertising content. In addition, the advertising device 100 acquires information about the LP that is the link destination of the submitted advertising content. In addition, the advertising device 100 acquires, from the search server 30, information related to the search query used for searching for LPs provided by the advertiser CL 10.

  Then, the advertising device 100 extracts keywords to be associated with the submitted advertising content based on the acquired information regarding the advertiser CL10 (step S13). Specifically, the advertising device 100 inputs information related to the advertiser CL10 to each extraction model and causes the extraction process to be executed. Thereby, the advertising device 100 extracts a bid keyword appropriate for the advertiser CL10.

  Then, the advertising device 100 presents the extracted keyword as a bid keyword candidate designated by the advertiser CL10 (step S14). Specifically, the advertising device 100 uses keywords that are expected to increase the probability that advertisement content will be displayed on the search result page 50 when the advertiser CL10 bids in the search-linked advertisement distribution service. Present to CL10.

  As described above, the advertising device 100 according to the embodiment is an extraction model that calculates an evaluation value as a keyword to be associated with content and extracts a keyword based on the calculated evaluation value, and uses different methods. For each extraction model for calculating the evaluation value, an index value for each extraction model is calculated based on the degree to which keywords (for example, keywords that are positive examples in the test data) that match a predetermined condition are extracted by the extraction model. . Then, the advertising device 100 extracts a keyword associated with the content or a bid keyword that is a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by the calculated index value. .

  That is, the advertising device 100 according to the embodiment extracts keywords suitable for the advertiser using a plurality of different extraction models with respect to bid keywords in the search-linked advertisement distribution service. For this reason, according to the advertising device 100, there are various information such as information on an advertiser who is currently using a search-linked advertisement distribution service and information on a search query that is a search request from a user received by the search server 30. Keywords can be extracted based on the information. Thereby, according to the advertising device 100, the advertiser can specify not only the bid keyword specified by the advertiser but also the keyword extracted based on various information as the bid keyword. Then, by bidding on a plurality of appropriate keywords, the advertiser can distribute more advertising content to users who are assumed to be interested in the advertising content to be provided. As a result, according to the advertising device 100 according to the embodiment, the advertising effect of the distributed advertising content can be enhanced.

[2. Configuration of Advertising Device 100]
Next, the configuration of the advertising device 100 according to the embodiment will be described with reference to FIG. FIG. 3 is a diagram illustrating a configuration example of the advertising device 100 according to the embodiment. As illustrated in FIG. 3, the advertising device 100 includes a communication unit 110, a storage unit 120, and a control unit 130.

(About the communication unit 110)
The communication unit 110 is connected to the network N by wire or wireless, and transmits / receives information to / from the user terminal 10, the advertiser terminal 20, and the search server 30. For example, the communication unit 110 is realized by a NIC (Network Interface Card) or the like.

(About the storage unit 120)
The storage unit 120 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. In the example of FIG. 3, the storage unit 120 includes a calculated data storage unit 121 and an advertisement information storage unit 126.

(Regarding the calculated data storage unit 121)
The calculation data storage unit 121 stores information used by each extraction model to calculate an evaluation value of a keyword and information used by the calculation unit 136 described later for calculation processing. As illustrated in FIG. 3, the calculation data storage unit 121 includes data tables such as a first table 122, a second table A 123, a second table B 124, and a third table 125.

(About the first table 122)
Here, FIG. 4 shows an example of the first table 122 according to the embodiment. FIG. 4 is a diagram illustrating an example of the first table 122 according to the embodiment. The first extraction unit 133 that executes processing based on the first extraction model performs keyword extraction based on information stored in the first table 122. As described above, the first extraction model is a method of extracting an appropriate bid keyword based on information about LP provided by the advertiser. For this reason, the first table 122 stores mainly information related to information about the LP among the information acquired by the advertising device 100. In the example illustrated in FIG. 4, the first table 122 includes items such as “bid keyword”, “LP”, “LP word vector”, and “keyword word vector”.

  “Bid keyword” indicates a bid keyword associated with each LP (in other words, each advertisement content). As shown in FIG. 4, since a bid keyword may be designated by a plurality of advertisers, a plurality of LPs may be associated with one bid keyword. The bid keyword is not limited to one word, and may be a combination of a plurality of words.

  “LP” indicates information about LP provided by the advertiser. The notation “LP01” shown in FIG. 4 represents identification information for identifying the LP and conceptually represents information about the LP. The information on the LP stored in the calculated data storage unit 121 may be an HTML (HyperText Markup Language) file or the like constituting the LP, or information indicating a URL (Uniform Resource Locator) where the LP data is placed. It may be. In the following, the information stored in the LP item shown in FIG. 4 may be used as the LP reference code. For example, an LP having information indicated by LP “LP01” may be expressed as “LP01”.

  “LP word vector” indicates a word vector corresponding to each LP. The notation “V01” shown in FIG. 4 is a conceptual representation of a word vector. Note that a word vector corresponding to each LP is generated by a first extraction unit 133 described later. The first extraction unit 133 stores the generated word vector in the first table 122 in association with the LP that is the generation source.

  “Keyword word vector” indicates a word vector corresponding to a bid keyword. The notation “G01” shown in FIG. 4 is a conceptual representation of a word vector. As will be described in detail later, the keyword word vector is generated by the first extraction unit 133 based on the LP word vector associated with the bid keyword.

  For example, in FIG. 4, the bid keyword “K101” is associated with LPs “LP01”, “LP02”, “LP03”,..., And the word vectors corresponding to each LP are LP word vectors. In this example, the word vector corresponding to the bid keyword “K101” is “V01”, “V02”, “V03”,.

(About the second table A123)
Next, FIG. 5 shows an example of the second table A123 according to the embodiment. FIG. 5 is a diagram illustrating an example of the second table A123 according to the embodiment. The second extraction unit 134 that executes processing based on the second extraction model performs keyword extraction based on information stored in the second table A123 and the second table B124. As described above, the second extraction model is a method for extracting an appropriate bid keyword based on the co-occurrence of the bid keyword actually submitted by the advertiser. For this reason, in the second table A123 and the second table B124, information on keywords that are mainly bid by the advertiser among the information acquired by the advertising device 100 is stored. In the example illustrated in FIG. 5, the second table A123 includes items such as “advertiser ID”, “advertisement content”, and “bid keyword”.

  The “advertiser ID” indicates identification information for identifying the advertiser or the advertiser terminal 20. In the following, the identification information stored in the advertiser ID shown in FIG. 5 may be used as an advertiser reference code. For example, the advertiser identified by the advertiser ID “CL21” may be referred to as “advertiser CL21”.

  “Advertising content” indicates the advertising content submitted from the advertiser terminal 20. FIG. 5 shows an example in which conceptual information such as “AD01” and “AD02” is stored in “advertising content”, but in actuality, text data that is advertising content of advertising content, or these Stores the file path name indicating the storage location.

  “Bid keyword” indicates a bid keyword associated with the advertising content. As shown in FIG. 5, the advertiser can bid on a plurality of keywords. For this reason, a plurality of bid keywords are associated with the advertising content.

  For example, in FIG. 5, in order for the advertiser CL21 identified by the advertiser ID “CL21” to display the advertisement content “AD01” on the search result page 50 in the search-linked advertisement distribution service, the bid keyword “K001” is displayed. ”,“ K002 ”,“ K003 ”,...

(About the second table B124)
Next, FIG. 6 shows an example of the second table B124 according to the embodiment. FIG. 6 is a diagram illustrating an example of the second table B124 according to the embodiment. As described above, the second extraction unit 134 performs keyword extraction using the information stored in the second table B124. As will be described in detail later, the second extraction unit 134 creates a combination of keywords for each advertiser who is bidding on the keyword, and calculates a weight value that is an evaluation value of the combination. For this reason, the second table B124 mainly stores information related to keyword combinations. In the example illustrated in FIG. 6, the second table B124 includes items such as “bid keyword (1)”, “bid keyword (2)”, and “weight value”.

  "Bid keyword (1)" and "Bid keyword (2)" correspond to the bid keyword shown in FIG. In the example illustrated in FIG. 6, the bid keyword indicated by the bid keyword (1) is included in the combination together with the bid keyword indicated by the bid keyword (2).

  “Weight value” indicates a weight value of a combination of keywords. The weight value is calculated for each combination by the second extraction unit 134 described later. Since the weight value changes as needed depending on the bid keyword received from the advertiser, the weight value may be updated at a predetermined timing (for example, once a day).

  For example, in FIG. 6, the weight value of the combination (K001, K002) formed by the bid keyword (1) “K001” and the bid keyword (2) “K002” is “0.6”, and the bid keyword (1 ) The weight value of the combination (K001, K003) formed by “K001” and the bid keyword (2) “K003” is “0.3”.

(About the third table 125)
Next, FIG. 7 shows an example of the third table 125 according to the embodiment. FIG. 7 is a diagram illustrating an example of the third table 125 according to the embodiment. The third extraction unit 135 that executes processing based on the third extraction model extracts keywords based on the information stored in the third table 125. As described above, the third extraction model is a method of extracting an appropriate bid keyword based on the tendency of the search query transmitted from the user terminal 10 to the search server 30. Therefore, in the third table 125, information related to the search request transmitted from the user mainly in the search service among the information acquired by the advertising device 100 is stored. In the example illustrated in FIG. 7, the third table 125 includes “advertiser ID”, “LP”, “search query”, “search count”, “re-search query”, “search count”, “trend value”, and the like. Have items.

  “Advertiser ID” corresponds to the item of advertiser ID shown in FIG. “LP” corresponds to the item of LP shown in FIG.

  “Search query” indicates a search request transmitted from the user terminal 10 in the search service. The “re-search query” indicates a search request further transmitted from the user terminal 10 that has transmitted the search query. The “number of searches” indicates the number of times a search query or re-search query has been searched. In other words, the search count indicates the number of times a search query or re-search query is transmitted from the user terminal 10.

  The “trend value” is a value indicating a search tendency such as what kind of content the re-search query is transmitted by the user terminal 10 that has transmitted the search query. The third extraction unit 135 treats the trend value as an evaluation value for the re-search query (ie, keyword). As illustrated in FIG. 7, the third extraction unit 135 calculates a trend value for each combination of a search query and a re-search query associated with LP. Note that the trend value changes from time to time depending on information related to the search query and re-search query acquired from the search server 30, and may be updated at a predetermined timing (for example, once per hour).

  For example, in FIG. 7, the web page indicated by LP “LP31” provided by the advertiser identified by the advertiser ID “CL31” is extracted by transmitting the search query “Diet” from the user terminal 10. This indicates that the web page is currently searched “20281” times by the search query “diet”. Further, the user terminal 10 displaying the LP 31 indicates that a re-search query “exercise”, “health food”, “yogurt”, “esthetic”, and the like have been transmitted. For example, the re-search query “exercise” indicates that the user terminal 10 that has transmitted the search query “diet” has been transmitted “4125” times at the present time. In addition, the user terminal 10 that displayed LP31 as a result of transmitting the search query “diet” has a tendency value “0.2” for transmitting the re-search query “exercise” and the re-search query “health food”. The tendency value to be transmitted is “0.2”, the tendency value to transmit the re-search query “yogurt” is “0.2”, and the tendency value to transmit the re-search query “este” is “0.1”. An example is shown.

(About the advertisement information storage unit 126)
The advertisement information storage unit 126 stores information related to the advertisement content submitted by the advertiser. Here, FIG. 8 illustrates an example of the advertisement information storage unit 126 according to the embodiment. FIG. 8 is a diagram illustrating an example of the advertisement information storage unit 126 according to the embodiment. In the example illustrated in FIG. 8, the advertisement information storage unit 126 includes items such as “advertiser ID”, “advertisement content”, “LP”, and “bid keyword”.

  “Advertiser ID” corresponds to the item of advertiser ID shown in FIG. "Advertising content" corresponds to the item of advertising content shown in FIG. “LP” corresponds to the item of LP shown in FIG. “Bid keyword” corresponds to the item of the bid keyword shown in FIG.

  For example, in FIG. 8, in order for the advertiser CL41 identified by the advertiser ID “CL41” to display the advertisement content “AD41” on the search result page 50 in the search-linked advertisement distribution service, the bid keyword “K401” is displayed. ”,“ K402 ”,“ K403 ”,... In addition, an example is shown in which the link destination of the advertisement content “AD41” is LP “LP41”.

  For example, the distribution unit 141 described later distributes the advertisement content stored in the advertisement information storage unit 126 to the user terminal 10 in accordance with an advertisement acquisition request transmitted from the user terminal 10.

(About the control unit 130)
The control unit 130 includes, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), or the like that stores various programs (an example of an extraction program) stored in a storage device inside the advertising device 100 as a RAM (Random Access Memory). Is implemented as a work area. The control unit 130 is realized by an integrated circuit such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

  As shown in FIG. 3, the control unit 130 includes an acquisition unit 131, an extraction control unit 132, a submission receiving unit 138, a presentation unit 139, a request receiving unit 140, and a distribution unit 141. Implements or executes the information processing functions and operations described in (1). Note that the internal configuration of the control unit 130 is not limited to the configuration illustrated in FIG. 3, and may be another configuration as long as the information processing described below is performed. Further, the connection relationship between the processing units included in the control unit 130 is not limited to the connection relationship illustrated in FIG. 3, and may be another connection relationship.

(About the acquisition unit 131)
The acquisition unit 131 acquires various types of information used for keyword extraction processing. Information acquired by the acquisition unit 131 is used for calculation processing and extraction processing described later. For example, the acquisition unit 131 acquires information related to the advertising content accepted by the submission receiving unit 138 and information related to the advertiser as test data for calculating the weight for each extraction model. In addition, the acquisition unit 131 acquires information related to the search query from the search server 30.

  Then, the acquisition unit 131 sends the acquired information to the extraction control unit 132 in accordance with a request from the extraction control unit 132. As illustrated in FIG. 3, the acquisition unit 131 may store the acquired information in the calculation data storage unit 121. In this case, the extraction control unit 132 uses information stored in the calculated data storage unit 121 for extraction processing and the like.

(About the extraction control unit 132)
The extraction control unit 132 controls keyword extraction processing. As illustrated in FIG. 3, the extraction control unit 132 includes a first extraction unit 133, a second extraction unit 134, a third extraction unit 135, a calculation unit 136, and a fourth extraction unit 137.

(About the first extraction unit 133)
The first extraction unit 133 executes keyword extraction processing based on the first extraction model. The first extraction model is a technique for extracting appropriate bid keywords based on information about LPs provided by advertisers and word vectors generated from LPs. Below, the process in which a 1st extraction model extracts a keyword is demonstrated.

  The first extraction unit 133 generates feature information of LPs and bid keywords acquired by the acquisition unit 131. For example, the first extraction unit 133 may include, as feature information, an LP provided by an advertiser who is currently accepting submission of advertising content, an LP provided by a new advertiser that is the target of the extracted keyword, Each word vector corresponding to the bid keyword currently designated by the advertiser is generated. Then, the first extraction unit 133 stores the generated word vector in the first table 122.

  Note that the word vector corresponding to LP is, for example, a representation of LP by a vector model after giving a predetermined weight to the type and number of appearances of words included in LP. As described above, the first extraction unit 133 generates the feature information of the LP by expressing the tendency of the word appearing in the LP by the word vector.

  Specifically, when generating the word vector corresponding to the LP, the first extraction unit 133 collects the LP data based on the LP URL, and acquires the collected LP data. Then, the first extraction unit 133 performs morphological analysis on the text included in the LP, for example. That is, the 1st extraction part 133 decomposes | disassembles the text contained in LP for every word. Then, the first extraction unit 133 generates a word vector corresponding to the LP based on the type of word included in the LP and the number of appearances of the word in the LP. For example, the first extraction unit 133 generates a word vector corresponding to LP by providing an element “1” for a word that appears in LP and an element “0” for a word that does not appear.

  The first extraction unit 133 also assigns a weight to each word and generates a word vector corresponding to LP. Here, with respect to the weighting of the word vector, for example, a method of extracting a feature word by a known feature extraction algorithm such as tf-idf (Term Frequency-Inverse Document Frequency) can be considered. In this case, for the word extracted as a feature word that characterizes LP, the first extraction unit 133 generates a word vector with a higher weight. Thereby, the 1st extraction part 133 can generate | occur | produce the word vector reflecting the characteristic of LP.

  Note that the first extraction unit 133 may use a technique called tf-ilf (Term Frequency-Inverse Label Frequency) instead of the above-described tf-idf. In the above-described tf-idf, df (Document Frequency) is used as a frequency indicating how many documents are commonly used in all documents (in this case, it corresponds to “document”). What you do is LP). On the other hand, in tf-ilf, a frequency (lf, Label Frequency) indicating how much a word to be processed included in LP is linked to a keyword (that is, a label) is used.

  In this case, for the word vector corresponding to LP, the first extraction unit 133 sums up the number of appearances of all words in the LP, the number of times a predetermined word appears in the LP, and the frequency with which the predetermined word is associated with the keyword. Based on the above, a word vector weighted for each predetermined word is generated. Specifically, the first extraction unit 133 determines the ratio between the total number of appearances of all words in the LP and the number of appearances of the predetermined word in the LP (that is, tf. The predetermined word in the LP). A word vector corresponding to LP is generated based on the frequency of appearance) and the frequency with which the word is associated with the keyword (ie, lf). Thus, the first extraction unit 133 generates a word vector that is appropriately weighted so that the weight of a common word that is very frequently associated with the keyword is not excessively weighted. For this reason, the 1st extraction part 133 can produce | generate the word vector which shows the characteristic of LP more appropriately.

  In addition, the first extraction unit 133 generates a word vector corresponding to the bid keyword. When there are a plurality of LPs associated with the bid keyword, the first extraction unit 133 generates a centroid vector of each word vector corresponding to the plurality of LPs associated with the bid keyword as a word vector corresponding to the bid keyword. To do. This point will be described with reference to FIG.

  FIG. 9 is a conceptual diagram illustrating an example of extraction processing by the first extraction model according to the embodiment. As shown in FIG. 9, LP01, LP02, LP03,... Are associated with the bid keyword K101. The first extraction unit 133 generates word vectors V01, V02, V03,... Corresponding to LP01, LP02, LP03,. And the 1st extraction part 133 produces | generates the gravity center vector of the word vector of each produced | generated LP as a word vector corresponding to the bid keyword K001. That is, the first extraction unit 133 generates the centroid vector of the word vector of each LP as the word vector G01 of the bid keyword K001. Thus, since the 1st extraction part 133 produces | generates the word vector corresponding to a bid keyword using the gravity vector of each word vector of several LP linked, the word which shows the characteristic of a bid keyword appropriately A vector can be generated.

  Next, an example in which the first extraction unit 133 extracts keywords from the LP that is actually the processing target will be described. First, the first extraction unit 133 calculates the similarity between the feature information of the existing LP linked to the keyword and the feature information of the processing target LP including an arbitrary word. Here, the processing target LP refers to an LP provided by an advertiser who desires to extract bid keywords in the advertisement distribution service.

  Since the advertising device 100 accepts bids for a plurality of keywords and submission of a plurality of advertisement contents in the advertisement distribution service, there are a plurality of combinations of bid keywords and existing LPs. At this time, the first extraction unit 133 calculates the similarity of the feature information for each keyword. Taking FIG. 4 as an example, when the bid keyword K101 is the processing target, the first extraction unit 133 includes the feature information of each of the existing LPs LP01, LP02, and LP03 associated with the bid keyword K101. The similarity with the feature information of the processing target LP is calculated. That is, the first extraction unit 133 includes, for each bid keyword, the feature information of all the existing LPs associated with the bid keyword and the processing target for a plurality of combinations associated with the bid keyword and the existing LP. The similarity with the feature information of LP is calculated.

  As described above, the first extraction unit 133 uses a word vector as feature information. For example, the first extraction unit 133 calculates the similarity between the word vector corresponding to the bid keyword and the word vector corresponding to the processing target LP, which is generated based on the existing LP feature information. Specifically, the first extraction unit 133 calculates the cosine similarity between the word vectors as the similarity between the word vector corresponding to the keyword associated with the existing LP and the word vector corresponding to the processing target LP. To do. That is, the 1st extraction part 133 calculates the cosine similarity of two word vectors used as the object of calculation processing. As described above, when the first extraction unit 133 gives an element “1” for a word that appears in LP and gives an element “0” for a word that does not appear, and generates a word vector corresponding to LP The first extraction unit 133 calculates a predetermined numerical value indicated in the range of “0 to 1” as the cosine similarity between the two word vectors.

  In addition, when calculating the degree of similarity, the first extraction unit 133 extracts word vectors that are assumed to have a relatively high degree of similarity, and calculates the degree of similarity between the extracted word vectors. Also good. For example, the first extraction unit 133 extracts a processing target using an algorithm such as Top-k retry (Top-k search). That is, the first extraction unit 133 extracts and extracts the top k word vectors (k is an arbitrary number) of word vectors ordered by a score given by a predetermined method (for example, tf-ilf described above). The similarity between word vectors is calculated. Thereby, since the 1st extraction part 133 skips a process about the process target with a low score, it can calculate a similarity at high speed.

  In addition, when there are a plurality of existing LPs associated with the keyword, the first extraction unit 133 determines the centroid vector of each word vector corresponding to the plurality of existing LPs associated with the keyword, and the processing target The similarity with the word vector corresponding to LP is calculated. Further, the first extraction unit 133 extracts keywords associated with a plurality of existing LPs whose calculated similarity is equal to or greater than a predetermined threshold. Thus, the 1st extraction part 133 averages and grasps the feature of a plurality of existing LPs which other advertisers use as a bid keyword. For this reason, the 1st extraction part 133 can calculate a similarity degree in consideration of the situation where a plurality of existing LPs are associated with keywords.

  As described above, the first extraction unit 133 calculates the degree of similarity between the feature information of the keyword associated with the existing LP and the feature information of the processing target LP, and thus the relationship between the keyword and the processing target LP. Sex can be shown numerically.

  Then, the first extraction unit 133 extracts a keyword corresponding to the feature information whose calculated feature information similarity is equal to or greater than a predetermined threshold. Specifically, the first extraction unit 133 extracts a keyword corresponding to a word vector whose cosine similarity is equal to or greater than a predetermined threshold based on the calculated cosine similarity of the word vector.

  For example, the higher the cosine similarity between the word vectors (that is, the closer to the value of “1”), the more the first extraction unit 133 has a relationship between the keyword corresponding to the calculated word vector and LP. Judge as high. Therefore, the first extraction unit 133 sets a predetermined threshold for the similarity, and extracts a keyword corresponding to the word vector related to the similarity equal to or higher than the predetermined threshold. The first extraction unit 133 can arbitrarily set a predetermined threshold value. For example, the first extraction unit 133 sets a predetermined threshold according to the number of keywords to be extracted. That is, the first extraction unit 133 may set the predetermined threshold value low when extracting a large number of keywords.

  As described above, the first extraction unit 133 according to the embodiment extracts a bid keyword associated with an existing LP having similar characteristics to the LP to be processed. For this reason, according to the 1st extraction part 133, it becomes processing object based on the information and keyword which the other advertiser using the search interlocking type | formula advertisement delivery service is bidding at present, and existing LP. It is possible to extract a keyword associated with the LP.

  In addition, as will be described in detail later, the first extraction unit 133 is extracted based on the test data acquired from the acquisition unit 131 when the calculation unit 136 intends to calculate a weight for the first extraction model. Information relating to the keyword and the similarity that is the evaluation value related to the extracted keyword is sent to the calculation unit 136.

(About the second extraction unit 134)
The second extraction unit 134 executes keyword extraction processing based on the second extraction model. The second extraction model is a technique for extracting an appropriate bid keyword based on the co-occurrence of bid keywords actually submitted by the advertiser. Below, the process in which a 2nd extraction model extracts a keyword is demonstrated.

  The second extraction unit 134 generates a combination of bid keywords based on the information regarding the bid keyword acquired by the acquisition unit 131. Note that the second extraction unit 134 generates a combination of bid keywords for each advertiser who has designated a bid keyword.

  Specifically, as illustrated in FIGS. 5 and 6, the second extraction unit 134 generates a combination (K001, K002) for the keyword K001 bid by the advertiser CL21 and the keyword K002. On the other hand, the second extraction unit 134 does not generate a combination (K001, K011) for the keyword K001 bid by the advertiser CL21 and the keyword K011 bid by the advertiser CL22. However, when the advertiser CL23 bids for the keyword K001 and the keyword K011, the second extraction unit 134 generates a combination (K001, K011). And the 2nd extraction part 134 stores the information regarding the produced | generated combination in 2nd table B124.

  As described above, the second extraction unit 134 does not generate a combination of all the bid keywords, but generates a combination of keywords that are bid for each advertiser. In other words, the second extraction unit 134 generates more combinations as the number of keyword sets that one advertiser is bidding simultaneously increases. For example, as the number of advertisers bidding on the keyword K002 together with the keyword K001 increases, the second extraction unit 134 generates more combinations (K001, K002).

  Then, the 2nd extraction part 134 calculates the evaluation value of each combination of the keyword produced | generated for every designated person who designated the keyword. Specifically, the second extraction unit 134 indicates a combination weight as an evaluation value of a combination of bid keywords, which is a combination generated for each advertiser and is a keyword associated with content provided by the advertiser. Calculate the weight value.

  For example, the second extraction unit 134 calculates a combination weight value according to the number of keyword combinations to be generated. In other words, the second extraction unit 134 calculates a higher weight value because the combination with a larger number of generated keywords is a combination of keywords that many advertisers bid simultaneously.

  In this case, the second extraction unit 134 may calculate a weight value of each combination by learning a linear model with keyword combinations as positive examples. This point will be described with reference to FIG.

  FIG. 10 is a conceptual diagram illustrating an example of calculation processing using the second extraction model according to the embodiment. As shown in FIG. 10, the advertiser CL21 has a bid keyword K001, a bid keyword K002, a bid keyword K003,. Similarly, the advertiser CL22 has a bid keyword K011, a bid keyword K002, a bid keyword K003,. Similarly, the advertiser CL23 has a bid keyword K001, a bid keyword K011, a bid keyword K111,.

  In this case, the second extraction unit 134 generates a combination (K001, K002), a combination (K001, K003), and a combination (K002, K003) based on the bid keyword of the advertiser CL21. Similarly, the second extraction unit 134 generates a combination (K011, K002), a combination (K011, K003), and a combination (K002, K003) based on the bid keyword of the advertiser CL22. Similarly, the second extraction unit 134 generates a combination (K001, K011), a combination (K001, K111), and a combination (K011, K111) based on the bid keyword of the advertiser CL23.

  As described above, the second extraction unit 134 generates a combination of keywords for each advertiser for all advertisers who bid on the advertising device 100 for keywords. Then, the second extraction unit 134 calculates a weight value of each combination by learning a linear model using the keyword combination as a positive example for all keyword combinations. As an example of a linear model, the second extraction unit 134 obtains a regression equation as shown in the following equation (1).

y = a ₁ · x ₁ + a ₂ · x ₂ + ... + a _n · x _n (1)

  In the above formula (1), “y” indicates whether there is an example having each combination. Here, an example having each combination is an advertiser, and the second extraction unit 134 learns by taking an advertiser having a combination of keywords as an example, so that “y” = “1” in principle.

In the above formula (1), “x _n ” (n is an arbitrary number) indicates a combination of keywords. For example, keyword combinations (kw _i , kw _j ) are assigned to “x ₁ ” to “x _n ”. For Figure 10, for example, the combination in _{"x 1"} (K001, K002) are assigned, is assigned a combination in the _{"x 2"} (K001, K003), the combination in the _{"x 3"} (K002, K003 ) is assigned, combining the _{"x 4"} (K011, K002) are allocated, _{"x 5"} combination (K011, K003) is assigned to the combination in the _{"x 6"} (K001, K011) are assigned, combinations (K001, K111) are assigned to the _{"x 7",} a combination (K011, K111) is assigned to _{"x 8".} Here, since the keyword combination is a positive example, for each advertiser, when the keyword combination (kw _i , kw _j ) exists, “x _n” = “1”, otherwise, “ 0 ". In the above formula (1), “a _n ” represents a coefficient of “x ₁ ” to “x _n ”. Here, “a _n ” indicates a weight value of “x ₁ ” to “x _n ” indicating each combination of keywords.

The second extraction unit 134, the weight value a _n of the above formula (1) as a parameter, for example, adjusted by regression analysis method, such as support vector regression using pattern recognition machine, such as of SVM (Support Vector Machine) I do. That is, the second extraction unit 134, for all advertisers represented by the above formula (1), so as to satisfy the above formula (1), the number of advertisers to be a value (sample weight value a _n is , the weight value a _n obtained finds also be) in approximation.

  For example, as shown in FIG. 10, the second extraction unit 134 sets “0.6” as the weight value of the combination (K001, K002), “0.3” as the weight value of the combination (K001, K003), “0.2” or the like is calculated as the weight value of the combination (K002, K003). And the 2nd extraction part 134 matches the calculated weight value with each combination, and stores it in 2nd table B124.

Next, an example in which the second extraction unit 134 extracts keywords for an advertiser that is actually a processing target will be described. Based on the weight value calculated as the evaluation value of the combination, the second extraction unit 134 selects keywords other than the specified bid keyword from the combination including the bid keyword newly specified by the advertiser who uses the advertisement distribution service. Extract. Specifically, it is assumed that the advertiser to be processed has designated the bid keyword “K001”. In this case, the second extraction unit 134, of the generated combination, to compare the weight values of all combinations are (K001, kw _j). Then, the second extraction unit 134 extracts keywords other than the designated bid keyword “K001”, which are keywords included in the combination having the highest weight value. That is, the extracted keyword is a keyword determined by the second extraction unit 134 as the most appropriate keyword to be added by the advertiser to be processed in addition to the bid keyword “K001”.

  In addition, when there are a plurality of bid keywords newly designated by the advertiser, the second extraction unit 134 determines a combination based on the sum of weight values of combinations of the newly designated bid keyword and other keywords. Extract other keywords included in. For example, the second extraction unit 134 selects a plurality of combinations that have the highest sum of weight values of a plurality of combinations including a new bid keyword among combinations generated from keywords already bid by the advertiser. Extract. Then, the second extraction unit 134 extracts keywords that are included in the extracted combinations and that are different from the bid keywords.

  Among the above processes, the second extraction unit 134 may extract a keyword to be presented to the advertiser based on a sum of index values in a combination of a keyword commonly included in each combination and a bid keyword. . For example, the second extraction unit 134 is a combination including a new bid keyword among the combinations generated from keywords already bid by the advertiser, and other keywords included in the combination are common with the bid keyword. The sum of the weight values between the combinations is compared. And the 2nd extraction part 134 extracts the keyword common to each combination based on the sum of the weight value of the compared several combination. Thereby, the 2nd extraction part 134 can extract a common keyword with higher co-occurrence with a bid keyword.

  The extraction process will be described with an example. For example, since the second extraction unit 134 performs the extraction process based on the sum of the weight values, the second extraction unit 134 obtains a function that is generated for each advertiser using the presence / absence of the combination and the weight value of the combination as variables. Then, when the second extraction unit 134 adds a variable corresponding to any combination including the newly specified bid keyword, the second extraction unit 134 selects a bid from the combinations corresponding to the variable that optimizes the value indicated by the function. Another keyword different from the keyword is extracted.

  For example, the second extraction unit 134 obtains a function represented by the following formula (2) as an example of a function generated using the presence / absence of the combination and the weight value of the combination as variables.

Advertiser score = Σ (w_kw _ij ) × (kw _i , kw _j ) (2)

In the above equation (2), “(w_kw _ij )” indicates a weight value corresponding to the combination (kw _i , kw _j ). Further, “(kw _i , kw _j )” indicates whether or not a combination (kw _i , kw _j ) exists. For example, when the advertiser to be processed has a combination (kw _i , kw _j ), “(kw _i , kw _j )” = “1”. On the other hand, when the advertiser does not have the combination (kw _i , kw _j ), “(kw _i , kw _j )” = “0”.

In the above formula (2), it is assumed that the second extraction unit 134 adds a variable corresponding to an arbitrary combination (kw _e , kw _j ) including the bid keyword “(kw _e )”. That is, the second extraction unit 134 multiplies the value “1” indicating that an arbitrary combination (kw _e , kw _j ) exists and the weight value corresponding to the arbitrary combination (kw _e , kw _j ). Variables are added to equation (2) above. In the above example, the larger the value of the added variable, the more the advertiser has a combination including appropriate keywords with co-occurrence with other advertisers. In other words, it can be said that the advertiser's score improved as the advertiser's score increased. And when an advertiser's score becomes the maximum, it can be said that the value which the said function shows was optimized. Therefore, the second extraction unit 134 adds one of the variables corresponding to the combination including the bid keyword “(kw _e )” and the arbitrary keyword “kw _j ” to the above formula (2), and increases the score most. A combination (kw _e , kw _j ) corresponding to the variable to be extracted is extracted. Then, the second extraction unit 134 extracts an arbitrary keyword “kw _j ” included in the combination (kw _e , kw _j ).

  The extraction process of the second extraction unit 134 described above will be specifically described with reference to FIG. FIG. 11 is a conceptual diagram illustrating an example of extraction processing by the second extraction model according to the embodiment. In the example illustrated in FIG. 11, it is assumed that the advertiser CL10 bids by newly specifying the keywords “K001” and “K011”. In addition, the weight value shown in FIG. 10 is used as the weight value of each combination.

  The second extraction unit 134 refers to the weight value of the combination with the arbitrary keyword calculated by the second extraction unit 134 for the bid keyword “K001” newly designated by the advertiser CL10. As shown in FIG. 11, the weight value of the combination (K001, K002) is “0.6”, the weight value of the combination (K001, K003) is “0.3”, and the combination (K001, K111) The weight value is “0.15”. Further, as an example of the weight value of the combination of the bid keyword “K011” newly designated by the advertiser CL10 and an arbitrary keyword, as shown in FIG. 11, the weight value of the combination (K011, K002) is “0”. .35 ”, the weight value of the combination (K011, K003) is“ 0.1 ”, and the weight value of the combination (K011, K111) is“ 0.05 ”.

  Here, the 2nd extraction part 134 calculates | requires the score of advertiser CL10 using the said Formula (2). In this example, the second extraction unit 134 extracts one common keyword as a keyword different from the bid keyword.

  Specifically, the second extraction unit 134 is a combination that newly includes the keywords “K001” and “K011”, and includes a common keyword “K002”. The combination (K001, K002) and the combination ( The variable corresponding to K011, K002) is added to the above equation (2). In this case, the second extraction unit 134 calculates the weight value “0.6” corresponding to the combination (K001, K002) and the weight value “0.35” corresponding to the combination (K011, K002) from the above equation (2). ", The score of the advertiser CL10 is requested to increase by" 0.95 ". Similarly, in the case of the common keyword “K003”, the second extraction unit 134 uses the weight value “0.3” corresponding to the combination (K001, K003) and the weight value corresponding to the combination (K011, K003). Based on “0.1”, the score of advertiser CL10 is requested to increase by “0.4”. Similarly, in the case of the common keyword “K111”, the second extraction unit 134 uses the weight value “0.15” corresponding to the combination (K001, K111) and the weight value corresponding to the combination (K011, K111). Based on “0.05”, the score of advertiser CL10 is requested to increase by “0.2”.

Note that the second extraction unit 134 does not use the combination (K001, K211) and the combination (K011, K311) illustrated in FIG. 11 for the extraction process. In this case, since the arbitrary keyword “kw _j ” is not common in the combination (kw _e , kw _j ), the second extraction unit 134 cannot compare the weight values of the combinations.

  Then, based on the above result, the second extraction unit 134 extracts the keyword “K002” that is commonly included in the combination that optimizes the advertiser's score. That is, the second extraction unit 134 extracts the keyword “K002” as an appropriate bid keyword to be added to the bid keywords “K001” and “K011” newly specified by the advertiser CL10.

  As described above, the second extraction unit 134 according to the embodiment uses a combination of keywords generated for each advertiser who actually bids for a keyword in the advertisement distribution service for the extraction process. For this reason, according to the 2nd extraction part 134, a new keyword can be extracted, after also including in the parameter | index the relevance of the keywords which an advertiser designates. For example, in a search-linked advertisement distribution service, even if a keyword that is easy to be input as a search query is presented to an advertiser, there is a discrepancy between the user's intention and the contents of the advertisement content if the keyword is not relevant to the advertisement content Therefore, there is a possibility that the effect of the advertising content is not sufficiently exhibited. On the other hand, the second extraction unit 134 extracts keywords by using a combination of keywords currently being bid by other advertisers. That is, the extracted keyword has co-occurrence with the keyword bid by other advertisers. As a result, the advertiser to be processed is a keyword extracted by the advertising device 100, and a keyword bid by another advertiser highly relevant to itself can be appropriately added to the bid keyword. it can.

  Similarly to the first extraction unit 133, the second extraction unit 134, based on the test data acquired from the acquisition unit 131, when the calculation unit 136 intends to calculate weights for the second extraction model. The extracted keyword and the weight value, which is the evaluation value related to the extracted keyword, or the information related to the advertiser's score described above are sent to the calculation unit 136.

(About the third extraction unit 135)
The third extraction unit 135 executes keyword extraction processing based on the third extraction model. The third extraction model is a technique for extracting an appropriate bid keyword based on the tendency of the search query transmitted from the user terminal 10 to the search server 30. Below, the process in which a 3rd extraction model extracts a keyword is demonstrated.

  The third extraction unit 135 uses information related to the search query and re-search query acquired by the acquisition unit 131 for processing. For example, the third extraction unit 135 determines the number of times that the search query and the re-search query are transmitted from the user terminal 10 to the search server 30, LPs and advertisement content corresponding to the search query and the re-search query, the search query and the re-search Information for identifying the user terminal 10 that transmitted the query is used for the processing.

  Then, the third extraction unit 135 calculates an evaluation value indicating the relevance between the search query and the re-search query. Specifically, in the search service, the third extraction unit 135 obtains a tendency value that is a value indicating a search tendency such as what kind of content the re-search query is transmitted by the user terminal 10 that has transmitted the search query. Calculate.

  For example, the third extraction unit 135 calculates a tendency value, which is an evaluation value indicating the relevance between the search query and the re-search query, based on the number of times the re-search query is transmitted after the search query is transmitted. That is, the third extraction unit 135 assumes that the re-search query acquired by the acquisition unit 131 is more likely to be transmitted after the search query is transmitted, and the re-search query tendency value for the search query is more likely to be transmitted. To be higher.

  Note that the third extraction unit 135 indicates the relationship between the search query and the re-search query further transmitted from the user terminal 10 that acquired the content (for example, LP or advertisement content) extracted based on the search query. The trend value that is the evaluation value is calculated. In other words, the third extraction unit 135 does not calculate the tendency value between the search query and the re-search query transmitted from any user terminal 10, but transmits from the user terminal 10 that acquired the content searched by the search query. Calculate the trend value with the re-search query. Thereby, since the 3rd extraction part 135 calculates the tendency value of the search query and re-search query corresponding to the content searched by the search service, it can calculate the search tendency in this content clearly.

  Further, the third extraction unit 135 calculates a tendency value between the search query and the re-search query for each content extracted based on the search query. For example, when the user terminal 10 transmits a search query to the search service, a plurality of contents can be extracted as search results. In this case, the user terminal 10 acquires content selected from a plurality of contents (for example, content clicked on a link destination in the search result page 50). And the 3rd extraction part 135 calculates the tendency value of a search query and a re-search query for every content selected by the user. Accordingly, the third extraction unit 135 can clearly calculate a search tendency in specific content even when a plurality of contents are extracted by the search service.

  As described above, the third extraction unit 135 calculates the tendency value based on the number of re-search queries transmitted for each predetermined LP extracted as a result of the transmission of the search query to the search server 30. As an example, the third extraction unit 135 calculates, as a trend value, a numerical value with the total number of search queries transmitted to display a predetermined LP as a denominator and the number of each re-search query transmitted as a numerator. . In this case, the calculated tendency value is expressed by a numerical value of “0” to “1”. Alternatively, the third extraction unit 135 may determine how much the re-search query is related to the search query based on, for example, the cosine ( You may calculate by the similarity expressed by cosine. Even in this case, the calculated tendency value is expressed by a numerical value of “0” to “1”. Or the 3rd extraction part 135 may treat the number by which the re-search query was transmitted as a tendency value as it is.

  Then, the third extracting unit 135 stores the calculated tendency value in the third table 125 in association with the content. In addition, when the information regarding the search query and the re-search query acquired from the search server 30 by the acquisition unit 131 changes at any time, the third extraction unit 135 determines the tendency value stored in the third table 125 as a predetermined value, for example. You may make it update at the timing of.

  And the 3rd extraction part 135 extracts the keyword linked | related with the predetermined content extracted based on a search query from a re-search query based on the calculated evaluation value. Specifically, the third extraction unit 135 extracts a re-search query corresponding to a numerical value having a higher calculated tendency value as an example of a bid keyword that is a keyword associated with the advertisement content or LP in the advertisement distribution service. To do. In other words, the third extraction unit 135 extracts bid keyword candidates to be presented to the advertiser.

  In addition, the 3rd extraction part 135 was acquired by the user terminal 10, when calculating the tendency value with the re-search query further transmitted from the user terminal 10 which acquired the content extracted based on the search query Extract keywords associated with content. That is, the third extraction unit 135 extracts a keyword by associating the calculated tendency value with the content used for calculating the tendency value.

  The extraction process of the third extraction unit 135 will be specifically described with reference to FIG. FIG. 12 is a conceptual diagram illustrating an example of extraction processing by the third extraction model according to the embodiment. Note that the LP, search query, and re-search query exemplified in FIG. 12 correspond to the LP, search query, and re-search query stored in the third table 125 shown in FIG. FIG. 12 illustrates an example in which the user terminal 10 that has acquired the LP 31 extracted by the search query “Diet” transmits a re-search query.

  First, the user terminal 10 transmits a search query “diet” in the search service. As a result, a link to the LP 31 is displayed on the search result page 50 as a search result of advertisement content or a web page with the LP 31 as a link destination. The user terminal 10 acquires the LP 31 by clicking on the advertisement content or the link and displays it on the display screen of the user terminal 10.

  Thereafter, the user terminal 10 uses the search service again. In this case, the user terminal 10 transmits a re-search query in the search service. In FIG. 12, the re-search query transmitted by the user terminal 10 that has searched for the search query “Diet” is indicated by a circle arranged around the circle labeled “Diet”. In the circle shown in FIG. 12, a word written in the circle indicates a re-search query. The numerical value described in a circle shows a tendency value. That is, FIG. 12 shows what words the user terminal 10 after searching for “diet” tends to transmit as a re-search query. In the example illustrated in FIG. 12, as the re-search query of the search query “diet”, “exercise”, “health food”, and “yogurt” have a tendency value “0.2”, which is higher than other re-search queries. Numerical values are shown.

  In this case, the third extraction unit 135 may extract all of “exercise”, “health food”, and “yogurt” as keywords to be linked to the LP 31, and calculated from these in more detail. Only one keyword may be extracted based on the trend value.

  Note that the user terminal 10 that has transmitted the re-search query “yogurt” may acquire the LP 32 extracted by the re-search query “yogurt”. And the user terminal 10 which acquired LP32 may transmit a re-search query further. In this case, “yogurt” is handled as a search query for LP32. That is, the third extraction unit 135 extracts a keyword associated with the LP 32 based on the tendency value corresponding to the re-search query transmitted from the user terminal 10 that acquired the LP 32 extracted by the search query “yogurt”. .

  In the example illustrated in FIG. 12, as a re-search query for the search query “yogurt”, “beautiful bowel” has a tendency value “0.4”, which is higher than other re-search queries. In this case, the third extraction unit 135 extracts the re-search query “beautiful intestine” as a keyword associated with the LP 32.

  Note that the third extraction unit 135 may not extract the re-search query “beautiful bowel” as a keyword associated with the LP 31. In other words, the third extraction unit 135 may not be used for the extraction process even if it is a re-search query transmitted later by the user terminal 10 that transmitted the search query. In other words, the third extraction unit 135 may treat only the search query transmitted immediately after the user terminal 10 that transmitted the search query as the re-search query. Accordingly, the third extraction unit 135 can prevent a re-search query that is assumed to be less relevant due to a chain of search queries from being extracted as a keyword.

  Thus, the 3rd extraction part 135 which concerns on embodiment extracts a keyword based on the query which a user actually inputs in an advertisement delivery service. In general, users who use the search service cannot obtain the information they want in the first search, or want to collect more information, so they have the same intention as the first search but use different queries. Perform a second search. From this, it can be said that the search query and the re-search query have a high probability of being words reflecting the same intention of the user. Therefore, the third extraction unit 135 obtains a keyword that is included in the re-search query and is difficult for the advertiser to recognize by obtaining a search tendency for the re-search query after the search query based on the above relationship. It is possible to extract. In other words, the third extraction unit 135 can extract not only a keyword that the advertiser himself assumes as a bid keyword but also a keyword reflecting an actual user's action history.

  Similarly to the first extraction unit 133, the third extraction unit 135 is based on the test data acquired from the acquisition unit 131 when the calculation unit 136 is to calculate the weight for the third extraction model. Information about the extracted keyword and a tendency value, which is an evaluation value related to the extracted keyword, is sent to the calculation unit 136.

(About the calculation unit 136)
For each extraction model that extracts keywords to be associated with content using different methods, the calculation unit 136 calculates an index value of the extraction model based on the degree of extraction of keywords that meet a predetermined condition by the extraction model. calculate. For example, as the predetermined condition, the calculation unit 136 may determine that the keyword extracted by the extraction model is actually associated with the advertisement content or LP, or a keyword (for example, artificial In addition, the index value is calculated using the fact that the predetermined keyword is determined to be a positive example in advance.

  Hereinafter, a method in which the calculation unit 136 calculates the weight for each extraction model will be described. For example, in the advertisement distribution service provided by the advertising device 100, the advertisement content from 100,000 advertisers (hereinafter, when the advertiser is counted, the unit of counting the advertiser is referred to as “account”). Is accepted. That is, it is assumed that the acquisition unit 131 has acquired information on advertisers for 100,000 accounts.

  First, the calculation unit 136 learns each extraction model using, for example, information of 80,000 accounts among the acquired 100,000 accounts. That is, the calculation unit 136 inputs information related to advertisers for 80,000 accounts to the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135. Then, the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 try to extract a keyword for each account by the extraction method of each extraction model based on the input information. Then, the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 use the bid keyword actually designated by the advertiser as a positive example, and extract the bid keyword actually designated by the advertiser. If possible, adjust the keyword rating value to be higher, and if the extracted keyword is not actually a bid keyword specified by the advertiser, the extracted keyword rating value will be lower The adjustment is repeated so that the extraction model is learned. Thus, each extraction model is learned and generated based on the common test data that is the same learning data.

  Then, the calculation unit 136 calculates a weighting index value to be assigned to each extraction model using information of 20,000 accounts among the 100,000 accounts that have been acquired. For example, the calculation unit 136 inputs, to the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135, information about 80% of the keywords bid by the advertiser in the information about the advertiser of 20,000 accounts. To do. Then, the calculation unit 136 calculates the degree to which each extraction model can accurately extract the hidden 20% bid keywords for each input account. In other words, the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 further add a bid keyword to be added to each account based on information corresponding to 80% of the keywords that the advertiser is actually bidding on. Extract. Then, the calculation unit 136 calculates the degree of whether or not the keywords extracted by the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 are bid keywords that are actually specified for each account. To do. For example, the calculation unit 136 calculates an index value that gives a higher weight to an extraction model that calculates a higher evaluation value for a bid keyword actually specified for each account.

  The calculation unit 136 calculates a weighting index value of each extraction model based on a numerical value obtained by normalizing the evaluation value of the keyword calculated by each extraction model when extracting the keyword. The calculation processing flow of the calculation unit 136 will be specifically described with reference to FIG. FIG. 13 is a conceptual diagram illustrating an example of calculation processing performed by the calculation unit 136 according to the embodiment.

  First, the calculation unit 136 inputs 20,000 account data in which 20% of bid keywords are hidden as test data to the first extraction model (step S21). The first extraction unit 133 calculates the above-described similarity based on information about the LP of each account, and extracts keywords. Similarly, the calculation unit 136 inputs test data to the second extraction model (step S22). The second extraction unit 134 calculates the above-described weight value based on information related to the combination of bid keywords for each account, and extracts keywords. Similarly, the calculation unit 136 inputs test data to the third extraction model (step S23). The third extraction unit 135 calculates the above-described tendency value based on information on the search query and the re-search query for the LP of each account, and extracts keywords.

  Then, the calculation unit 136 normalizes the similarity, the weight value, and the trend value, which are evaluation values calculated by each extraction model, using a known method (step S24). The calculation unit 136 does not use the similarity, weight value, or trend value itself as an evaluation value for each extraction model, but calculates a score for a keyword determined based on the similarity, weight value, or trend value. If so, the score may be normalized. Thereby, the calculation part 136 can compare the evaluation value which each extraction model calculates.

  Then, the calculation unit 136 weights each extraction model by calculating an index value indicating weighting based on the degree to which positive examples can be extracted from the test data as described above (step S25). ). For example, the calculation unit 136 calculates weights represented by the following formula (3).

    Keyword score = α · X + β · Y + γ · Z (3)

  In the above formula (3), “keyword score” indicates the score of the extracted keyword. For example, the score corresponds to an evaluation value calculated by each extraction model. In the above formula (3), “α” represents the weight of the first extraction model, and “X” represents the keyword score calculation method based on the first extraction model. Similarly, “β” indicates the weight of the second extraction model, and “Y” indicates a keyword score calculation method based on the second extraction model. Similarly, “γ” represents the weight of the third extraction model, and “Z” represents a keyword score calculation method based on the third extraction model. “Α”, “β”, and “γ” can dynamically change in accordance with the above-described weighting calculation process. The sum of “α”, “β”, and “γ” is “1”. For example, calculating a keyword score with a single extraction model without using a plurality of extraction models means that any one of “α”, “β”, and “γ” is “1”, and the remaining weight is It is the same as being “0”.

  In the above description, the calculation unit 136 has shown an example in which the index value for weighting is calculated using the bid keyword actually designated by the advertiser as a positive example. However, the calculation unit 136 may not use a bid keyword actually designated by the advertiser as a positive example but may use a keyword determined by a predetermined agreement as a positive example. For example, the calculation unit 136 accepts in advance a setting such that a keyword assumed to be appropriate for each account is determined as a positive example by human judgment. Then, the calculation unit 136 may calculate an index value for weighting so that the value becomes higher with respect to the extraction model that can extract the keyword that is determined as a positive example by the agreement.

  In this way, the calculation unit 136 generates an arithmetic expression for extracting a keyword using a plurality of extraction models by assigning an appropriate weight to each extraction model. The fourth extraction unit 137, which will be described later, can extract an appropriate bid keyword based on the keyword score calculated by the arithmetic expression as shown in the above formula (3).

(About the fourth extraction unit 137)
The fourth extraction unit 137 extracts a keyword associated with the content or a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by the index value calculated by the calculation unit 136. . Specifically, the fourth extraction unit 137 receives a plurality of extraction models based on information related to the advertisement content and the advertiser when receiving the submission of the advertisement content from an advertiser who newly uses the advertisement distribution service. The bid keyword suitable for the new advertiser is extracted by the extraction processing method using. In other words, the fourth extraction unit 137 extracts a keyword as a candidate for a new advertiser to bid.

  For example, the fourth extraction unit 137 extracts a keyword based on the score calculated by the arithmetic expression such as the above expression (3) based on the information related to the account to be processed. That is, the fourth extraction unit 137 adds the weight corresponding to the index value calculated by the calculation unit 136, which is an evaluation value for each keyword calculated by the plurality of extraction models, as shown in the above formula (3). Based on the sum of the evaluated values, a keyword associated with the advertising content or a keyword for extracting the advertising content and LP is extracted.

(About the submission reception part 138)
The submission receiving unit 138 receives submission of advertisement content. In addition, the submission receiving unit 138 receives bid keywords specified by the advertiser as well as the submission of the advertising content. Then, the submission receiving unit 138 stores the received advertisement content and bid keyword in the calculated data storage unit 121.

(About the presentation unit 139)
The presentation unit 139 presents the keywords extracted by the fourth extraction unit 137. Specifically, the presenting unit 139 presents the keyword extracted by the fourth extracting unit 137 to the advertiser as a bid keyword associated with the advertising content in the advertisement distribution service.

  Note that the presentation unit 139 may not present the keyword as a bid keyword to the advertiser, even if the keyword is extracted by the fourth extraction unit 137. For example, the presentation unit 139 does not have to present the excluded keyword that can be set in the advertising content. Examples of the exclusion keyword that can be set in the advertisement content include a company name that competes with an advertiser who is a provider of the advertisement content, and a product name handled by the competing company.

(About the request reception unit 140)
The request reception unit 140 receives a request regarding distribution of advertisement content. For example, the request reception unit 140 receives an HTTP (Hypertext Transfer Protocol) request included in a web page displayed by the user terminal 10 as an advertisement content acquisition request.

(About distribution unit 141)
The distribution unit 141 distributes advertising content. For example, the distribution unit 141 distributes the advertising content to the user terminal 10 in accordance with the HTTP request received by the request reception unit 140. In addition, when the request reception unit 140 receives an advertisement content acquisition request, the distribution unit 141 selects the distribution candidate advertisement content from the advertisement information storage unit 126.

  Here, when the search query received by the request receiving unit 140 and the bid keyword are the same or similar, the distribution unit 141 selects the advertisement content associated with the bid keyword as a distribution candidate. Then, the distribution unit 141 distributes the selected advertisement content to the user terminal 10 that has transmitted the search query to the search server 30.

[3. Processing procedure)
Next, the procedure of the calculation process executed by the advertising device 100 according to the embodiment will be described with reference to FIG. FIG. 14 is a flowchart illustrating a calculation processing procedure by the advertising device 100 according to the embodiment.

  As illustrated in FIG. 14, the advertising device 100 acquires test data for learning an extraction model including information on an advertiser who already uses an advertisement distribution service, for example (step S101). Then, the advertising device 100 calculates an evaluation value for each account included in the acquired test data using each extraction model (step S102).

  Then, the advertising device 100 normalizes the evaluation value calculated by each extraction model (step S103). The advertising device 100 calculates the weight of each extraction model based on the normalized evaluation value and the degree to which the extracted keyword is a positive example (step S104).

  Next, the procedure of the extraction process executed by the advertising device 100 according to the embodiment will be described with reference to FIG. FIG. 15 is a flowchart illustrating an extraction processing procedure by the advertising device 100 according to the embodiment.

  As shown in FIG. 15, the advertising device 100 determines whether or not a submission of advertising content has been received from an advertiser who newly uses the advertising distribution service (step S201). When the advertising device 100 has not accepted the submission of the advertising content (step S201; No), the advertising device 100 waits until it is accepted.

  On the other hand, when the advertising device 100 accepts the submission of the advertising content (step S201; Yes), the advertising device 100 extracts a keyword by using a plurality of extraction models based on the received information on the advertiser or the like ( Step S202). Then, the advertising device 100 presents the extracted keyword as a bid keyword to the advertiser who has submitted the advertising content (step S203).

  In FIG. 15, the advertising device 100 may extract a different keyword after presenting the keyword to the advertiser in step S203. That is, the advertising device 100 may repeatedly execute Step S202 and Step S203 depending on the advertiser's correspondence to the extracted keyword. For example, when the advertiser desires a keyword other than the keyword presented, the advertising device 100 further extracts a keyword other than the keyword previously extracted as the bid keyword. Then, the advertising device 100 presents the newly extracted keyword to the advertiser.

[4. (Modification)
The embodiment described above may be implemented in various different forms other than the above embodiment. Accordingly, other embodiments will be described below.

[4-1. (Dynamic weighting)
In the embodiment described above, the advertising device 100 has shown an example in which the weight of each extraction model is given based on the normalized evaluation value and the degree of extraction of a positive example keyword in the test data. Here, the advertising device 100 may dynamically change the weight of each extraction model in conjunction with the extraction process. This point will be described with reference to FIG.

  FIG. 16 is a diagram illustrating an example of calculation processing by the advertising device 100 according to the modification. In FIG. 16, “α” and “β” in the above equation (3) are shown by a two-dimensional graph. Here, since the sum of “α”, “β”, and “γ” is “1” as described above, coordinates (α, β) = (1, 0), ( When one point (α, β) is defined inside a triangle surrounded by three points (0, 0) and (0, 1), the values that (α, β, γ) can take are uniquely determined. . For example, when the point P1 (α, β) = (0.4, 0.3) shown in FIG. 16 is determined, “γ” = 0.3. It is assumed that the above weights (α, β, γ) are determined by calculation processing by the calculation unit 136.

  Here, it is assumed that the advertising device 100 performs keyword extraction processing on an advertiser who newly uses the advertisement distribution service, using the weighting determined above. Then, the advertising device 100 presents the extracted keyword to the advertiser. It is assumed that the advertiser designates the presented keyword as a bid keyword. Thereafter, the advertising device 100 acquires an index value indicating the effect as a bid keyword for the extracted keyword. For example, the advertising device 100 displays the number of times a keyword associated with the advertising content is input from the user as a search query for searching for advertising content, the number of times the advertising content is selected by the user, or the advertising content. Information on the profits obtained by the advertiser.

  In other words, the advertising device 100 acquires an index value indicating an improvement in some advertising effect on the advertising content by designating an actually extracted keyword as a bid keyword. Then, for example, the advertising device 100 adjusts the extraction model that extracts the keyword that has the most improvement in the advertising effect so as to increase the weight. For example, when the advertising device 100 acquires a predetermined index value indicating that the advertising effect of the advertising content is improved by the keyword extracted by the first extraction model, the numerical value “α” indicating the weight of the first extraction model To increase. For example, when the advertising device 100 sets the weight in increments of “0.1”, “α” is adjusted from “0.4” to “0.5”. In this case, as shown in FIG. 16, the point P1 moves to the point P5. That is, the weight (α, β, γ) is adjusted to (0.5, 0.3, 0.2).

  Then, the advertising device 100 uses the newly calculated weight (α, β, γ) = (0.5, 0.3, 0.2) to the advertiser who newly uses the advertisement distribution service. To extract keywords. By repeating this process, the advertising device 100 can dynamically calculate the weight assumed to increase the advertising effect for each extraction model. For example, since the advertising device 100 adjusts the weight according to the result of specifying the extracted keyword as a bid keyword, in the example of FIG. 16, the point P1 indicating the weight is the point P2, the point P3, Adjust the weight to be moved to P4.

  As described above, the advertising device 100 can dynamically adjust the weight given to each extraction model based on the result such as the advertising effect. Thereby, the advertising device 100 can execute the extraction process using each extraction model to which an appropriate weight is given in accordance with the actual situation.

[4-2. Information used for extraction)
In the above-described embodiment, the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 are based on the bid keyword related to the advertiser who has already submitted the advertising content, or information on the search query. An example of extracting is shown. Here, the 1st extraction part 133, the 2nd extraction part 134, and the 3rd extraction part 135 may extract a keyword not only using an existing bid keyword etc. but using other information.

  For example, as described above, the first extraction unit 133 generates a word vector corresponding to the existing LP, but the first extraction unit 133 may generate a word vector corresponding to the advertising content.

  That is, the first extraction unit 133 performs morphological analysis on text data included in the advertisement content, not the existing LP. Then, the first extraction unit 133 generates a word vector corresponding to the advertising content. In addition, the first extraction unit 133 extracts a keyword by generating a word vector corresponding to a bid keyword associated with the advertising content (that is, a keyword associated with LP) based on the word vector corresponding to the advertising content. May be.

  Further, as described above, the second extraction unit 134 generates a combination of keywords based on the existing bid keywords. However, the second extraction unit 134 determines the combination of keywords based on the advertisement content and information about the LP. It may be generated. For example, the second extraction unit 134 performs morphological analysis on the text data included in the advertising content or LP, thereby extracting a keyword that can be an element for generating a combination from the text data included in the advertising content or LP. Then, the second extraction unit 134 generates a combination of extracted keywords. And the 2nd extraction part 134 extracts the keyword linked | related with a predetermined content based on the produced | generated combination.

  Note that the keyword included in the advertising content or LP is preferably a keyword that characterizes the advertising content or LP. In this case, the second extraction unit 134 determines, for example, words that are included in the advertising content or the text data of the LP, or words that are unlikely to appear in other advertising content or LP, as keywords that characterize the advertising content or LP. Then, such keywords are extracted.

  Thus, according to the 1st extraction part 133 and the 2nd extraction part 134, it is not restricted to the existing bid keyword, By using advertising content and LP, the suitable keyword which an advertiser bids can be extracted. it can.

  Moreover, the 1st extraction part 133 and the 2nd extraction part 134 may perform said extraction process using the parameter | index which shows the advertising effect of advertisement content. For example, in the combination weight value calculation process, the second extraction unit 134 may calculate the combination weight value by increasing the weight for an advertiser providing advertisement content having a high CTR (Click Through Rate). . For example, the second extraction unit 134 may adjust the value of “y” for each advertiser in the above equation (1). In this case, the second extraction unit 134 increases the value of “y” for the advertiser that provides the advertising content with a high CTR, and consequently increases the weight value of each combination that the advertiser has. Can do. Thereby, the 2nd extraction part 134 becomes easy to extract the keyword which the advertiser who provides advertising content with high CTR specifies. Or the 2nd extraction part 134 may add a predetermined value to the weight value calculated with respect to the combination which the advertiser who provides advertising content with high CTR has.

  In addition, the 1st extraction part 133 and the 2nd extraction part 134 may calculate the evaluation value of a keyword based on the parameter | index which shows advertising effects other than CTR. For example, the first extraction unit 133 and the second extraction unit 134 may use a CVR (Conversion Rate) indicating a percentage of conversion to the number of times the advertising content is displayed, or advertising content in the user terminal 10. The actual number of impressions (number of impressions), CPC (Cost Per Click), which is the amount charged for clicks on advertising content, and eCPM (effective cost per mill), which is the amount charged per 1000 times of advertising content The evaluation value may be calculated using information related to the advertisement content.

  Moreover, the 3rd extraction part 135 may calculate a tendency value not only based on the user terminal 10 which acquired LP but based on the frequency | count that the re-search query was transmitted from the user terminal 10 which acquired advertising content. In addition, when the user selects the advertisement content, the third extraction unit 135 is transmitted from the user terminal 10 corresponding to the user, assuming that the user is more interested in the advertiser than the other users. A predetermined weight value may be given to the re-search query. In the above embodiment, an example in which the search query is an LP word or a bid keyword has been shown. However, the search query may be a word included in the advertisement content.

  As described above, the first extraction unit 133, the second extraction unit 134, and the third extraction unit 135 are not limited to the examples shown in the above embodiment, and can extract keywords using various information.

[4-3. Advertiser)
In the above embodiment, the advertiser has shown an example in which the advertising content is submitted to the advertising device 100 using the advertiser terminal 20. However, the advertiser may request such processing from the agency without submitting the advertising content to the advertising device 100 itself. In this case, it is an agency that performs the processing related to the advertising device 100. That is, the notation “advertiser” is a concept including not only the advertiser but also the agency, and the notation “advertiser terminal” includes not only the advertiser terminal 20 but also the agency terminal used by the agency. It is assumed to be a concept.

[4-4. (Advertisement delivery)
In addition, when there are a plurality of advertisement contents associated with the same bid keyword in the distribution of the advertisement content, the distribution unit 141 may be an advertiser with a high bid price of the bid keyword, an advertisement content with a high CTR, a bid price and a CTR. It is also possible to preferentially select advertising content that is both high. Then, the distribution unit 141 distributes the selected advertisement content to the user terminal 10.

[4-5. Example of re-search query)
In the above embodiment, as an example of a re-search query used by the third extraction unit 135, the search server 30 is a user terminal 10 that has transmitted a search query, and is transmitted again from the user terminal 10 that has acquired the LP. An example of handling a query as a re-search query was shown. However, the search server 30 may treat a search query transmitted in another situation as a re-search query.

  For example, after the search query is transmitted from the user terminal 10, the search server 30 may treat a search query transmitted within a predetermined time (for example, within 10 minutes) as a re-search query. Further, the search server 30 is transmitted to the search server 30 by being newly input to the search query input unit 52 immediately after displaying the search result page 50 in the user terminal 10 that has acquired the search result page 50. May be treated as a re-search query. In this case, the search server 30 may not determine whether or not the user has browsed the LP extracted by the search query. Note that even if the search query is transmitted again from the user terminal 10 that acquired the LP, the search server 30 is outside the predetermined time after the search query is transmitted from the user terminal 10 (for example, transmission of the previous search query). The search query transmitted after 24 hours) may not be treated as a re-search query.

  Further, even when the user terminal 10 that has transmitted the search query does not click the advertisement content on the search result page 50 and clicks a link of another web page, the search server 30 then returns to the search result page. You may treat the search query transmitted from the user terminal 10 which displayed 50 as a re-search query. As described above, the search server 30 can handle search queries before and after it is assumed that a user is searching with the same search intention as a search query and a re-search query.

  In addition, the advertising device 100 may receive query information from the user terminal 10. For example, the advertising device 100 acquires information related to the search query based on a request included in the advertisement acquisition request transmitted when the user terminal 10 displays the search result page 50. In this case, after the search query is transmitted from the user terminal 10, the third extraction unit 135 according to the advertising device 100 uses a new search query transmitted from the user terminal 10 within a predetermined time as a new search query. You may calculate the tendency value of a search query and a re-search query. The identity of the user terminal 10 is identified based on, for example, terminal identification information included in a cookie transmitted from the user terminal 10 or the like.

[4-6. Other uses)
In the above embodiment, the advertising device 100 has shown an example in which bid keywords are extracted in a search-linked advertisement distribution service. However, the advertising device 100 may be used for an advertisement distribution service other than the search-linked type as described above. For example, the advertising device 100 may be used for a so-called advertising content matching type advertising service that accepts bids for keywords and distributes the advertising content to web pages having high affinity with the keywords.

  Moreover, in the said embodiment, the advertisement content and LP were mentioned as an example of content, and the bid keyword was mentioned as an example as an example of a keyword. However, the advertising device 100 may be used not only for extracting bid keywords associated with advertising content and LPs but also for other purposes.

  For example, when the extraction device according to the present application corresponds to the search server 30, the search server 30 may display a keyword that assists the search query on the search result page 50. For example, the search server 30 extracts a keyword corresponding to a predetermined LP that is a search target. Then, when the user uses the search service, the search server 30, together with the search query, when the search query input by the user and the extracted keyword are related (for example, when they are similar in meaning). The extracted keyword may be presented to the user. For example, the search server 30 extracts the extracted keyword as a query for searching for a predetermined web page instead of the search query previously input by the user. Then, the search server 30 may present the keyword extracted together with the search result on the search result page 50. The user may be able to efficiently search for a target web page by executing a search based on the presented keyword. In this case, the search server 30 includes processing units corresponding to the acquisition unit 131, the extraction control unit 132, and the presentation unit 139 according to the above embodiment, in addition to the control unit that executes the search service. As described above, the fourth extraction unit 137 according to the extraction device uses the extraction model weighted by the index value calculated by the calculation unit 136 to extract not only the keyword associated with the content but also the content. Keywords can be extracted.

[4-7. Others]
In addition, among the processes described in the above embodiment, all or part of the processes described as being automatically performed can be performed manually, or the processes described as being performed manually can be performed. All or a part can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

  Further, each component of each illustrated apparatus is functionally conceptual, and does not necessarily need to be physically configured as illustrated. In other words, the specific form of distribution / integration of each device is not limited to that shown in the figure, and all or a part thereof may be functionally or physically distributed or arbitrarily distributed in arbitrary units according to various loads or usage conditions. Can be integrated and configured.

  For example, the calculation data storage unit 121 and the advertisement information storage unit 126 illustrated in FIG. 3 may be held by a storage server or the like (not shown) without being held by the advertising device 100. In this case, the advertising device 100 acquires information related to the advertisement from the storage server.

  Further, for example, in the above-described embodiment, an example in which the advertising device 100 acquires information related to a search query and the like and extracts a keyword based on the acquired information has been described, but the processing executed by the advertising device 100 is performed by the advertising device 100. It may be executed by an individual device that realizes the above functions. For example, the advertising device 100 described above includes an information acquisition device having an acquisition unit 131 and a submission reception unit 138, an extraction device having an extraction control unit 132, a presentation device having a presentation unit 139, and a request reception unit 140. It may be distributed to a distribution device having the distribution unit 141.

[5. Hardware configuration)
Further, the advertising device 100 according to the embodiment described above is realized by a computer 1000 having a configuration as shown in FIG. 17, for example. Hereinafter, the advertisement apparatus 100 will be described as an example. FIG. 17 is a hardware configuration diagram illustrating an example of a computer 1000 that implements the functions of the advertising device 100. The computer 1000 includes a CPU 1100, RAM 1200, ROM 1300, HDD 1400, communication interface (I / F) 1500, input / output interface (I / F) 1600, and media interface (I / F) 1700.

  The CPU 1100 operates based on a program stored in the ROM 1300 or the HDD 1400 and controls each unit. The ROM 1300 stores a boot program executed by the CPU 1100 when the computer 1000 is started up, a program depending on the hardware of the computer 1000, and the like.

  The HDD 1400 stores a program executed by the CPU 1100, data used by the program, and the like. The communication interface 1500 receives data from other devices via the communication network 500 (corresponding to the network N) and sends the data to the CPU 1100, and transmits data generated by the CPU 1100 to other devices via the communication network 500. .

  The CPU 1100 controls an output device such as a display and a printer and an input device such as a keyboard and a mouse via the input / output interface 1600. The CPU 1100 acquires data from the input device via the input / output interface 1600. Further, the CPU 1100 outputs the data generated via the input / output interface 1600 to the output device.

  The media interface 1700 reads a program or data stored in the recording medium 1800 and provides it to the CPU 1100 via the RAM 1200. The CPU 1100 loads such a program from the recording medium 1800 onto the RAM 1200 via the media interface 1700, and executes the loaded program. The recording medium 1800 is, for example, an optical recording medium such as a DVD (Digital Versatile Disc) or PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), a tape medium, a magnetic recording medium, or a semiconductor memory. Etc.

  For example, when the computer 1000 functions as the advertising device 100 according to the embodiment, the CPU 1100 of the computer 1000 implements the function of the control unit 130 by executing a program loaded on the RAM 1200. The HDD 1400 stores data in the storage unit 120. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them, but as another example, these programs may be acquired from other devices via the communication network 500.

[6. effect〕
As described above, the advertising device 100 according to the embodiment includes the calculation unit 136 and the fourth extraction unit 137. The calculation unit 136 is an extraction model that calculates an evaluation value as a keyword to be associated with an advertising content (an example of content) and extracts a keyword based on the calculated evaluation value, and uses different methods to evaluate the evaluation value The index value for each extraction model is calculated based on the degree to which keywords that match the predetermined condition are extracted by the extraction model for each extraction model for calculating. The fourth extraction unit 137 selects a keyword associated with the advertising content or a keyword for extracting the advertising content based on the evaluation value calculated by the extraction model weighted by the index value calculated by the calculation unit 136. Extract.

  As described above, the advertising device 100 according to the embodiment extracts a keyword appropriate for the advertiser using a plurality of different extraction models with respect to the bid keyword in the search-linked advertisement distribution service. For this reason, according to the advertising device 100, the advertiser can specify keywords extracted by a plurality of extraction models using various information as bid keywords. As a result, the advertiser can distribute more advertising content to users who are expected to be interested in the advertising content to be provided. That is, the advertising device 100 according to the embodiment can extract a keyword that enhances the appeal effect of the advertising content.

  The fourth extraction unit 137 is an evaluation value for each keyword calculated by a plurality of extraction models, and is based on the sum of evaluation values to which weights corresponding to the index values calculated by the calculation unit 136 are added. A keyword associated with the advertising content or a keyword for extracting the advertising content is extracted.

  Thus, the advertising device 100 according to the embodiment can extract bid keywords in the advertisement distribution service based on the keyword evaluation values calculated by a plurality of different methods. Therefore, according to the advertising device 100, it is possible to automatically select and extract an appropriate bid keyword from keywords extracted with different tendencies depending on each extraction model.

  In addition, the calculation unit 136 calculates an index value for each extraction model using a keyword that is actually associated with the advertising content or a keyword that is determined in advance to meet the condition as a keyword that satisfies the predetermined condition. To do.

  As described above, the advertising device 100 according to the embodiment calculates weights using, as positive examples, keywords that are actually set as bid keywords by other advertisers, keywords that are identified as appropriate as bid keywords, and the like. . Thereby, the advertising device 100 can make it easy to extract keywords that are assumed to be highly effective, such as many advertisers bidding. Therefore, the advertising device 100 can extract a bid keyword appropriate for the advertiser.

  Further, the calculation unit 136 calculates an index value for each extraction model generated from the same learning data.

  As described above, the advertising device 100 according to the embodiment learns each extraction model from the common test data, and thus can learn without biasing to specific information and execute the extraction process using the generated extraction model. .

  Further, the calculation unit 136 extracts, as an extraction model, a keyword based on similarity between LP feature information (for example, a word vector) associated with the keyword and feature information corresponding to the keyword. Model, LP or second extraction model that extracts keywords based on a predetermined index value of each combination of keywords generated for each advertiser who has specified a keyword for LP or advertisement content, or a search transmitted from user terminal 10 A third extraction model that extracts a keyword based on a predetermined index value indicating a relationship between the query and a re-search query that is a search query further transmitted from the user terminal 10 after transmitting the search query; The index value of one extraction model is calculated.

  As described above, the advertising device 100 according to the embodiment calculates a weight for an extraction model that extracts keywords using different methods. Thereby, the advertising device 100 can extract bid keywords based on various information and without biasing to specific information.

  In addition, the calculation unit 136 may determine whether the keyword extracted by the fourth extraction unit 137 is input by the user as a search query for extracting advertisement content or LP, the number of times the advertisement content is selected by the user, or the advertisement content. The index value for each extraction model is calculated based on at least one of the information on the profits obtained by the advertiser by displaying.

  As described above, the advertising device 100 according to the embodiment can calculate the weight for each extraction model in consideration of the current advertising effect. For this reason, the advertising device 100 can extract an appropriate keyword in accordance with the current situation.

  In addition, the calculation unit 136 calculates an index value for each extraction model that calculates an evaluation value as a keyword for which an advertiser bids for advertisement distribution.

  Thereby, the advertising device 100 according to the embodiment can extract an appropriate bid keyword in the advertisement distribution service.

  Further, the calculation unit 136 calculates an index value indicating a weight for each extraction model, using the advertisement content submitted by the advertiser or the web page that is the link destination of the advertisement content as the content.

  Thereby, the advertising device 100 according to the embodiment can extract an appropriate bid keyword in a keyword match type advertisement distribution service such as a search-linked type.

  In addition, the advertising device 100 according to the embodiment includes a presentation unit 139 that presents a keyword extracted by the fourth extraction unit 137 to a content provider (for example, an advertiser) as a keyword associated with predetermined content. Further prepare.

  As described above, the advertising device 100 according to the embodiment presents an appropriate keyword to an advertiser who bids for a keyword. Therefore, according to the advertising device 100, the advertiser can know an appropriate keyword as a newly added bid keyword. Further, the manager of the advertisement distribution service can promote bidding by increasing the number of keywords bid by the advertiser, and can activate the service.

  As described above, some of the embodiments of the present application have been described in detail based on the drawings. It is possible to implement the present invention in other forms with improvements.

  In addition, the “section (module, unit)” described above can be read as “means” or “circuit”. For example, the calculation unit can be read as calculation means or a calculation circuit.

DESCRIPTION OF SYMBOLS 1 Extraction system 10 User terminal 20 Advertiser terminal 30 Search server 100 Advertising apparatus 110 Communication part 120 Storage part 121 Calculation data storage part 122 1st table 123 2nd table A
124 Second table B
125 third table 126 advertisement information storage unit 130 control unit 131 acquisition unit 132 extraction control unit 133 first extraction unit 134 second extraction unit 135 third extraction unit 136 calculation unit 137 fourth extraction unit 138 submission reception unit 139 presentation unit 140 Request reception unit 141 Distribution unit

Claims (10)

An extraction model that calculates an evaluation value as a keyword to be linked to content and extracts the keyword based on the calculated evaluation value, and for each of the extraction models that calculates the evaluation value using different methods The index value for each extraction model is calculated based on the degree to which keywords that are actually linked to the content by the extraction model or keywords that are determined in advance to meet predetermined conditions are extracted. A calculation unit;
An extraction unit for extracting a keyword associated with the content or a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by the index value calculated by the calculation unit; ,
An extraction device comprising: The extraction unit includes:
The evaluation value for each keyword calculated by a plurality of extraction models, and associated with the content based on a sum of the evaluation values to which a weight corresponding to the index value calculated by the calculation unit is added A keyword or a keyword for extracting the content is extracted;
The extraction device according to claim 1. The calculation unit includes:
Calculating the index value for each of the extraction models generated by the same learning data;
The extraction apparatus according to claim 1 or 2 , wherein The calculation unit includes:
As the extraction model, a first extraction model for extracting the keyword based on the similarity between the feature information of the content associated with the keyword and the feature information corresponding to the keyword, and the keyword specified for the content A second extraction model for extracting the keyword based on a predetermined evaluation value of each combination of the keywords generated for each content provider, or a search query transmitted from a user terminal and the search query are transmitted And at least any two of the third extraction models that extract the keyword based on a predetermined evaluation value indicating a relevance with a re-search query that is a search query further transmitted from the user terminal Calculating the index value of the model;
The extraction device according to any one of claims 1 to 3 , wherein The calculation unit includes:
The number of times the keyword extracted by the extraction unit is input by the user as a search query for extracting the content, the number of times the content is selected by the user, or the content is displayed by displaying the content Calculating the index value for each of the extraction models based on at least one of the information on the profit obtained by
The extraction device according to any one of claims 1 to 4 , wherein The calculation unit includes:
Calculating the index value for each of the extraction models for calculating an evaluation value as a keyword bid by the advertiser in the advertisement distribution;
The extraction apparatus according to any one of claims 1 to 5 , wherein The calculation unit includes:
The index value for each extraction model is calculated using the advertising content submitted by the advertiser as the content, or the web page that is the link destination of the advertising content.
The extraction apparatus according to any one of claims 1 to 6 , wherein A presentation unit that presents a keyword extracted by the extraction unit as a keyword to be associated with the content to a provider who provides the content;
Extraction device according to any one of claims 1 to 7, further comprising a. An extraction method performed by a computer,
An extraction model that calculates an evaluation value as a keyword to be linked to content and extracts the keyword based on the calculated evaluation value, and for each of the extraction models that calculates the evaluation value using different methods The index value for each extraction model is calculated based on the degree to which keywords that are actually linked to the content by the extraction model or keywords that are determined in advance to meet predetermined conditions are extracted. A calculation process;
An extraction step of extracting a keyword associated with the content or a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by the index value calculated by the calculation step; ,
The extraction method characterized by including. An extraction model that calculates an evaluation value as a keyword to be linked to content and extracts the keyword based on the calculated evaluation value, and for each of the extraction models that calculates the evaluation value using different methods The index value for each extraction model is calculated based on the degree to which keywords that are actually linked to the content by the extraction model or keywords that are determined in advance to meet predetermined conditions are extracted. Calculation procedure,
An extraction procedure for extracting a keyword associated with the content or a keyword for extracting the content based on the evaluation value calculated by the extraction model weighted by the index value calculated by the calculation procedure; ,
An extraction program characterized by causing a computer to execute.

Images