JP2024004703A - Query forming system, method for forming query, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To form a conversion query in a search linkage advertisement to a proper state.

SOLUTION: An acquisition unit (201) of a relation estimation unit (1) acquires a conversion query in which a search linkage advertisement has caused a conversion in a predetermined service. A query forming unit (203) forms a conversion query on the basis of a learning model which has learned training data regarding formation of a query. A processing execution unit (204) executes predetermined processing on the basis of the conversion query formed by the query formation unit.

SELECTED DRAWING: Figure 4

COPYRIGHT: (C)2024,JPO&INPIT

Description

The present disclosure relates to a query formatting system, a query formatting method, and a program.

BACKGROUND ART Search-linked advertisements have been known that display search results corresponding to a query entered by a user as well as advertisements corresponding to the query. Conversion queries that have resulted in conversions through search-linked advertising are extremely useful information because they can be used for a variety of purposes. For example, a conversion query can be used as an index or display keyword in search-linked advertising, or used in marketing to analyze user trends.

In paid search advertising, conversion queries are not always obtained in the appropriate state. For example, conversion queries may not be inserted at all, or may be filled with unnecessary spaces, so it is important to format conversion queries appropriately. For example, Patent Document 1 describes that after removing spaces from a search log query, the query is formatted by dividing the query into a plurality of character strings.

Japanese Patent Application Publication No. 2012-141681

However, Patent Document 1 merely describes formatting a query of a search log in which a large number of queries input in the past are stored regardless of conversion. Therefore, the technology of Patent Document 1 does not format conversion queries in search-linked advertisements. Conventional technology has not been able to appropriately format conversion queries in paid search advertising.

One of the purposes of the present disclosure is to appropriately format a conversion query in search-linked advertising.

A query shaping system according to the present disclosure includes an acquisition unit that acquires a conversion query in which a conversion related to a search-linked advertisement has occurred, and a query shaping system that formats the conversion query based on a learning model in which training data related to query shaping has been learned. and a processing execution unit that executes a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit.

According to the present disclosure, a conversion query in a search-linked advertisement can be appropriately formatted.

FIG. 1 is a diagram illustrating an example of the overall configuration of a query formatting system. FIG. 2 is a diagram illustrating an example of a flow in which a search-linked advertisement is displayed in a search service and a conversion occurs in an EC service. FIG. 3 is a diagram illustrating an example of a flow of formatting a conversion query. It is a figure showing an example of the function realized by a 1st embodiment. It is a figure showing an example of a product database. FIG. 3 is a diagram showing an example of a conversion database. FIG. 3 is a diagram illustrating an example of a learning method for a query shaping model. It is a diagram showing an example of an advertisement database. It is a flow diagram showing an example of processing performed in a 1st embodiment. It is a figure showing an example of a relevance estimation model in a 2nd embodiment. It is a figure which shows an example of the function implement|achieved in 2nd Embodiment. It is a figure which shows an example of the process performed in 2nd Embodiment. It is a figure which shows an example of the function in the modification regarding 1st Embodiment. 7 is a diagram illustrating an example of query formatting in Modification 1-2. FIG.

[1. First embodiment]
A first embodiment, which is an example of an embodiment according to the present disclosure, will be described. In the first embodiment, a case will be exemplified in which an operator of an EC (Electronic Commerce) service uses a query formatting system. The query shaping system is also applicable to services other than EC services. Examples of application to other services will be explained in modified examples below.

[1-1. Overall configuration of query formatting system]
FIG. 1 is a diagram showing an example of the overall configuration of a query formatting system. For example, the query shaping system 1 includes an EC server 10 and an advertiser terminal 20. A search server 30 and a user terminal 40 exist outside the query formatting system 1. Each of the EC server 10, advertiser terminal 20, search server 30, and user terminal 40 can be connected to a network N such as the Internet or a LAN.

The EC server 10 is a server computer in the EC service. For example, the EC server 10 includes a control section 11, a storage section 12, and a communication section 13. Control unit 11 includes at least one processor. The storage unit 12 includes volatile memory such as RAM and nonvolatile memory such as flash memory. The communication unit 13 includes at least one of a communication interface for wired communication and a communication interface for wireless communication.

The advertiser terminal 20 is a computer of an advertiser that posts a search-linked advertisement on a search service. In the first embodiment, a case will be described in which the operator of the EC service is an advertiser, but the advertiser may be another entity such as a store or a product manufacturer that opens a store in the EC service. For example, the advertiser terminal 20 is a personal computer, a smartphone, a tablet terminal, or a wearable terminal.

For example, the advertiser terminal 20 includes a control section 21, a storage section 22, a communication section 23, an operation section 24, and a display section 25. The physical configurations of the control section 21, the storage section 22, and the communication section 23 are the same as those of the control section 11, the storage section 12, and the communication section 13, respectively. The operation unit 24 is an input device such as a touch panel or a mouse. The display section 25 is a liquid crystal display or an organic EL display.

The search server 30 is a server computer in a search service. In the first embodiment, a web page search service is taken as an example, but the search service may be a service for searching other search targets such as content, products, services, or documents. For example, the search server 30 includes a control section 31, a storage section 32, and a communication section 33. The physical configurations of the control section 31, the storage section 32, and the communication section 33 are the same as those of the control section 11, the storage section 12, and the communication section 13, respectively.

User terminal 40 is a user's computer. For example, the user terminal 40 is a personal computer, a smartphone, a tablet terminal, or a wearable terminal. For example, the user terminal 40 includes a control section 41, a storage section 42, a communication section 43, an operation section 44, and a display section 45. The physical configurations of the control unit 41, storage unit 42, communication unit 43, operation unit 44, and display unit 45 are the same as those of the control unit 11, storage unit 12, communication unit 13, operation unit 24, and display unit 25, respectively. be.

Note that the programs stored in the storage units 12, 22, 32, and 42 may be supplied via the network N. In addition, at least one of a reading section (for example, a memory card slot) for reading a computer-readable information storage medium and an input/output section (for example, a USB port) for inputting and outputting data with an external device is connected to each computer. may be included in For example, a program stored on an information storage medium may be supplied via at least one of a reading section and an input/output section.

Further, the query formatting system 1 may include at least one computer, and is not limited to the example shown in FIG. 1. For example, the query shaping system 1 may include only the advertiser terminal 20 without including the EC server 10. In this case, the EC server 10 exists outside the query formatting system 1. For example, the query shaping system 1 may include only the EC server 10 without including the advertiser terminal 20. In this case, the advertiser terminal 20 exists outside the query shaping system 1, and the EC server 10 has the same functions as the advertiser terminal 20.

[1-2. Overview of query formatting system]
In the first embodiment, an operator of an EC service requests a search service to post a search-linked advertisement. Search-linked advertisements are advertisements that are displayed along with search results in search services. A search-linked advertisement is an advertisement related to a query. A query is a search condition input by a user. Queries are sometimes called keywords. Various types of search engines, queries, and search-linked advertisements can be used in search services. Queries can be entered in any language such as Japanese, Chinese, or English.

FIG. 2 is a diagram showing an example of a flow in which a search-linked advertisement is displayed in a search service and a conversion occurs in an EC service. Conversion is the result of paid search advertising. In the first embodiment, a case will be described in which the purchase of a product corresponds to a conversion, but a conversion may be any state in which a target action is performed, and is not limited to the purchase of a product. For example, a conversion may be selecting a paid search advertisement, displaying a particular web page, downloading particular data, or adding a product to a shopping cart.

For example, when the user terminal 40 accesses the search server 30, the top screen SC1 of the search service is displayed on the display unit 45. The user inputs a query into the input form F10 on the top screen SC1. When the user selects button B11, the search server 30 executes a search for web pages and a search for search-linked advertisements based on the query input by the user. When the user terminal 40 receives these execution results from the search server 30, a search result screen SC2 is displayed on the display unit 45.

For example, the search result screen SC2 includes a display area A20 for displaying search-linked advertisements and a display area A21 for displaying search results. If there is no search-linked advertisement that hits the query, no search-linked advertisement is displayed in the display area A20. The search-linked advertisement includes a link to an EC service. The link includes a query and an advertisement ID that can identify the search-linked advertisement as arguments. The link argument allows the EC service operator to specify which search-linked advertisement was selected by what query.

In the first embodiment, it is assumed that a search service performs morphological analysis on a query input by a user. In a search service, a query entered by a user is divided into morphemes, unnecessary words are removed, and a search is executed. The operator of the EC service is different from the operator of the search service. Therefore, it is assumed that the operator of the EC service cannot grasp what kind of morphological analysis will be executed on the search service side.

For example, the display area A20 displays a query divided into morphemes and a link to an EC service. The query included in the link argument is a query divided into morphemes. The input form F22 displays the query before it is divided into morphemes. However, internally in the search service, the search may be executed based on the query displayed in the display area A20 instead of the input form F22. The user can also execute a search by inputting a query again into the input form F22.

For example, when the user selects the search-linked advertisement in the display area A20, the user terminal 40 accesses the EC server 10 and the product screen SC3 is displayed on the display unit 45. The product screen SC3 is a web page linked to a search-linked advertisement. When the user adds products to the shopping cart using the EC service and completes the order procedure, an order completion screen SC4 is displayed on the display unit 45. The product ordering procedure itself may be the same as that of a known EC service.

In the first embodiment, a conversion is when a user selects a search-linked advertisement, accesses an EC server, and completes an order for a product. Hereinafter, a query that causes a conversion will be referred to as a conversion query. In the display area A20, the conversion query is displayed divided into morphemes, which may make the user feel uncomfortable. Furthermore, since the conversion query displayed in the display area A20 does not look good, the appeal effect of the search-linked advertisement may be reduced.

For example, it is conceivable that the search service side displays a conversion query (conversion query in the same state as the input form F22) before being divided into morphemes in the display area A20. However, even if an EC service operator requests such modifications, there is a possibility that the search service operator will not comply with the modification. Even if the search service operator agrees to the modification, the burden will be on the search service operator.

In this regard, if the operator of the EC service can obtain the conversion query before it is divided into morphemes, by registering this conversion query in the database dedicated to advertisers in the search service, the operator of the EC service can adjust the display in the display area A20 appropriately. can be in a state. However, since the conversion query in a state after being divided into morphemes is included in the link argument, the operator of the EC service cannot obtain the conversion query in an appropriate state. Therefore, the query formatting system 1 formats the conversion query divided into morphemes so that they return to their original form.

FIG. 3 is a diagram illustrating an example of the flow of formatting a conversion query. In the first embodiment, a word unit included in a conversion query is called a token. Tokens are words separated by separators such as spaces or symbols. In the example of FIG. 3, the conversion query Q1 input by the user includes tokens T10 to T12. A separator exists between tokens T10 and T11 and between tokens T11 and T12.

For example, the conversion query Q2 after being divided into morphemes includes tokens T20 to T27. Between tokens T20 and T21, between tokens T21 and T22, between tokens T22 and T23, between tokens T23 and T24, between tokens T24 and T25, between tokens T25 and T26, and between tokens T26 and T27 , there is a separator. The EC service operator obtains the conversion query Q2.

For example, detokenization is performed on conversion query Q2. Detokenization is the removal of all or some separators. In the example of FIG. 3, all separators of conversion query Q2 are deleted. Therefore, all tokens T20 to T27 are combined. Detokenized conversion query Q3 includes only one token T30.

In the first embodiment, a query shaping model M1 using a machine learning technique is prepared. The query shaping model M1 has been trained to perform appropriate tokenization on the detokenized conversion query Q3. Tokenization is the insertion of at least one separator. In the example of FIG. 3, the conversion query Q4 shaped by the query shaping model M1 is the same as the conversion query Q1.

As described above, the query formatting system 1 performs detokenization on the conversion query Q2 after being divided into morphemes, and then performs tokenization using the query formatting model M1. Get Q1. This allows conversion queries in paid search ads to be formatted appropriately. Hereinafter, the details of the query formatting system 1 will be explained. Note that when there is no need to refer to FIG. 3, the reference numbers for conversion queries Q1 to Q4 are omitted.

[1-3. Functions realized in the first embodiment]
FIG. 4 is a diagram illustrating an example of functions realized in the first embodiment.

[1-3-1. Functions realized by EC server]
For example, the EC server 10 includes a data storage section 100 and an order processing section 101. The data storage section 100 is realized by the storage section 12. The order processing section 101 is realized by the control section 11.

[Data storage unit]
The data storage unit 100 stores data necessary for providing EC services. For example, the data storage unit 100 stores a product database DB1 and a conversion database DB2.

FIG. 5 is a diagram showing an example of the product database DB1. The product database DB1 is a database that stores various information regarding products in the EC service. For example, the product database DB1 stores store IDs, product IDs, product titles, product attributes, product prices, product detailed descriptions, and search keywords. The product database DB1 may store any information regarding products, and may also store information other than that shown in FIG.

The store ID is an example of store identification information that can identify a store in the EC service. The store identification information may be other information such as a store name, telephone number, or email address. The product ID is an example of product identification information that can identify the product. The product identification information may be other information such as a product name or JAN code. The title, attribute, price, detailed description, and keyword are used as an index when searching in the EC service. Even in EC services, rather than search services, users can search for products by entering a query.

FIG. 6 is a diagram showing an example of the conversion database DB2. The conversion database DB2 is a database that stores various information regarding conversions that occur in EC services. For example, the conversion query database stores an advertisement ID, a product ID, a user ID, a conversion query divided into morphemes, and the date and time of conversion. The conversion database DB2 may store any information regarding conversion, and may also store information other than that shown in FIG. 6.

The advertisement ID is an example of advertisement identification information that can identify a search-linked advertisement. The advertisement identification information may be any information that can identify the search-linked advertisement in some way, and may be other information called by a name other than the advertisement ID. The conversion database DB2 stores advertisement IDs of search-linked advertisements that are the cause of conversions. The advertisement ID is included in the link argument of the search-linked advertisement. The EC server 10 can identify which search-linked advertisement has been selected by acquiring the advertisement ID included in the argument.

For example, the conversion database DB2 contains the product ID of the product ordered by conversion, the user ID of the user who ordered the product, the conversion query included in the link argument of the search-linked advertisement, and the current date and time when the conversion occurred. is stored. In the first embodiment, each time a conversion occurs, the order processing unit 101, which will be described later, creates a new record in the conversion database DB2, and stores various information regarding the conversion.

Note that the data stored in the data storage unit 100 is not limited to the above example. For example, the data storage unit 100 may store a user database that stores various information regarding users who have registered as members of the EC service. The user database stores information such as a user ID, password, name, address, telephone number, email address, product browsing history, and product purchase history.

[Order processing department]
The order processing unit 101 executes order processing for accepting orders for products from users. The order processing itself can utilize the processing of a known EC service. For example, when a user places a product in a shopping cart and processes an order, the order processing unit 101 executes a payment process necessary for paying for the product and a shipping process necessary for shipping the product. The payment process is a process for making a payment using a payment method such as a credit card specified by the user. The shipping process is a process for transmitting information regarding the product ordered by the user to the store. As described above, when a conversion occurs, the order processing unit 101 updates the conversion database DB2.

[1-3-2. Functions realized on advertiser terminal]
For example, the advertiser terminal 20 includes a data storage section 200, an acquisition section 201, a detokenization execution section 202, a query shaping section 203, and a processing execution section 204. The data storage section 200 is realized by the storage section 12. The acquisition unit 201, detokenization execution unit 202, query shaping unit 203, and processing execution unit 204 are realized by the control unit 11.

[Data storage unit]
The data storage unit 200 stores data necessary for query shaping. For example, the data storage unit 200 stores a learned query shaping model M1. The query shaping model M1 includes a program part that executes query shaping processing and a parameter part that is adjusted through learning. For example, the parameters are weights and biases. The parameters may be those that are compatible with the machine learning method used as the query shaping model M1.

In the first embodiment, a query shaping model M1 using Seq2Seq (Sequence to Sequence) is taken as an example. Seq2Seq includes an encoder into which input data to be processed is input, and a decoder which acquires the feature amount output from the encoder and outputs output data according to the feature amount. The feature amount is information regarding the feature of input data. In the first embodiment, a case will be described where the feature amount is expressed as a multidimensional vector, but the feature amount may be expressed in other formats such as an array or a single numerical value.

FIG. 7 is a diagram illustrating an example of a learning method for the query shaping model M1. For example, the training data for the query shaping model M1 is a pair of a training query corresponding to input data and a training query corresponding to output data. The training query corresponding to the input data is the training query before shaping. The training query corresponding to the output data is a formatted training query. The pairs included in the training data may be prepared depending on the type of shaping to be performed by the query shaping model M1.

In this embodiment, a detokenized conversion query is input to the query shaping model M1, and a tokenized conversion query is output from the query shaping model M1, so the pairs included in the training data are also It depends on the pair. In the example of FIG. 7, the training query corresponding to the input data is a training query that includes only one token by detokenizing the training query that includes multiple tokens. The training query corresponding to the output data is a training query that includes the plurality of tokens.

For example, the advertiser terminal 20 adjusts the parameters of the query shaping model M1 so that when a training query corresponding to input data is input, a training query corresponding to training output data is output. , the query shaping model M1 is made to learn the training data. For the learning itself, various algorithms used in machine learning techniques can be used, and for example, error backpropagation or gradient descent may be used. Learning of the query shaping model M1 may be performed by another computer instead of the advertiser terminal 20.

In the first embodiment, a query actually input in an EC service is used as a training query. This query does not have to be a conversion query. For example, the advertiser terminal 20 acquires a query actually input in the EC service from the EC server 10. The advertiser terminal 20 generates training data based on this query. If this query includes a separator at an appropriate position, the pair of this query and the detokenized one becomes the training data. If this query does not include a separator at an appropriate position, the EC service operator may manually format the query as appropriate.

Note that a query actually input in a search service instead of an EC service may be used as a training query. This query does not have to be a conversion query. For example, the advertiser terminal 20 obtains from the search server 30 a query actually input in a search service. The advertiser terminal 20 may generate training data based on this query. Further, the data storage unit 200 may store other data. For example, the data storage unit 200 may store a conversion database DB2.

Further, the query shaping model M1 can use various machine learning methods other than Seq2Seq. The query shaping model M1 may be a supervised learning model, a semi-supervised learning model, or an unsupervised learning model. For example, a model using a neural network or a model called Transformer may be used. The query shaping model M1 may be a model using a machine learning method. Furthermore, the query shaping model M1 may be a Seq2Seq model that uses RNNs such as LSTM and GRU as encoders and decoders.

[Acquisition part]
The acquisition unit 201 acquires a conversion query in which a conversion related to a search-linked advertisement has occurred. In the first embodiment, since the conversion query is stored in the conversion database DB2, the acquisition unit 201 acquires the conversion query to be subjected to query shaping from the conversion database DB2. If the conversion query is stored in a database other than the conversion database DB2 or in a computer other than the EC server 10, the acquisition unit 201 acquires the conversion query from the other database or computer.

In the first embodiment, the search-linked advertisement is an advertisement for allowing the user to access an EC service different from the search service in which the search-linked advertisement is displayed. For example, the acquisition unit 201 acquires a conversion query that caused a conversion in an EC service. EC service is an example of other services. The other services may be any service other than the search service and are not limited to the EC service. For example, other services may be travel reservation services, online flea market services, video distribution services, payment services, financial services, or communication services.

In the search service of the first embodiment, the conversion query input during the search is changed, so the acquisition unit 201 acquires the changed conversion query that was changed in the search service. Modifying a conversion query means adding a separator, letter, number, or other symbol to the conversion query, or removing a separator, letter, number, or other symbol from the conversion query. Moving some words of the conversion query may correspond to changing the conversion query.

In the first embodiment, the search service performs morphological analysis, so dividing the conversion query into morphemes and adding a separator corresponds to changing the conversion query. Therefore, the portion explaining morphological analysis can be replaced with other arbitrary changes. Various methods can be used for morphological analysis itself. For example, a morphological analysis tool such as MeCab, JUMAN, or JANOME may be used. The morphological analysis tool is not limited to Japanese, and may be a tool compatible with other languages such as Chinese or English.

[Detokenization execution department]
The detokenization execution unit 202 executes detokenization regarding a conversion query. For example, the detokenization execution unit 202 executes detokenization by deleting at least part of the separators included in the changed conversion query. In the first embodiment, a case will be described in which the detokenization execution unit 202 deletes all separators included in a changed conversion query, but only some separators included in a changed conversion query are deleted. You may. In the example of FIG. 3, detokenization results in a state where there is only one token.

[Query formatter]
The query shaping unit 203 formats the conversion query based on the query shaping model M1. The query shaping model M1 is an example of a learning model from which training data related to query shaping has been learned. Therefore, the portion describing the query shaping model M1 can be read as a learning model. The training data may be created according to the machine learning method used as the learning model.

For example, the query formatting unit 203 formats the modified conversion query so that the modified conversion query returns to the state at the time of input. The state at the time of input is the state before the change by the search service. The query shaping unit 203 inputs the changed conversion query to the trained query shaping model M1. The query shaping model M1 calculates the feature amount of the changed conversion query and outputs a conversion query according to the calculated feature amount.

For example, the output conversion query has the same or substantially the same content as the conversion query when input in the search service. Substantially the same content means that the token and separator differ by about 1 to 3 characters. In addition, for example, the output conversion query may be a natural content from which unnecessary separators included in an unnatural conversion query when input into a search service are removed, or an unnatural conversion query when input into a search service. It can be natural content with any necessary separators added that are not included in the query.

In the search service of the first embodiment, the conversion query input at the time of search is changed by performing morphological analysis. Format the modified conversion query so that it reverts to the previous state. For example, the query formatting unit 203 formats the modified conversion query so that a token divided into a plurality of morphemes by morphological analysis is returned as one token. In other words, the query formatting unit 203 formats the modified conversion query so as to delete the separator inserted by the morphological analysis.

In the first embodiment, the training query is a query actually input in the EC service, so the query shaping unit 203 converts the training query actually input in the EC service into a modified training query. , and the training query before modification, the conversion query is formatted based on the query formatting model M1 in which the relationship between the following is learned as training data. The learning method for causing the query shaping model M1 to learn a query actually input in the EC service as a training query is as described above.

In the first embodiment, the query shaping unit 203 formats the executed conversion query by performing tokenization on the conversion query that has been detokenized based on the query shaping model M1. For example, the query shaping unit 203 performs tokenization on a conversion query that has become one token through detokenization, by inserting a separator at an appropriate position. The appropriate location is determined by query shaping model M1.

[Processing execution unit]
The processing execution unit 204 executes a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit 203 . The predetermined process may be any process that corresponds to the purpose of the formatted conversion query. For example, the predetermined process may be a recording process that records the formatted conversion query in the data storage unit 200, or an output process that outputs the formatted conversion query to the outside.

In the first embodiment, the processing execution unit 204 executes a registration process of registering a formatted conversion query in a search service as a predetermined process. For example, an index and a display keyword are registered for each search-linked advertisement in a search service, such as an advertisement database DB3 described later. In the search service, search-linked advertisements are searched based on the index, and display keywords are displayed together with the search-linked advertisements.

For example, the processing execution unit 204 executes a process of registering a formatted conversion query as a display keyword in a search service as a registration process. The processing execution unit 204 executes the registration process by transmitting the formatted conversion query to the search server 30. When the search server 30 receives the formatted conversion query, it associates it with the operator of the EC service and registers the received formatted conversion query as a display keyword in the advertising database DB3.

Note that the formatted conversion query may be registered in association with the search-linked advertisement that is the source of the conversion, or may be registered in association with another search-linked advertisement. For example, an operator of an EC service may specify a search-linked advertisement to be associated with a formatted conversion query. Further, the processing execution unit 204 may execute, as a predetermined process, a registration process of registering a formatted conversion query as an index of a search-linked advertisement.

[1-3-3. Functions realized by search server]
For example, the search server 30 includes a data storage section 300 and a search section 301. The data storage section 300 is realized by the storage section 32. The search unit 301 is realized by the control unit 31.

[Data storage unit]
The data storage unit 300 stores data necessary for providing search services. For example, the data storage unit 300 stores an advertisement database DB3.

FIG. 8 is a diagram showing an example of the advertisement database DB3. The advertisement database DB3 is a database that stores various information regarding search-linked advertisements. For example, the advertisement database DB3 stores advertisement IDs, advertiser IDs, actual data of search-linked advertisements, indexes, and display keywords. The advertisement database DB3 may store any information related to search-linked advertisements, and may also store information other than that shown in FIG. 8.

The advertiser ID is an example of advertiser identification information that can identify the advertiser. The advertiser identification information may be other information such as an advertiser name, telephone number, or email address. In the first embodiment, it is assumed that the user wishes to publish search-linked advertisements of advertisers other than the EC service operator. Therefore, the advertisement database DB3 stores information regarding search-linked advertisements of various advertisers. The actual data is data such as character strings, images, and link destination URLs that indicate the content of the advertisement. In the first embodiment, it is assumed that the argument of the link destination URL includes a query divided into morphemes and an advertisement ID.

An index is information that is compared with a query. Any information can be used as an index. For example, a part of the title of the product, an attribute of the product, a keyword extracted from the detailed description of the product, a character string specified by the operator of the EC service, or a combination thereof is used as the index. The display keyword is a character string used when displaying a search-linked advertisement in the display area A20. If nothing is specified as a display keyword as shown in FIG. 8, a query divided into morphemes is displayed in the display area A20. When the display keyword is specified, the display keyword is displayed in the display area A20.

Note that the data stored in the data storage unit 300 is not limited to the above example. The data storage unit 300 can store arbitrary data. For example, the data storage unit 300 may store a web page database necessary for searching web pages. For example, a web page database stores URLs and indexes of various web pages.

[Search section]
When a user inputs a query, the search unit 301 executes a search for web pages and a search for search-linked advertisements based on the query. These searches themselves may be performed based on a known search engine. For example, the search unit 301 executes a search based on a query input by the user, an index stored in the web page database, and an index stored in the advertisement database DB3. The search unit 301 obtains search results for web pages and search results for search-linked advertisements. The search unit 301 transmits display data of a search result screen SC2 showing these search results to the user terminal 40.

Note that the display data may be in any format; for example, if the search result screen SC2 is to be displayed on a browser, it is HTML data. If the search result screen SC2 is to be displayed on an application dedicated to a search service, any data may be in a format supported by the application. In addition, the search unit 301 can transmit display data of various screens of the search service to the user terminal 40.

[1-3-4. Functions realized on user terminal]
For example, the user terminal 40 includes a data storage section 400, a display control section 401, and an operation reception section 402. The data storage section 400 is realized by the storage section 42. The display control unit 401 and the operation reception unit 402 are realized by the control unit 41.

[Data storage unit]
The data storage unit 400 stores data necessary for providing search services and EC services. For example, the data storage unit 400 stores a browser or a dedicated application for displaying each screen.

[Display control section]
The display control unit 401 causes the display unit 45 to display various screens.

[Operation reception department]
The operation reception unit 402 receives various operations from the operation unit 44.

[1-4. Processing executed in the first embodiment]
FIG. 9 is a flow diagram illustrating an example of processing executed in the first embodiment. This process is executed by the control units 11, 21, 31, 41 operating according to programs stored in the storage units 12, 22, 32, 42. Hereinafter, the process when a conversion by a user occurs will be explained. For this reason, the point at which user input occurs is called a conversion query.

As shown in FIG. 9, the search server 30 executes processing for displaying the top screen SC1 of the search service with the user terminal 40 (S1). When the process of S1 is executed, the user inputs an arbitrary character string as a conversion query into the input form F10 and selects the button B11. The user terminal 40 transmits the conversion query input by the user to the search server 30 (S2). When the search server 30 receives the conversion query, it executes morphological analysis using a morphological analysis tool (S3).

The search server 30 executes a search for web pages and a search for search-linked advertisements based on the conversion query divided into morphemes and the advertisement database DB3 (S4). The search server 30 generates display data for the search result screen SC2 based on the search execution result in S4 and transmits it to the user terminal 40 (S5). In S5, the search server 30 includes the conversion query divided into morphemes in S3 and the advertisement ID of the search-linked advertisement as arguments in the link of the search-linked advertisement.

When the user terminal 40 receives the display data, it displays the search result screen SC2 on the display unit 45 (S6). When the user selects the search-linked advertisement, the user terminal 40 executes processing to display the product screen SC3 with the EC server 10 (S7). Using the argument included in the link of the search-linked advertisement, the EC server 10 can obtain the conversion query after it has been divided into morphemes in S3. The EC server 10 executes processing with the user terminal 40 to receive an order for a product and display an order completion screen SC4 (S8). In S8, the EC server 10 stores the conversion query in the conversion database DB2.

The advertiser terminal 20 executes a process with the EC server 10 to obtain a conversion query to be subjected to query shaping (S9). A conversion query to be subjected to query shaping may be manually specified. All or part of the conversion queries stored in the conversion database DB2 may be acquired. The advertiser terminal 20 formats the conversion query acquired in S9 based on the learned query formatting model M1 (S10). The advertiser terminal 20 transmits the formatted conversion query to the search server 30 (S11). When the search server 30 receives the formatted conversion query, it stores the formatted conversion query as a display keyword in the advertisement database DB3 (S12), and this process ends. Here, the display keyword corresponds to a search query that is a trigger in search-linked advertising.

The query formatting system 1 of the first embodiment formats a conversion query based on the query formatting model M1. The query formatting system 1 executes predetermined processing based on the formatted conversion query. This allows conversion queries in paid search advertising to be formatted appropriately. Therefore, predetermined processing can be executed after the conversion query is in an appropriate state, making it easier to achieve the purpose of using the conversion query. For example, as in the first embodiment, when using a conversion query for the purpose of preventing the display of search-linked advertisements displayed in the display area A20 from becoming unnatural, it is necessary to make the display more natural. Can be done.

The query shaping system 1 also acquires a conversion query in which a conversion occurred in an EC service different from the search service in which the search-linked advertisement is displayed. This allows conversion queries to be formatted appropriately even if conversions occur across services.

In addition, the query formatting system 1 executes a registration process for registering a formatted conversion query in a search service as a predetermined process. As a result, even if a conversion occurs across services, an appropriately formatted conversion query can be registered with the search service.

In addition, the query formatting system 1 executes a process of registering a formatted conversion query as a display keyword in a search service as a registration process. Thereby, it is possible to display search-linked advertisements in a search service in a natural manner.

In addition, the query formatting system 1 formats the modified conversion query so that the modified conversion query returns to the state at the time of input in the search service. This way, even if a search service changes the conversion query to an unnatural state, it can be restored to its original state.

Furthermore, the query formatting system 1 formats the modified conversion query so that the morphemes included in the modified conversion query return to the state at the time of input. As a result, even if a search service splits a conversion query into morphemes, it can be restored to its original state.

The query shaping system 1 also processes a query shaping system in which the relationship between a modified training query obtained by modifying a training query actually input in the EC service and a training query before modification is learned as training data. A conversion query is formatted based on model M1. Thereby, a conversion query can be formatted using the query formatting model M1 that has been trained by effectively utilizing information held by the EC service operator.

Furthermore, the query shaping system 1 formats the executed conversion query by performing tokenization on the executed conversion query that has undergone detokenization based on the query shaping model M1. As a result, by detokenization, no matter what state the conversion query is in, query formatting can be performed after converting it to the same state of one token, thereby increasing the accuracy of query formatting for the conversion query.

[2. Second embodiment]
Next, a second embodiment, which is another embodiment of the query formatting system 1, will be described. For example, in addition to registering a conversion query as a display keyword, it can also be used as an index for search-linked advertising or for marketing purposes. In this regard, with paid search advertising, if there is no relationship between the conversion query and the product where the conversion occurred, it is inappropriate for the conversion query to be used for the above purpose and may result in noise. be.

For example, suppose that a conversion query entered by a user hits a search-linked advertisement associated with an index that has a completely different meaning but is similar in wording. There is a possibility that a user may become interested in the search-linked advertising product by chance and convert. Conversion queries in this case are likely not very useful. Therefore, the query shaping system 1 estimates the relevance between a conversion query and an EC service based on a relevance estimation model using machine learning.

FIG. 10 is a diagram illustrating an example of a relevance estimation model in the second embodiment. In the second embodiment, training data including words related to EC services is learned by the relevance estimation model M2. In FIG. 10, the words included in the training data are described as training queries, but the training queries are words used as virtual queries. The training query does not have to be an actually input query.

For example, training queries may include various words such as queries actually entered in an EC service, product titles, product attributes, words included in detailed product descriptions, and keywords specified as product indexes. It's okay. The training data is a pair of a training query as input data and information regarding relevance. Based on this training data, learning of the relevance estimation model M2 is performed so as to be able to estimate whether a word included in the input data is related to the EC service.

For example, when a conversion query is input, the relevance estimation model M2 calculates the feature amount of the conversion query. The relevance estimation model M2 outputs whether the conversion query is relevant to the EC service based on the calculated feature amount. In the second embodiment, a case will be explained in which a first value indicating that the conversion query is relevant to the EC service or a second value indicating that the conversion query is not relevant to the EC service is output. However, instead of these binary values, a score indicating the degree of relevance may be output.

For example, the advertiser terminal 20 registers a conversion query that is estimated to be related to the EC service in the search service. Even if a conversion occurs, the advertiser terminal 20 does not register in the search service a conversion query that is estimated to be unrelated to the EC service. As described above, in the second embodiment, relevance is estimated using the relevance estimation model M2. Hereinafter, details of the second embodiment will be described. In the second embodiment, descriptions of the same points as in the first embodiment will be omitted.

[2-1. Functions realized in the second embodiment]
FIG. 11 is a diagram illustrating an example of functions realized in the second embodiment. In the second embodiment, descriptions of functions similar to those in the first embodiment will be omitted. In the second embodiment, in addition to the functions described in the first embodiment, the advertiser terminal 20 includes a relevance estimation unit 205. The relevance estimation unit 205 is realized by the control unit 11.

The data storage unit 200 of the second embodiment stores the relevance estimation model M2. The relevance estimation model M2 includes a program part that executes relevance estimation processing and a parameter part that is adjusted through learning. For example, the parameters are weights and biases. The parameters may be those that are compatible with the machine learning method used as the relevance estimation model M2. In the second embodiment, since binary classification is performed, a machine learning method capable of binary classification may be used.

In the second embodiment, a relevance estimation model M2 using a neural network will be taken as an example. For example, the advertiser terminal 20 configures the relevance estimation model M2 to output a first value indicating that the EC service is relevant when training data regarding words that are relevant to the EC service is input. Execute learning. For example, error backpropagation or gradient descent may be used in learning. Learning of the relevance estimation model M2 may be performed not by the advertiser terminal 20 but by another computer.

Note that the relevance estimation model M2 can use various machine learning methods other than neural networks. In the second embodiment, since it is sufficient to be able to estimate the presence or absence of a relationship or the probability thereof, any machine learning method that is used for some sort of classification in the field of natural language processing may be used. For example, machine learning techniques can be used to estimate the genre of a document. The relevance estimation model M2 may be a supervised learning model, a semi-supervised learning model, or an unsupervised learning model. The machine learning method exemplified in the description of the query shaping model M1 of the first embodiment may be used in the relevance estimation model M2.

In the second embodiment as well, the acquisition unit 201 acquires a conversion query in which a conversion occurred in an EC service due to a search-linked advertisement. The EC service is an example of a predetermined service. Therefore, the portion described as EC service can be read as a predetermined service. In the second embodiment, the predetermined service is a service different from the search service that displays search-linked advertisements. Examples other than EC services are as described in the first embodiment.

The relevance estimation unit 205 estimates the relevance between the conversion query and the EC service based on the relevance estimation model M2. Relevance is the correlation between the conversion query and the EC service. For example, the number of hits when searching within an EC service using a conversion query is one indicator of relevance. The higher the relevance, the higher the number of hits; the lower the relevance, the lower the number of hits. The relationship is expressed by a binary value indicating whether or not there is a relationship, or by a score indicating the strength of the relationship. The score is not binary, but has an intermediate value. For example, the higher the score, the stronger the relationship.

The relevance estimation model M2 is an example of a learning model in which training queries that are relevant to the EC service have been learned. Therefore, the portion describing the relevance estimation model M2 can be read as a learning model. The training query may be any query that can be used as training data. When training data, which is a training query, is input, the relevance estimation model M2 is trained to output that it is relevant to the EC service.

For example, the relevance estimation unit 205 calculates the relevance based on the relevance estimation model M2 that has learned training queries that are actually input in the search service and have relevance to the EC service. Estimate. In the second embodiment, similarly to the first embodiment, a case will be described in which morphological analysis is executed on the search service side. Therefore, in order to use the data as training data for the relevance estimation model M2, it is assumed that it is necessary to perform query shaping.

In the second embodiment, a conversion query that has been subjected to query shaping by the query shaping unit 203 described in the first embodiment is used as training data for the relevance estimation model M2. Since some of these conversion queries are unrelated, the unrelated and formatted conversion queries are manually excluded by the EC service operator.

Note that the training data for the relevance estimation model M2 may not be actually input in the search service, but may be a training query manually created by the administrator of the EC service. Further, the query shaping unit 203 of the second embodiment may format a conversion query that is estimated to be relevant based on the query shaping model M1. Conversion queries that are estimated to be unrelated may not be subject to query formatting.

The processing execution unit 204 of the second embodiment executes a predetermined process based on the relevance estimated by the relevance estimation unit 205. The predetermined process may be any process similar to the first embodiment. For example, the predetermined process may be a process of recording a conversion query estimated to be relevant in the data storage unit 200, or a process of outputting a conversion query estimated to be relevant externally. There may be. The predetermined process may be a process in which a conversion query that is estimated to be relevant is targeted for query shaping.

For example, the processing execution unit 204 executes, as a predetermined process, a registration process in which a conversion query that is estimated to be relevant is registered in a search service. The processing execution unit 204 executes, as a registration process, a process of registering the formatted conversion query, which is estimated to be relevant and has been formatted by the query formatting unit 203, in the search service. The processing execution unit 204 differs from the first embodiment in that it targets conversion queries that are estimated to be relevant, but is similar in other respects. When a relevance score is output, a score equal to or greater than a threshold value corresponds to relevance.

[2-2. Processing executed in second embodiment]
FIG. 12 is a diagram illustrating an example of processing executed in the second embodiment. The processing from S21 to S28 in FIG. 12 is similar to the processing from S1 to S8 in the first embodiment.

The advertiser terminal 20 acquires a conversion query whose relevance is to be estimated from the EC server 10 (S29). The advertiser terminal 20 estimates the relevance between the conversion query acquired in S29 and the EC service based on the learned relevance estimation model M2 (S30). If it is estimated that there is no relevance (S30: N), this process ends. On the other hand, if it is estimated that there is a relationship (S30: Y), the processes of S31 to S33, which are similar to the processes of S10 to S12 of the first embodiment, are executed.

The query shaping system 1 of the second embodiment estimates the relevance between the conversion query and the EC service based on the relevance estimation model M2. The query shaping system 1 executes predetermined processing based on the relevance. This makes it possible to estimate the relationship between the conversion query and the EC service. Therefore, predetermined processing can be executed based on conversions that are related to the EC service, making it easier to achieve the purpose of using conversion queries. For example, it is possible to prevent a conversion query that has no relevance to the EC service from being displayed in the display area A20 as part of a search-linked advertisement. For example, it is possible to prevent search-linked advertisements that are unrelated to the EC service from being displayed, thereby reducing the exposure of the EC service to reputation risk.

Furthermore, in the query shaping system 1, the EC service is a service different from the search service that displays search-linked advertisements. As a result, even if conversion occurs across services, it is possible to estimate the relationship between the conversion query and the EC service.

In addition, the query shaping system 1 executes a registration process of registering a conversion query estimated to be relevant in a search service as a predetermined process. As a result, even if conversion occurs across services, conversion queries related to the EC service can be registered in the search service.

In addition, the query formatting system 1 executes a process of registering a formatted conversion query that is estimated to be relevant in a search service as a registration process. As a result, even if a conversion occurs across services, a conversion query that is relevant to the EC service and has been appropriately formatted can be registered in the search service.

In addition, the query shaping system 1 calculates the relevance based on the relevance estimation model M2 that has learned training queries that are actually input in the search service and have relevance to other services. Estimate. Relevance can be estimated by the relevancy estimation model M2 created by effectively utilizing information actually input in the search service.

[3. Modified example]
The present disclosure is not limited to the first embodiment and second embodiment described above. Changes can be made as appropriate without departing from the spirit of the present disclosure.

[3-1. Modification example regarding the first embodiment]
FIG. 13 is a diagram illustrating an example of functions in a modified example of the first embodiment. In Modification 1-2, which will be described later, the advertiser terminal 20 includes a division execution unit 206. The division execution unit 206 is realized by the control unit 11. In the modification of the first embodiment, detokenization does not need to be performed, so the detokenization execution unit 202 is not shown in FIG.

[Modification 1-1]
For example, if the operator of an EC service knows what kind of morphological analysis is performed on the search service side, the query shaping model can be created by reproducing the morphological analysis performed on the search service side. M1 training data may be generated. In modification 1-1, it is assumed that the operator of the EC service has acquired the same morphological analysis tool as the search service.

The query shaping unit 203 of Modification Example 1-1 determines the relationship between the training query that has been changed in accordance with the change of the conversion query in the search service and the training query before the change as training data. The modified conversion query is formatted based on the query formatting model M1 learned as follows. In Modification 1-1, as in the first embodiment, morphological analysis of a conversion query corresponds to a change to the conversion query. A training query before morphological analysis is performed may be obtained in the same manner as in the first embodiment.

For example, in modification 1-1, the advertiser terminal 20 performs morphological analysis of the training query based on the same morphological analysis tool as the search service. The advertiser terminal 20 acquires the training query after being divided into morphemes as a modified training query. The advertiser terminal 20 acquires a pair of a training query after being divided into morphemes and a training query before being divided into morphemes as training data, and causes the query shaping model M1 to learn the pair. The query shaping model M1 of Modification 1-1 may be the same as the query shaping model M1 of Modification 1-3.

Note that in Modification 1-1, the method for acquiring training data is different from the first embodiment, but the learning method for the query shaping model M1 is the same as in the first embodiment. Further, the process of formatting a conversion query using the learned query formatting model M1 is also the same as in the first embodiment. Additionally, if the search service makes any other changes to the conversion query other than morphological analysis, the EC system operator must have acquired a change tool to make the other changes. shall be. Training data may be generated based on the modification tool.

The query shaping system 1 of Modification 1-1 performs query shaping by shaping a changed conversion query based on a query shaping model M1 that has learned training data that reproduces changes to a conversion query in a search service. Increases accuracy.

[Modification 1-2]
For example, in the first embodiment, a case has been described in which query shaping is executed by performing tokenization after detokenization, but the query shaping method is different from that in the first embodiment. This example is not limited to. Query formatting can be done by formatting the conversion query in an appropriate format in some way. In modification 1-2, contrary to the first embodiment, a case will be exemplified in which detokenization is performed after tokenization is performed.

FIG. 14 is a diagram illustrating an example of query formatting in Modification 1-2. The query shaping system 1 of Modification 1-2 includes a division execution unit 206. The division execution unit 206 divides the conversion query into predetermined division units. In the example of FIG. 14, a case will be described in which the division unit is one character, but the division unit may be a predetermined number of characters, and may be two or more characters. The division execution unit 206 divides the conversion query into characters and inserts separators between the characters.

The query shaping model M1 of Modification 1-2 is trained on training data including a training query divided into character units. In the example of FIG. 14, the training data includes the authenticity of separators between characters included in the training query. For example, the authenticity of the separator may be manually assigned by the EC service operator, or the separator included in the training query before division may be set to true, and the separator included in the training query before division may be set to true. A separator that is not present may be set to false.

When the divided training query is input, the query shaping model M1 outputs the truth or falsity of the separator in the training query. The parameters of the query shaping model M1 are adjusted so that the output truth/false becomes the truth/false that is the correct answer included in the training data. For example, the query shaping unit 203 formats the divided conversion query by estimating the appropriateness of division for the divided conversion query divided for each division unit based on the query shaping model M1.

In the example of FIG. 14, some of the pseudo-inserted separators are estimated to be false, so the query shaping unit 203 formats the conversion query by leaving only the separators estimated to be true. Shape. That is, the query shaping unit 203 shapes the conversion query by deleting separators that are estimated to be false from among the separators included in the conversion query after being divided into division units.

Query formatting system 1 of Modified Example 1-2 divides the conversion query into division units, so that no matter what state the conversion query is in, it can format the query after making it into the same state of one token. This increases the accuracy of query formatting for conversion queries.

[Modification 1-3]
For example, the query shaping model M1 may execute query shaping using other methods instead of tokenization as in the first embodiment or determining the authenticity of separators as in Modification 1-2. . In modification 1-3, it is assumed that the query shaping model M1 has learned training data that is a pair of a training query before shaping and a training query after shaping.

The training query after formatting is a query in an optimal state in which morphological analysis and the like have not been performed. The training query before shaping may be a query after morphological analysis has been performed using some morphological analysis tool. For example, as in Modification 1-1, a training query before formatting may be obtained by performing morphological analysis using another morphological analysis tool rather than the same morphological analysis tool as that used on the search service side.

For example, when the advertiser terminal 20 inputs the training query before shaping included in the training data to the query shaping model M1, the training query after shaping included in the training data is output from the query shaping model M1. The query shaping model M1 is trained as follows. The query shaping unit 203 of modification example 1-3 inputs a conversion query to the query shaping model M1 in which the relationship between the training query before shaping and the training query after shaping has been learned as training data, The conversion query is formatted by obtaining the formatted conversion query output from the query formatting model M1.

For example, the query shaping unit 203 inputs a conversion query to the query shaping model M1. The query shaping model M1 calculates the feature amount of the input conversion query, and outputs a formatted conversion query according to the calculated feature amount. In the first embodiment, a formatted conversion query is output after detokenization is performed, but in modification 1-3, detokenization is not performed and the conversion query is divided into morphemes. The query is input as is to the query shaping model M1.

The query formatting system 1 of Modification 1-3 formats the conversion query based on the query formatting model M1 in which the relationship between the training query before formatting and the training query after formatting is learned as training data. . This allows you to properly format your conversion queries.

[Modification 1-4]
For example, position information regarding the position of a token included in a training query may be learned in the query shaping model M1. The location information itself may be any information, for example, a BIO (Begin Inside Outside) tag or a BIOES (Begin Inside Outside End Single) tag. In modification 1-4, a case where the location information is a BIOES tag will be explained. The training data is a pair of a training query and location information where the first character of each token is given a B tag and the last character is given an E tag when the training query is properly tokenized. Become.

The query shaping unit 203 of Modified Example 1-4 inputs a conversion query to the query shaping model M1 in which position information regarding the position of a token included in a training query has been learned as training data, and inputs a conversion query to the query shaping model M1, which is output from the learning model. Shape conversion queries based on location. The query shaping model M1 adds a B tag to a character estimated to be the first character of a token, and an E tag to a character estimated to be the last character of a token, for an input conversion query. Intermediate characters may be given an I tag, and separators may be given an O tag. The query formatting unit 203 formats the conversion query so that the characters from the B tag to the E tag become one token.

The query shaping system 1 of Modification 1-4 shapes the conversion query based on the query shaping model M1 in which position information regarding the position of the token included in the training query is learned as training data. This allows you to properly format your conversion queries.

[Other modifications of the first embodiment]
For example, in the first embodiment, a case has been described in which the conversion query is changed on the search service side and query shaping is executed on the EC service side, but the conversion query does not need to be changed on the search service side. For example, a user may enter a conversion query without any spaces, or may enter a conversion query with spaces in unnatural positions. In this case as well, the query shaping unit 203 may reformat the unnatural conversion query input by the user by executing the same process as in the first embodiment.

Note that when formatting an unnatural conversion query input by the user, it is assumed that the conversion query input by the user is displayed as is in the display area A20. If a display keyword is associated with the search-linked advertisement, this display keyword is displayed in the display area A20. Even if the user inputs an unnatural conversion query, a natural display keyword shaped by the query shaping unit 203 can be displayed in the display area A20.

[3-2. Modification example regarding second embodiment]
A modification of the second embodiment will be described.

[Modification 2-1]
For example, in the second embodiment, a case has been described in which a training query actually input in a search service is learned by the relevance estimation model M2. A training query may be learned. The relevance estimation model differs from the second embodiment in whether the training query used as training data is actually input in a search service or in an EC service. The learning method of M2 itself is the same as in the second embodiment.

The relevance estimating unit 205 of Modification 2-1 estimates the relevance based on the relevance estimation model M2 that has learned the training query actually input in the EC service. Although the training query used in learning the relevance estimation model M2 is different from the second embodiment, the process of estimating the relevance based on the trained relevance estimation model M2 is the same as in the second embodiment. be. It is assumed that the training queries actually input in the EC service are stored in the data storage unit 100 as a usage log of the EC service.

The query shaping system 1 of Modification 2-1 estimates the relevance based on a learning model learned from a training query actually input in an EC service. Thereby, the relevance estimation model M2 learns the training query actually input in the EC service, thereby increasing the accuracy of relevance estimation.

[Modification 2-2]
For example, the relevance estimation unit 205 calculates a first training query that has relevance to the EC service and a second training query that has no relevance to the EC service based on the relevance estimation model M2 that has been learned. The relevance may be estimated by That is, the relevance estimation model M2 may learn not only the first training query that has relevance to the EC service but also the second training query that has no relevance to the EC service.

The second training query that has no relevance to the EC service may be manually prepared by the operator of the EC service. The output corresponding to the second training query is a second value indicating that it has no relevance to the EC service. The learning method based on the first training query is as described in the second embodiment. For example, when the second training query is input to the relevance estimation model M2, the relevance estimation model M2 is trained so as to output that there is no relevance.

The query shaping system 1 of Modification 2-2 uses a relevance estimation model that has learned a first training query that has relevance to the EC service and a second training query that has no relevance to the EC service. By estimating the relevance based on M2, the accuracy of estimating the relevance increases.

[Modification 2-3]
For example, in the second embodiment, a case has been described in which it is estimated whether the conversion query has some relevance to the EC service, but even if the conversion query has some relevance to the EC service, the conversion It may have nothing to do with the product. Therefore, the relationship between the conversion query and the converted product that caused the conversion may be estimated. The meaning of relevance is as explained in the second embodiment. For example, when a conversion query is input in an EC service, the more products that are hit in the search, the higher the relevance.

The relevance estimation unit 205 of Modified Example 2-3 calculates the conversion query and the conversion in the EC service based on the relevance estimation model M2 in which the training query that is relevant to the training product in the EC service has been learned. Estimate the relevance of the conversion product that occurred. Training products are an example of training items. A conversion product is an example of a conversion item. Therefore, the parts described as training products or conversion products can be read as training items or conversion items, respectively.

The item is not limited to a product, and may be anything that is a target of search-linked advertising. For example, in the case of a travel reservation service, accommodation facilities correspond to items. For example, in the case of a video distribution service, content such as videos or songs corresponds to the item. Although this embodiment differs from the second embodiment in that the object of relevance estimation is a conversion product rather than the entire EC service, the general mechanism of the learning method of the relevance estimation model M2 and the processing at the time of estimation are the same. be.

For example, the training data for the relevance estimation model M2 includes, as input data, a training query and product information (e.g., title, attributes, all or part of the detailed description, product name, or other keywords) and. The training data includes a first value indicating relevance as output data. The relevance estimation model M2 is trained so that when a combination of a training query and product information is input, a first value indicating that there is relevance is output. The relevance estimation unit 205 estimates the relevance by inputting the combination of the conversion query and the product information regarding the converted product into the relevance estimation model M2, and acquiring the output from the relevance estimation model M2.

The query shaping system 1 of Modified Example 2-3 uses a conversion query and a conversion in the EC service based on the relevance estimation model M2 in which a training query that is related to a training product in the EC service has been learned. Estimate the relevance of converted products. This makes it possible to estimate whether the combination of a conversion query and a conversion product is appropriate, thereby increasing the accuracy of estimating relevance.

[Modification 2-4]
For example, in Modification 2-3, the case where the relationship between a conversion query and a conversion product is estimated is explained, but more specifically, the relationship between a conversion query and an attribute of a conversion product is estimated. may be done. In modification 2-4, the product information described in modification 2-3 indicates attributes. Attributes are product classifications. Attributes are sometimes called categories or genres.

The relevance estimation unit 205 of Modified Example 2-4 calculates the conversion query and the first attribute related to the conversion product based on the relevance estimation model M2 in which the training query that is related to the first attribute related to the training product is learned. Estimate the relationship between the two attributes. The learning method for the relevance estimation model M2 is the same as in Modification 2-3. Product information in the description of Modification 2-3 may be read as attribute. The relevance estimation unit 205 estimates the relevance by inputting the combination of the conversion query and the attribute of the converted product into the relevance estimation model M2, and acquiring the output from the relevance estimation model M2.

The query shaping system 1 of Modified Example 2-4 is based on the relevance estimation model M2 in which the training query that is related to the first attribute related to the training product is learned, and the conversion query and the second attribute related to the conversion item. Estimate the relationship between attributes. This makes it possible to estimate whether the combination of attributes of the conversion query and conversion product is appropriate, thereby increasing the accuracy of estimating relevance.

[Modification 2-5]
For example, the predetermined process executed by the process execution unit 204 is not limited to the example described in the second embodiment. The processing execution unit 204 may execute a display process of displaying a conversion query based on the relevance as a predetermined process. In Modified Example 2-5, the processing execution unit 204 takes as an example a display process of displaying information related to relevance (for example, the presence or absence of relevance or a score) on the display unit 25 of the advertiser terminal 20. Display processing on computers other than 20 may also be executed. The processing execution unit 204 may perform a display process of sorting and displaying a plurality of conversion queries based on scores indicating relevance.

The query formatting system 1 of Modification 2-5 executes a display process of displaying a conversion query based on relevance as a predetermined process. This makes it easier for the EC service operator to understand the relationship between the conversion query and the EC service.

[Other variations regarding the second embodiment]
For example, the query shaping system 1 of the second embodiment is an example of a relevance estimation system. Therefore, the section describing the query shaping system 1 of the second embodiment can be read as the relevance estimation system. In the second embodiment, the query shaping function described in the first embodiment may not be provided. Therefore, the relevance estimation system may include only the acquisition unit 201, the processing execution unit 204, and the relevance estimation unit 205, without including the detokenization execution unit 202 and the query shaping unit 203.

[3-3. Other variations]
For example, the modifications described above may be combined.

For example, the query shaping system 1 may be applied to services other than EC services. Other services that cooperate with the search service may be any service, such as a travel reservation service, a video distribution service, a financial service, a communication service, or a payment service. Any other service may be used as long as it posts search-linked advertisements on a search service and generates some kind of conversion. The search-linked advertisement may be an advertisement within an EC service. In this case, search-linked advertisements are displayed when searching for products in the EC service.

For example, the functions described as being realized by each of the EC server 10, advertiser terminal 20, search server 30, and user terminal 40 may be realized by other computers, or may be shared among multiple computers. Good too. For example, the data stored in the data storage units 100, 200, 300, and 400 may be stored in a database server not shown in FIG.

[4. Additional notes]
For example, the query formatting system can also have the following configuration.
(1)
an acquisition unit that acquires a conversion query in which a conversion related to search advertising has occurred;
a query formatting unit that formats the conversion query based on a learning model in which training data related to query formatting is learned;
a processing execution unit that executes a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit;
A query formatting system including:
(2)
The search-linked advertisement is an advertisement for causing the user to access another service different from the search service in which the search-linked advertisement is displayed,
The acquisition unit acquires the conversion query in which a conversion occurred in the other service.
The query formatting system described in (1).
(3)
The processing execution unit executes a registration process of registering the formatted conversion query in the search service as the predetermined process.
The query formatting system described in (2).
(4)
In the search service, indexes and display keywords are registered for each search-linked advertisement,
In the search service, the search-linked advertisement is searched based on the index, and the display keyword is displayed together with the search-linked advertisement,
The processing execution unit executes a process of registering the formatted conversion query as the display keyword in the search service as the registration process.
The query formatting system described in (3).
(5)
The search service changes the conversion query entered at the time of search,
The acquisition unit acquires the changed conversion query that has been changed in the search service,
The query formatting unit formats the modified conversion query so that the modified conversion query returns to the state at the time of input.
The query formatting system according to any one of (2) to (4).
(6)
In the search service, the conversion query input at the time of search is changed by performing morphological analysis,
The query formatting unit formats the modified conversion query so that the morphemes included in the modified conversion query return to the state at the time of input.
The query formatting system described in (5).
(7)
The query shaping unit determines, as the training data, the relationship between the training query that has been changed in accordance with the change of the conversion query in the search service, and the training query before the change. formatting the modified conversion query based on the learned learning model;
The query formatting system according to (5) or (6).
(8)
The query shaping unit is configured to learn, as the training data, a relationship between the training query that is a modified training query that was actually input in the other service, and the training query before the modification. formatting the conversion query based on the learned model,
The query formatting system according to any one of (2) to (7).
(9)
The query shaping system further includes a detokenization execution unit that performs detokenization regarding the conversion query,
The query shaping unit formats the executed conversion query by performing tokenization on the executed conversion query on which the detokenization has been performed based on the learning model.
The query formatting system according to any one of (1) to (8).
(10)
The query shaping system further includes a division execution unit that divides the conversion query into predetermined division units,
The query shaping unit formats the divided conversion query by estimating suitability for dividing the divided conversion query divided into each division unit based on the learning model.
The query formatting system according to any one of (1) to (9).
(11)
The query shaping unit inputs the conversion query to the learning model in which the relationship between the training query before shaping and the training query after shaping is learned as the training data, and converts the conversion query from the learning model. formatting the conversion query by obtaining the output formatted conversion query;
The query formatting system according to any one of (1) to (10).
(12)
The query shaping unit inputs the conversion query to the learning model in which position information regarding the position of the token included in the training query has been learned as the training data, and inputs the conversion query into the position information output from the learning model. formatting the conversion query based on;
The query formatting system according to any one of (1) to (11).

1 Query shaping system, N network, 10 EC server, 11, 21, 31, 41 Control unit, 12, 22, 32, 42 Storage unit, 13, 23, 33, 43 Communication unit, 20 Advertiser terminal, 24, 44 Operation unit, 25, 45 Display unit, 30 Search server, 40 User terminal, M1 Query formatting model, M2 Relevance estimation model, 100, 200, 300, 400 Data storage unit, 101 Order processing unit, 201 Acquisition unit, 202 De tokenization execution unit, 203 query shaping unit, 204 processing execution unit, 205 relevance estimation unit, 206 division execution unit, 300 data storage unit, 301 search unit, 400 data storage unit, 401 display control unit, 402 operation reception unit , A20, A21 display area, DB1 product database, DB2 conversion database, DB3 advertising database, SC1 top screen, SC2 search result screen, SC3 product screen, SC4 order completion screen.

Claims (14)

an acquisition unit that acquires a conversion query in which a conversion related to search advertising has occurred;
a query formatting unit that formats the conversion query based on a learning model in which training data related to query formatting is learned;
a processing execution unit that executes a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit;
A query formatting system including: The search-linked advertisement is an advertisement for causing the user to access another service different from the search service in which the search-linked advertisement is displayed,
The acquisition unit acquires the conversion query in which a conversion occurred in the other service.
The query formatting system according to claim 1. The processing execution unit executes a registration process of registering the formatted conversion query in the search service as the predetermined process.
The query formatting system according to claim 2. In the search service, indexes and display keywords are registered for each search-linked advertisement,
In the search service, the search-linked advertisement is searched based on the index, and the display keyword is displayed together with the search-linked advertisement,
The processing execution unit executes a process of registering the formatted conversion query as the display keyword in the search service as the registration process.
The query formatting system according to claim 3. The search service changes the conversion query entered at the time of search,
The acquisition unit acquires the changed conversion query that has been changed in the search service,
The query formatting unit formats the modified conversion query so that the modified conversion query returns to the state at the time of input.
The query formatting system according to any one of claims 2 to 4. In the search service, the conversion query input at the time of search is changed by performing morphological analysis,
The query formatting unit formats the modified conversion query so that the morphemes included in the modified conversion query return to the state at the time of input.
The query formatting system according to claim 5. The query shaping unit determines, as the training data, the relationship between the training query that has been changed in accordance with the change of the conversion query in the search service, and the training query before the change. formatting the modified conversion query based on the learned learning model;
The query formatting system according to claim 5. The query shaping unit is configured to learn, as the training data, a relationship between the training query that is a modified training query that was actually input in the other service, and the training query before the modification. formatting the conversion query based on the learned model,
The query formatting system according to any one of claims 2 to 4. The query shaping system further includes a detokenization execution unit that performs detokenization regarding the conversion query,
The query shaping unit formats the executed conversion query by performing tokenization on the executed conversion query on which the detokenization has been performed based on the learning model.
A query formatting system according to any one of claims 1 to 4. The query shaping system further includes a division execution unit that divides the conversion query into predetermined division units,
The query shaping unit formats the divided conversion query by estimating suitability for dividing the divided conversion query divided into each division unit based on the learning model.
A query formatting system according to any one of claims 1 to 4. The query shaping unit inputs the conversion query to the learning model in which the relationship between the training query before shaping and the training query after shaping is learned as the training data, and converts the conversion query from the learning model. formatting the conversion query by obtaining the output formatted conversion query;
A query formatting system according to any one of claims 1 to 4. The query shaping unit inputs the conversion query to the learning model in which position information regarding the position of the token included in the training query has been learned as the training data, and inputs the conversion query into the position information output from the learning model. formatting the conversion query based on;
A query formatting system according to any one of claims 1 to 4. an acquisition step of acquiring a conversion query in which a conversion related to paid search advertising occurred;
a query shaping step of shaping the conversion query based on a learning model in which training data regarding query shaping has been learned;
a processing execution step of executing a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit;
Query formatting methods, including: an acquisition unit that acquires a conversion query in which a conversion related to search advertising has occurred;
a query shaping unit that formats the conversion query based on a learning model in which training data regarding query shaping has been learned;
a processing execution unit that executes a predetermined process based on the formatted conversion query that has been formatted by the query formatting unit;
A program that allows a computer to function as a computer.