JP6979986B2 - Information processing equipment, information processing methods and information processing programs - Google Patents

Description

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

In recent years, with the rapid spread of the Internet, techniques related to analysis using various information on the Internet have been provided. For example, a technique for predicting a trend in a predetermined field based on the analysis of a search query has been proposed.

Japanese Patent Publication No. 2015-534180

However, in the above-mentioned conventional technique, it is not always possible to appropriately extract a character string indicating an object belonging to a predetermined category. For example, in the above-mentioned prior art, an innovator in a predetermined category is identified, and a trend in the predetermined category is only predicted based on a search query searched by the specified innovator, and an object belonging to the predetermined category is shown. It is not always possible to properly extract a character string. For example, in the above-mentioned conventional technique, when the character string input as a search query is a new term that has just appeared in the world, it is difficult to specify the target indicated by the term. For example, even if an attempt is made to look up the meaning of a new term in a dictionary or the like, the term may not be listed in the dictionary or the like yet. In such a case, it is difficult to identify the object indicated by the new term, and it is also difficult to identify the category to which the object indicated by the new term belongs.

The present application has been made in view of the above, and an object thereof is to provide an information processing apparatus, an information processing method, and an information processing program capable of appropriately extracting a character string indicating an object belonging to a predetermined category. do.

The information processing apparatus according to the present application uses a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. An acquisition unit that acquires the second learning model that predicts the category to which the predetermined search query belongs from the generated second learning model, and a second learning model acquired by the acquisition unit. Is used to estimate the category to which the target indicated by the character string input as the search query belongs, and the extraction belonging to the target field from the character string based on the category estimated by the estimation unit. It is characterized by having an extraction unit for extracting a target character string indicating a target.

According to one aspect of the embodiment, there is an effect that a character string indicating an object belonging to a predetermined category can be appropriately extracted.

FIG. 1 is a diagram showing an example of information processing according to an embodiment. FIG. 2 is a diagram showing a configuration example of an information processing system according to an embodiment. FIG. 3 is a diagram showing a configuration example of the information processing apparatus according to the embodiment. FIG. 4 is a diagram showing an example of a query information storage unit according to an embodiment. FIG. 5 is a diagram showing an example of an unnecessary character string storage unit according to the embodiment. FIG. 6 is a diagram showing an example of an unnecessary category storage unit according to the embodiment. FIG. 7 is a diagram showing an example of a model information storage unit according to an embodiment. FIG. 8 is a diagram showing an example of information processing according to the embodiment. FIG. 9 is a flowchart showing an information processing procedure according to the embodiment. FIG. 10 is a diagram showing an example of a generation process of the first learning model according to the embodiment. FIG. 11 is a diagram showing an example of a generation process of the first learning model according to the embodiment. FIG. 12 is a diagram showing an example of a generation process of the second learning model according to the embodiment. FIG. 13 is a diagram showing a configuration example of the generator according to the embodiment. FIG. 14 is a diagram showing an example of the query information storage unit according to the embodiment. FIG. 15 is a diagram showing an example of a vector information storage unit according to an embodiment. FIG. 16 is a diagram showing an example of a classification definition storage unit according to an embodiment. FIG. 17 is a diagram showing an example of a category information storage unit according to an embodiment. FIG. 18 is a diagram showing an example of a model information storage unit according to an embodiment. FIG. 19 is a diagram showing an example of the first learning model according to the embodiment. FIG. 20 is a diagram showing an example of the second learning model according to the embodiment. FIG. 21 is a flowchart showing a generation processing procedure of the first learning model according to the embodiment. FIG. 22 is a flowchart showing a generation processing procedure of the second learning model according to the embodiment. FIG. 23 is a diagram showing an example of the hardware configuration of the computer that executes the program.

Hereinafter, an information processing apparatus, an information processing method, and an embodiment for implementing an information processing program (hereinafter referred to as “embodiments”) according to the present application will be described in detail with reference to the drawings. Note that this embodiment does not limit the information processing apparatus, information processing method, and information processing program according to the present application. Further, in each of the following embodiments, the same parts are designated by the same reference numerals, and duplicate description will be omitted.

[1. Embodiment]
[1-1. An example of information processing]
First, an example of information processing according to an embodiment will be described with reference to FIG. FIG. 1 is a diagram showing an example of information processing according to an embodiment. The information processing according to the embodiment is performed by the information processing apparatus 100 shown in FIG. In the example shown in FIG. 1, the information processing apparatus 100 indicates a target belonging to the fashion field (category related to fashion), which is the target field, from among the miscellaneous character strings included in the inflow search query group flowing into the fashion site. Extract the character string.

Prior to the description of FIG. 1, the configuration of the information processing system according to the embodiment will be described with reference to FIG. FIG. 2 is a diagram showing a configuration example of an information processing system according to an embodiment. As shown in FIG. 2, the information processing system 1 includes a generation device 50, an information processing device 100, and a service server 200. The generation device 50, the information processing device 100, and the service server 200 are connected to each other via a predetermined network N so as to be communicable by wire or wirelessly. The information processing system 1 shown in FIG. 2 may include an arbitrary number of generation devices 50, an arbitrary number of information processing devices 100, and an arbitrary number of service servers 200.

The generation device 50 is generated by using a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. It is an information processing device that generates a second learning model that predicts a category to which a predetermined search query belongs from a predetermined search query, which is a two-learning model. The details of the generation process of the second learning model by the generation device 50 will be described later. The generation device 50 transmits the model data MDT2 of the second learning model to the information processing device 100 in response to the request of the information processing device 100.

The information processing device 100 is a server device that extracts a target character string indicating an extraction target belonging to a target field. In the example shown in FIG. 1, the information processing apparatus 100 extracts a character string indicating an object belonging to the fashion field (hereinafter, also referred to as a fashion category) from the character strings included in the inflow search query that has flowed into the fashion-related site. do.

Here, in general, the inflow search query that has flowed into the site related to a predetermined category within a predetermined period includes various search queries. Specifically, the inflow search query includes general terms, site names, etc. that are different from the character strings indicating the target belonging to the predetermined category. As described above, since the inflow search query contains many character strings different from the character strings indicating the target belonging to the predetermined category, it is necessary to remove these character strings.

For example, in the inflow search query that has flowed into a fashion-related site, general terms such as ladies', popularity, coordination, size, etc., and inflow as a character string different from the character string indicating the target belonging to the fashion category. Contains a string indicating the name of the previous site. These general terms and character strings indicating the name of the inflow destination site can be registered in advance as unnecessary character strings and removed on a dictionary basis. On the other hand, if the character string entered as an inflow search query contains a new term that has just appeared in the world (hereinafter, also referred to as an unknown term), it is difficult to remove the term. For example, the inflow search query that has flowed into a fashion site includes a character string that indicates a person name, a brand name, a store name, ..., etc. as a character string different from the character string that indicates a target belonging to a fashion category. There is. There are many unknown terms in the character strings indicating these personal names, brand names, store names, etc., because new personal names, new brand names, new store names, etc. appear one after another in the world. Therefore, it is difficult to register all the character strings indicating the person name, brand name, store name, ..., Etc. in advance. Therefore, conventionally, it has been difficult to remove a character string other than a character string indicating a target belonging to a predetermined category from an inflow search query.

In addition, the character string to be extracted may be an unknown term. For example, the inflow search query that has flowed into a fashion-related site includes a new term (hereinafter, also referred to as an unknown term) that is a character string indicating an object belonging to a fashion category and has just appeared in the world. In particular, in fields such as the fashion field where new terms indicating new targets appear one after another, it is not uncommon for the character string to be extracted to be an unknown term. Therefore, in order to determine whether or not the unknown term is a character string to be extracted, a method of looking up the meaning of the term in a dictionary or the like can be considered, but the unknown term that has just appeared in the world is still published in the dictionary or the like. It may not be. In such a case, since it is difficult to specify the object indicated by the unknown term, it is also difficult to specify the category to which the object indicated by the unknown term belongs. That is, conventionally, since it is difficult to specify the category to which the object indicated by the unknown term belongs, it is difficult to extract the unknown term as a character string belonging to the target field.

Therefore, in the information processing apparatus 100 according to the present invention, the object indicated by the character string input as the inflow search query belongs to the second learning model that predicts the category to which the predetermined search query belongs from the predetermined search query. Estimate the category. As a result, the information processing apparatus 100 according to the present invention can estimate the category to which the object indicated by the unknown term belongs even when the character string input as the inflow search query is an unknown term. It is possible to accurately extract a character string indicating an object belonging to a predetermined category. Specifically, in the information processing apparatus 100, when a category indicating a non-target field different from the target field is estimated for an unknown term, the unknown term is other than a character string indicating a target belonging to the target category. It can be presumed to be a character string. Therefore, the information processing apparatus 100 can remove the unknown term as an unnecessary character string. On the other hand, even if the character string input as the inflow search query is an unknown term, the information processing apparatus 100 uses the unknown term when a category indicating a target field is estimated for the unknown term. It can be estimated that it is a target character string belonging to the target field. Therefore, the information processing apparatus 100 can extract the unknown term as a target character string. Therefore, the information processing apparatus 100 according to the present invention can appropriately extract a character string indicating an object belonging to a predetermined category.

The service server 200 is a server device that provides a site related to a target field. Specifically, the service server 200 stores information about the inflow search query that has flowed into the site related to the target field. In the example shown in FIG. 1, the service server 200 stores information about an inflow search query that has flowed into a site related to the fashion field. The service server 200 transmits information regarding an inflow search query to the information processing device 100 in response to a request from the information processing device 100. The service server 200 may provide sites related to various fields, not limited to the fashion field.

From here, the flow of information processing will be described with reference to FIG. In FIG. 1, the information processing apparatus 100 transmits a request for acquiring information regarding an inflow search query that has flowed into a fashion site to a service server 200. The service server 200 transmits information regarding an inflow search query to the information processing device 100 in response to a request from the information processing device 100.

The information processing apparatus 100 acquires information regarding an inflow search query that has flowed into a fashion site from the service server 200. Specifically, the information processing apparatus 100 acquires each character string input as an inflow search query (step S1).

In the example shown on the left side of FIG. 1, the information processing apparatus 100 acquires the character string “Ladies unknown term L1” input as the inflow search query Q100. Here, it is assumed that the "unknown term L1" is a new term (for example, a fashion term) that has just appeared in the world and has not yet been published in a dictionary or the like.

Further, the information processing apparatus 100 acquires the character string "unknown term L1 M size store name T1" input as the inflow search query Q200. The "store name T1" is a phrase indicating the name of a specific store (for example, a specific apparel maker), and is, for example, a phrase indicating a new store name that has just appeared in the world.

Further, the information processing apparatus 100 acquires the character string "Y-shirt 20s" input as the inflow search query Q300.

Further, the information processing apparatus 100 acquires the character string "Y-shirt person name M1" input as the inflow search query Q400. The "personal name M1" is a phrase indicating the name of a specific person (for example, a fashion model, an entertainer, etc.), and is, for example, a phrase indicating a new personal name that has just appeared in the world.

Further, the information processing apparatus 100 acquires the character string "unknown term L2 coordination" input as the inflow search query Q500. Here, it is assumed that the "unknown term L2" is a new term (for example, a fashion term) that has just appeared in the world and has not yet been published in a dictionary or the like, like the "unknown term L1". ..

Further, the information processing apparatus 100 acquires the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600. The "brand name B1" is a phrase indicating the name of a specific parallel brand, and is, for example, a phrase indicating a new brand name that has just appeared in the world.

Subsequently, when the information processing apparatus 100 acquires the inflow search query group, the information processing apparatus 100 removes the unnecessary character string registered as an unnecessary character string from each character string input as each search query included in the acquired inflow search query group. Get the first character string. For example, the information processing apparatus 100 acquires a list of unnecessary character strings registered as unnecessary character strings. Subsequently, the information processing apparatus 100 refers to the acquired list of unnecessary character strings, and whether the unnecessary character strings are included in each character string input as each search query included in the acquired inflow search query group. Judge whether or not. Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the information processing apparatus 100 acquires the first character string obtained by removing the unnecessary character string from the character string input as the search query.

For example, the information processing apparatus 100 refers to the list of unnecessary character strings, and determines that the character string "ladies unknown term L1" input as the inflow search query Q100 includes the unnecessary character string "ladies". Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the first character obtained by removing the unnecessary character string "ladies" from the character string "ladies unknown term L1" input as the inflow search query Q100. The column "unknown term L1" (first character string L1) is acquired.

Further, the information processing apparatus 100 refers to the list of unnecessary character strings, and includes the unnecessary character string "M size" in the character string "unknown term L1 M size store name T1" input as the inflow search query Q200. It is determined that the information is processed. Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the information processing device 100 selects the unnecessary character string "M size" from the character string "unknown term L1 M size store name T1" input as the inflow search query Q200. The removed first character string "unknown term L1 store name T1" (first character string L2) is acquired.

Further, the information processing apparatus 100 refers to the list of unnecessary character strings, and determines that the character string "Y-shirt 20s" input as the inflow search query Q300 includes the unnecessary character string "20s". .. Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the first character string "Y-shirt 20s" input as the inflow search query Q300 is removed from the unnecessary character string "20s". The character string "Y shirt" (first character string L3) is acquired.

Further, the information processing apparatus 100 refers to the list of unnecessary character strings, and determines that the character string "unknown term L2 coordination" input as the inflow search query Q500 includes the unnecessary character string "coordination". Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the first character obtained by removing the unnecessary character string "corde" from the character string "unknown term L2 coordination" input as the inflow search query Q500. The column "unknown term L2" (first character string L5) is acquired.

Further, when the information processing apparatus 100 refers to the list of unnecessary character strings, the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600 includes "popularity" which is an unnecessary character string. judge. Subsequently, when the information processing apparatus 100 determines that the unnecessary character string is included, the information processing apparatus 100 removes the unnecessary character string "popularity" from the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600. The first character string "unknown term L2 brand name B1" (first character string L6) is acquired.

Subsequently, when the information processing apparatus 100 acquires the first character string, the information processing apparatus 100 learns the characteristics of the plurality of search queries as if the plurality of search queries input by the same user within a predetermined time have similar characteristics. A second learning model (second learning model M2) that predicts a category to which a predetermined search query belongs is acquired from a predetermined search query, which is a second learning model generated by using the first learning model. Subsequently, when the information processing apparatus 100 acquires the second learning model, the information processing apparatus 100 estimates the category to which the object indicated by the first character string belongs by using the second learning model. The information processing apparatus 100 estimates a plurality of categories to which the object indicated by the first character string belongs. Specifically, the information processing apparatus 100 outputs the probability that the object indicated by the first character string belongs to each category for each category. More specifically, in the information processing apparatus 100, by inputting the first character string as the input information of the second learning model M2, the target indicated by the first character string as the output information of the second learning model M2 is each. The probability of belonging to a category is output for each category (step S2).

In the example shown in the middle of FIG. 1, when the information processing apparatus 100 acquires the first character string L1, the information processing apparatus 100 inputs the first character string L1 as the input information of the second learning model M2, so that the second learning model M2 As output information, the probability that the target indicated by the first character string L1 belongs to each category is output for each category. For example, the information processing apparatus 100 outputs the probability that the object represented by the first character string L1 belongs to the fashion category C1 (hereinafter, also referred to as fashion category C1) as 100%, and the probability that the object belongs to another category is 0%. ..

Further, when the information processing apparatus 100 acquires the first character string L2, the information processing apparatus 100 inputs the first character string L2 as the input information of the second learning model M2, thereby inputting the first character string as the output information of the second learning model M2. The probability that the object indicated by L2 belongs to each category is output for each category. For example, the information processing apparatus 100 has a 70% probability that the object indicated by the first character string L2 belongs to the fashion category C1, a 30% probability that the object belongs to the category C4 related to the store name, and a 0% probability that the object belongs to another category. Output.

Further, when the information processing apparatus 100 acquires the first character string L3, the information processing apparatus 100 inputs the first character string L3 as the input information of the second learning model M2, thereby inputting the first character string as the output information of the second learning model M2. The probability that the object indicated by L3 belongs to each category is output for each category. For example, the information processing apparatus 100 outputs 100% the probability that the object indicated by the first character string L3 belongs to the fashion category C1 and 0% the probability that the object belongs to another category.

Further, when the information processing apparatus 100 acquires the first character string L4, the information processing apparatus 100 inputs the first character string L4 as the input information of the second learning model M2, thereby inputting the first character string as the output information of the second learning model M2. The probability that the object indicated by L4 belongs to each category is output for each category. For example, the information processing apparatus 100 outputs that the probability that the object indicated by the first character string L4 belongs to the fashion category C1 is 50%, the probability that the object belongs to the category C2 related to the personal name is 50%, and the probability that the object belongs to another category is 0%. do.

Further, when the information processing apparatus 100 acquires the first character string L5, the information processing apparatus 100 inputs the first character string L5 as the input information of the second learning model M2, thereby inputting the first character string as the output information of the second learning model M2. The probability that the object indicated by L5 belongs to each category is output for each category. For example, the information processing apparatus 100 outputs 100% the probability that the object indicated by the first character string L5 belongs to the fashion category C1 and 0% the probability that the object belongs to another category.

Further, when the information processing apparatus 100 acquires the first character string L6, the information processing apparatus 100 inputs the first character string L6 as the input information of the second learning model M2, thereby inputting the first character string as the output information of the second learning model M2. The probability that the object indicated by L6 belongs to each category is output for each category. For example, in the information processing apparatus 100, the probability that the object indicated by the first character string L6 belongs to the fashion category C1 is 60%, the probability that the object belongs to the category C3 related to the brand name is 40%, and the probability that the object belongs to the other category is 0%. Output.

Subsequently, when the information processing apparatus 100 estimates the category to which the object indicated by the first character string belongs, the information processing apparatus 100 extracts from the first character string to belong to the fashion category C1 which is the target field, based on the estimated category. The target character string indicating the target is extracted (step S3). In the example shown on the right side of FIG. 1, the information processing apparatus 100 extracts the first character string "unknown term L1" (first character string L1) from the first character string as the target character string W1. Further, the information processing apparatus 100 extracts the first character string "Y-shirt" (first character string L3) from the first character string as the target character string W2. Further, the information processing apparatus 100 extracts the first character string "unknown term L2" (first character string L5) from the first character string as the target character string W3.

Specifically, when the information processing apparatus 100 estimates the category to which the object indicated by the first character string belongs, the first character indicates whether or not the estimated category includes the fashion category C1 which is the target field. Judgment is made for each column. Subsequently, when the information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the estimated category, the information processing apparatus 100 extracts the first character string thereof. That is, the information processing apparatus 100 extracts a first character string indicating an object belonging to the fashion category C1. Subsequently, when the information processing apparatus 100 extracts the first character string indicating the object belonging to the fashion category C1, the first first character string is extracted as to whether or not the estimated category includes an unnecessary category registered as an unnecessary category. Judgment is made for each character string. Subsequently, when the information processing apparatus 100 determines that the estimated category does not include the unnecessary category registered as an unnecessary category, the information processing apparatus 100 extracts the first character string as the target character string. That is, the information processing apparatus 100 extracts the first character string other than the first character string indicating the target belonging to the unnecessary category as the target character string.

For example, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L1. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L1. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L1. Subsequently, the information processing apparatus 100 determines that the category estimated for the first character string L1 does not include an unnecessary category. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is not included, the first character string L1 is extracted as the target character string W1.

Further, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L2. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L2. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L2. Subsequently, the information processing apparatus 100 determines that the category C3 related to the store name, which is an unnecessary category, is included in the category estimated for the first character string L2. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is included, it is determined not to extract the first character string L2 as the target character string.

Further, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L3. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L3. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L3. Subsequently, the information processing apparatus 100 determines that the category estimated for the first character string L3 does not include an unnecessary category. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is not included, the first character string L3 is extracted as the target character string W2.

Further, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L4. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L4. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L4. Subsequently, the information processing apparatus 100 determines that the category C2 relating to the personal name, which is an unnecessary category, is included in the category estimated for the first character string L4. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is included, it is determined not to extract the first character string L4 as the target character string.

Further, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L5. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L5. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L5. Subsequently, the information processing apparatus 100 determines that the category estimated for the first character string L5 does not include an unnecessary category. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is not included, the first character string L5 is extracted as the target character string W3.

Further, the information processing apparatus 100 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L6. The information processing apparatus 100 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L6. Subsequently, since the information processing apparatus 100 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L6. Subsequently, the information processing apparatus 100 determines that the category C3 relating to the brand name, which is an unnecessary category, is included in the category estimated for the first character string L6. Subsequently, since the information processing apparatus 100 has determined that the unnecessary category is included, it is determined not to extract the first character string L6 as the target character string.

As described above, the information processing apparatus 100 uses a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. The second training model generated by using the above, which predicts the category to which the predetermined search query belongs from the predetermined search query, is acquired. Further, the information processing apparatus 100 uses the acquired second learning model to estimate the category to which the target indicated by the character string input as the search query belongs. Then, the information processing apparatus 100 extracts a target character string indicating an extraction target belonging to the target field from the character string based on the estimated category.

As a result, the information processing apparatus 100 can estimate the category to which the target indicated by the unknown term belongs even when the character string input as the search query is an unknown term, and therefore belongs to a predetermined category. The character string indicating the target can be extracted with high accuracy. For example, the information processing apparatus 100 can remove an unknown term as an unnecessary character string when a category indicating a non-target field different from the target field is estimated for the unknown term. Further, even if the character string input as the inflow search query is an unknown term, the information processing apparatus 100 uses the unknown term when a category indicating a target field is estimated for the unknown term. It can be extracted as a target character string belonging to the target field. Therefore, the information processing apparatus 100 according to the present invention can appropriately extract a character string indicating an object belonging to a predetermined category.

[1-2. Information processing device configuration]
Next, the configuration of the information processing apparatus 100 according to the embodiment will be described with reference to FIG. FIG. 3 is a diagram showing a configuration example of the information processing apparatus 100 according to the embodiment. As shown in FIG. 3, the information processing apparatus 100 includes a communication unit 110, a storage unit 120, and a control unit 130. The information processing device 100 includes an input unit (for example, a keyboard, a mouse, etc.) that receives various operations from the administrator of the information processing device 100, and a display unit (for example, a liquid crystal display, etc.) for displaying various information. You may have.

(Communication unit 110)
The communication unit 110 is realized by, for example, a NIC (Network Interface Card) or the like. Then, the communication unit 110 is connected to the network by wire or wirelessly, and for example, information is transmitted / received between the generation device 50 and the service server 200.

(Memory unit 120)
The storage unit 120 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. As shown in FIG. 3, the storage unit 120 includes a query information storage unit 121, an unnecessary character string storage unit 122, an unnecessary category storage unit 123, and a model information storage unit 124.

(Query information storage unit 121)
The query information storage unit 121 stores various information related to the search query that has flowed into the site related to the target field. For example, the query information storage unit 121 stores various information related to the search query that has flowed into the fashion-related site. FIG. 4 shows an example of the query information storage unit according to the embodiment. In the example shown in FIG. 4, the query information storage unit 121 has items such as "search query ID", "date and time", "inflow site name", and "character string".

The "search query ID" indicates identification information that identifies the search query that has flowed into the site related to the target field. "Date and time" indicates the date and time when the search query entered the site related to the target field. The "inflow site name" indicates the site name of the inflow destination to which the search query has flowed. "Character string" indicates a character string entered as a search query.

In the example shown in the first record of FIG. 4, the search query (search query Q1) identified by the search query ID "Q1" is a search query that has flowed into the site name "site name N1" that has flowed into the date and time "DT1". Show that. Further, the character string "site name N1 cosash" indicates that the character string input as the search query Q1 is "site name N1 cosash".

(Unnecessary character string storage unit 122)
The unnecessary character string storage unit 122 stores various information related to the unnecessary character string. FIG. 5 shows an example of an unnecessary character string storage unit according to the embodiment. In the example shown in FIG. 5, the unnecessary character string storage unit 122 has items such as “unnecessary character string ID” and “unnecessary character string”.

The "unnecessary character string ID" indicates identification information for identifying the unnecessary character string. "Unnecessary character string" indicates a character string registered as an unnecessary character string.

In the example shown in the first record of FIG. 5, it is shown that the unnecessary character string identified by the unnecessary character string ID “UL11” is “site name N1”. For example, "site name N1" indicates the site name of a fashion-related site.

(Unnecessary category storage unit 123)
The unnecessary category storage unit 123 stores various information related to the unnecessary category. FIG. 6 shows an example of an unnecessary category storage unit according to the embodiment. In the example shown in FIG. 6, the unnecessary category storage unit 123 has items such as “unnecessary category ID” and “unnecessary category”.

The "unnecessary category ID" indicates identification information for identifying the unnecessary category. "Unnecessary category" indicates a category registered as an unnecessary category.

In the example shown in the first record of FIG. 6, the unnecessary category "person name" indicates that the unnecessary category identified by the unnecessary category ID "C2" is a category related to a person name.

(Model information storage unit 124)
The model information storage unit 124 stores various information about the learning model generated by the generation device 50. FIG. 7 shows an example of the model information storage unit according to the embodiment. In the example shown in FIG. 7, the model information storage unit 124 has items such as "model ID" and "model data".

The "model ID" indicates identification information for identifying the learning model generated by the generation device 50. The "model data" indicates the model data of the learning model generated by the generation device 50. For example, "model data" stores data for converting a search query into the probability that the search query belongs to each category.

In the example shown in the first record of FIG. 7, the learning model identified by the model ID “M1” corresponds to the first model M1 described later. Further, the model data "MDT1" indicates model data (model data MDT1) of the first model M1 generated by the information processing apparatus 100.

The model data MDT1 includes an input layer into which a search query is input, an output layer, a first element which is any layer from the input layer to the output layer and belongs to a layer other than the output layer, and the first element and the first element. The output layer outputs a distributed representation of the search query input to the input layer according to the search query input to the input layer, including the second element whose value is calculated based on the weight of one element. As such, the information processing device 100 may be made to function.

Here, it is assumed that the model data MDT1 is realized by the regression model shown by "y = a1 * x1 + a2 * x2 + ... + ai * xi". In this case, the first element included in the model data MDT1 corresponds to input data (xi) such as x1 and x2. Further, the weight of the first element corresponds to the coefficient ai corresponding to xi. Here, the regression model can be regarded as a simple perceptron having an input layer and an output layer. When each model is regarded as a simple perceptron, the first element corresponds to any node of the input layer, and the second element can be regarded as the node of the output layer.

Further, it is assumed that the model data MDT1 is realized by a neural network having one or a plurality of intermediate layers such as DNN (Deep Neural Network). In this case, the first element included in the model data MDT1 corresponds to either the node of the input layer or the intermediate layer. Further, the second element corresponds to the node of the next stage, which is the node to which the value is transmitted from the node corresponding to the first element. Further, the weight of the first element corresponds to a connection coefficient which is a weight considered for the value transmitted from the node corresponding to the first element to the node corresponding to the second element.

The information processing apparatus 100 calculates the distributed representation using a model having an arbitrary structure such as the regression model and the neural network described above. Specifically, the model data MDT1 is set with a coefficient so as to output a distributed representation when a search query is input. The information processing apparatus 100 uses such model data MDT1 to calculate a distributed representation.

In the above example, the model data MDT1 is a model (hereinafter referred to as model X1) that outputs a distributed representation of the search query when the search query is input. However, the model data MDT1 according to the embodiment may be a model generated based on the result obtained by repeating the input / output of data to the model X1. For example, the model data MDT1 may be a model (hereinafter referred to as model Y1) trained to input a search query and output a distributed expression output by the model X1. Alternatively, the model data MDT1 may be a model trained to input a search query and output the output value of the model Y1.

Further, when the information processing apparatus 100 performs estimation processing using GAN (Generative Adversarial Networks), the model data MDT1 may be a model constituting a part of GAN.

In the example shown in the second record of FIG. 7, the learning model identified by the model ID “M2” corresponds to the second model M2 shown in FIG. Further, the model data "MDT2" indicates model data (model data MDT2) of the second model M2 generated by the information processing apparatus 100.

The model data MDT2 includes an input layer into which a search query is input, an output layer, a first element which is any layer from the input layer to the output layer and belongs to a layer other than the output layer, and the first element and the first element. An output layer that includes a second element whose value is calculated based on the weight of one element, and the probability that the search query input to the input layer belongs to each category according to the search query input to the input layer. The information processing apparatus 100 may be made to function so as to output from.

Here, it is assumed that the model data MDT2 is realized by the regression model shown by "y = a1 * x1 + a2 * x2 + ... + ai * xi". In this case, the first element included in the model data MDT2 corresponds to input data (xi) such as x1 and x2. Further, the weight of the first element corresponds to the coefficient ai corresponding to xi. Here, the regression model can be regarded as a simple perceptron having an input layer and an output layer. When each model is regarded as a simple perceptron, the first element corresponds to any node of the input layer, and the second element can be regarded as the node of the output layer.

Further, it is assumed that the model data MDT2 is realized by a neural network having one or a plurality of intermediate layers such as DNN (Deep Neural Network). In this case, the first element included in the model data MDT2 corresponds to either the node of the input layer or the intermediate layer. Further, the second element corresponds to the node of the next stage, which is the node to which the value is transmitted from the node corresponding to the first element. Further, the weight of the first element corresponds to a connection coefficient which is a weight considered for the value transmitted from the node corresponding to the first element to the node corresponding to the second element.

The information processing apparatus 100 calculates the probability that the search query belongs to each category by using a model having an arbitrary structure such as the regression model and the neural network described above. Specifically, in the model data MDT2, when a search query is input, a coefficient is set so as to output the probability that the search query belongs to each category. The information processing apparatus 100 uses such model data MDT2 to calculate the probability that the search query belongs to each category.

In the above example, the model data MDT2 is a model (hereinafter referred to as model X2) that outputs the probability that the search query belongs to each category when the search query is input. However, the model data MDT2 according to the embodiment may be a model generated based on the result obtained by repeating the input / output of data to the model X2. For example, the model data MDT2 may be a model trained to input a search query and output a probability output by the model X2 (hereinafter referred to as a model Y2). Alternatively, the model data MDT2 may be a model trained to input a search query and output the output value of the model Y2.

Further, when the information processing apparatus 100 performs estimation processing using GAN (Generative Adversarial Networks), the model data MDT2 may be a model constituting a part of GAN.

(Control unit 130)
Returning to the description of FIG. 3, the control unit 130 is a controller, and is stored in a storage device inside the information processing device 100 by, for example, a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). Various programs (corresponding to an example of an information processing program) are realized by executing the RAM as a work area. Further, the control unit 130 is a controller, and is realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

As shown in FIG. 3, the control unit 130 includes an acquisition unit 131, a processing unit 132, an estimation unit 133, and an extraction unit 134, and realizes or executes the information processing operation described below. The internal configuration of the control unit 130 is not limited to the configuration shown in FIG. 3, and may be any other configuration as long as it is configured to perform information processing described later.

(Acquisition unit 131)
The acquisition unit 131 acquires various types of information. Specifically, the acquisition unit 131 acquires an unnecessary character string which is a character string indicating an object different from the character string indicating the object belonging to the target field. For example, the acquisition unit 131 acquires a list of unnecessary character strings registered as unnecessary character strings. For example, the acquisition unit 131 acquires a list of unnecessary character strings registered by the administrator of the information processing apparatus 100. When the acquisition unit 131 acquires an unnecessary character string, the acquisition unit 131 stores the acquired unnecessary character string in the unnecessary character string storage unit 122.

In addition, the acquisition unit 131 acquires information about an unnecessary category, which is a category indicating a non-target field different from the target field. For example, the acquisition unit 131 acquires a list of unnecessary categories registered as unnecessary categories. For example, the acquisition unit 131 acquires a list of unnecessary categories registered by the administrator of the information processing apparatus 100. When the acquisition unit 131 acquires an unnecessary category, the acquisition unit 131 stores the acquired unnecessary category in the unnecessary category storage unit 123.

Further, the acquisition unit 131 acquires various information from an external information processing device. Specifically, the acquisition unit 131 uses a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. A second learning model that predicts the category to which the predetermined search query belongs is acquired from the generation device 50 from the predetermined search query. Subsequently, when the acquisition unit 131 acquires the second learning model, the acquisition unit 131 stores the acquired second learning model in the model information storage unit 124.

Further, the acquisition unit 131 uses a first learning model in which the characteristics of a plurality of search queries are learned so that the distributed expressions of a pair of search queries input by the same user within a predetermined time have similar characteristics. A second learning model that predicts a category to which a predetermined search query belongs is acquired from a predetermined search query.

Further, the acquisition unit 131 acquires a second learning model generated by using a first learning model that outputs a distributed representation of the predetermined search query as output information when a predetermined search query is input as input information. do.

Further, the acquisition unit 131 acquires a second learning model that outputs the probability that the search query belongs to a category as output information when the search query is input to the second learning model as input information.

Further, the acquisition unit 131 learns that a plurality of search queries including a character string delimited by a predetermined delimiter have similar characteristics as a plurality of search queries input by the same user within a predetermined time. By doing so, it is assumed that a plurality of search queries have similar characteristics, and it is a second learning model generated by using the first learning model in which the characteristics possessed by the plurality of search queries are learned, from a predetermined search query. Acquire a second training model that predicts the category to which a given search query belongs.

Further, the acquisition unit 131 is a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned, and is random. It is a second learning model generated by using the first learning model that learned the characteristics of the plurality of search queries by learning that the plurality of search queries extracted from the above have different characteristics. A second learning model that predicts the category to which a predetermined search query belongs is acquired from the search query of.

Further, the acquisition unit 131 is a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned, and is random. It is a second learning model generated by using the first learning model that learned the characteristics of a plurality of search queries by learning so that the distributed expressions of the pair of search queries extracted in the above are different. A second learning model that predicts the category to which a predetermined search query belongs is acquired from the search query of.

Further, the acquisition unit 131 is generated by using a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. In the second learning model, when the search query is input to the second learning model, the classification result of the distributed expression output by the second learning model is learned so as to correspond to the category to which the search query belongs. Then, the second learning model that predicts the category to which the predetermined search query belongs is acquired from the predetermined search query.

(Processing unit 132)
The processing unit 132 acquires the inflow search query that has flowed into the site related to the target field, and acquires the first character string obtained by removing the unnecessary character string registered as an unnecessary character string from the character string input as the inflow search query. do. Specifically, the processing unit 132 acquires information regarding the inflow search query that has flowed into the site related to the target field from the service server 200. More specifically, the processing unit 132 acquires each character string input as each inflow search query. When the processing unit 132 acquires the character string input as the inflow search query, the processing unit 132 stores the acquired character string in the query information storage unit 121.

In the example shown in FIG. 1, the processing unit 132 acquires information regarding the inflow search query that has flowed into the fashion site from the service server 200. Specifically, the processing unit 132 acquires each character string input as an inflow search query.

In the example shown on the left side of FIG. 1, the processing unit 132 acquires the character string “ladies unknown term L1” input as the inflow search query Q100. Here, it is assumed that the "unknown term L1" is a new term (for example, a fashion term) that has just appeared in the world and has not yet been published in a dictionary or the like.

Further, the processing unit 132 acquires the character string “unknown term L1 M size store name T1” input as the inflow search query Q200. The "store name T1" is a phrase indicating the name of a specific store (for example, a specific apparel maker), and is, for example, a phrase indicating a new store name that has just appeared in the world.

Further, the processing unit 132 acquires the character string "Y-shirt 20s" input as the inflow search query Q300.

Further, the processing unit 132 acquires the character string "Y-shirt person name M1" input as the inflow search query Q400. The "personal name M1" is a phrase indicating the name of a specific person (for example, a fashion model, an entertainer, etc.), and is, for example, a phrase indicating a new personal name that has just appeared in the world.

Further, the processing unit 132 acquires the character string "unknown term L2 coordination" input as the inflow search query Q500. Here, it is assumed that the "unknown term L2" is a new term (for example, a fashion term) that has just appeared in the world and has not yet been published in a dictionary or the like, like the "unknown term L1". ..

Further, the processing unit 132 acquires the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600. The "brand name B1" is a phrase indicating the name of a specific parallel brand, and is, for example, a phrase indicating a new brand name that has just appeared in the world.

Subsequently, when the inflow search query group is acquired, the processing unit 132 removes the unnecessary character string registered as an unnecessary character string from each character string input as each search query included in the acquired inflow search query group. Get the first character string. For example, the processing unit 132 refers to the unnecessary character string storage unit 122 to acquire a list of unnecessary character strings registered as unnecessary character strings. Subsequently, the processing unit 132 refers to the acquired list of unnecessary character strings, and determines whether or not the unnecessary character strings are included in each character string input as each search query included in the acquired inflow search query group. Is determined. Subsequently, when it is determined that the unnecessary character string is included, the processing unit 132 acquires the first character string obtained by removing the unnecessary character string from the character string input as the search query.

For example, the processing unit 132 refers to the list of unnecessary character strings, and determines that the character string “ladies unknown term L1” input as the inflow search query Q100 includes the unnecessary character string “ladies”. Subsequently, when the processing unit 132 determines that the unnecessary character string is included, the first character string obtained by removing the unnecessary character string "ladies" from the character string "ladies unknown term L1" input as the inflow search query Q100. Acquire the "unknown term L1" (first character string L1).

Further, the processing unit 132 refers to the list of unnecessary character strings, and the character string "unknown term L1 M size store name T1" input as the inflow search query Q200 includes the unnecessary character string "M size". Is determined. Subsequently, when the processing unit 132 determines that the unnecessary character string is included, the processing unit 132 removes the unnecessary character string "M size" from the character string "unknown term L1 M size store name T1" input as the inflow search query Q200. The first character string "unknown term L1 store name T1" (first character string L2) is acquired.

Further, the processing unit 132 refers to the list of unnecessary character strings, and determines that the character string "Y-shirt 20s" input as the inflow search query Q300 includes the unnecessary character string "20s". Subsequently, when the processing unit 132 determines that the unnecessary character string is included, the first character obtained by removing the unnecessary character string "20's" from the character string "Y-shirt 20's" input as the inflow search query Q300. The column "Y shirt" (first character string L3) is acquired.

Further, the processing unit 132 refers to the list of unnecessary character strings, and determines that the character string "unknown term L2 coordination" input as the inflow search query Q500 includes the unnecessary character string "coordination". Subsequently, when the processing unit 132 determines that the unnecessary character string is included, the first character string obtained by removing the unnecessary character string "corde" from the character string "unknown term L2 coordination" input as the inflow search query Q500. Acquire the "unknown term L2" (first character string L5).

Further, the processing unit 132 refers to the list of unnecessary character strings, and determines that the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600 includes "popularity" which is an unnecessary character string. do. Subsequently, when the processing unit 132 determines that the unnecessary character string is included, the processing unit 132 removes the unnecessary character string "popularity" from the character string "unknown term L2 brand name B1 popularity" input as the inflow search query Q600. Acquire one character string "unknown term L2 brand name B1" (first character string L6).

(Estimation unit 133)
The estimation unit 133 uses the second learning model acquired by the acquisition unit 131 to estimate the category to which the target indicated by the character string input as the search query belongs. The estimation unit 133 estimates the category to which the object indicated by the first character string acquired by the processing unit 132 belongs. Specifically, when the first character string is acquired by the processing unit 132, the estimation unit 133 refers to the model information storage unit 124 and acquires the second learning model (second learning model M2). Subsequently, when the estimation unit 133 acquires the second learning model, the estimation unit 133 estimates the category to which the object indicated by the first character string belongs by using the second learning model. The estimation unit 133 estimates a plurality of categories to which the object indicated by the character string belongs. For example, the estimation unit 133 estimates a plurality of categories to which the object indicated by the first character string belongs.

More specifically, the estimation unit 133 outputs the probability that the object indicated by the character string belongs to each category for each category. For example, the estimation unit 133 outputs the probability that the object indicated by the first character string belongs to each category for each category. For example, by inputting the first character string as the input information of the second learning model M2, the estimation unit 133 determines the probability that the target indicated by the first character string belongs to each category as the output information of the second learning model M2. Output for each category.

In the example shown in the middle of FIG. 1, when the estimation unit 133 acquires the first character string L1, the estimation unit 133 inputs the first character string L1 as the input information of the second learning model M2, thereby outputting the second learning model M2. As information, the probability that the target indicated by the first character string L1 belongs to each category is output for each category. For example, the estimation unit 133 outputs the probability that the object represented by the first character string L1 belongs to the fashion category C1 (hereinafter, also referred to as fashion category C1) as 100%, and the probability that the object belongs to another category is 0%.

Further, when the estimation unit 133 acquires the first character string L2, the estimation unit 133 inputs the first character string L2 as the input information of the second learning model M2, so that the first character string L2 is used as the output information of the second learning model M2. The probability that the target indicated by is in each category is output for each category. For example, the estimation unit 133 outputs the probability that the object indicated by the first character string L2 belongs to the fashion category C1 is 70%, the probability that the object belongs to the category C4 regarding the store name is 30%, and the probability that the object belongs to another category is 0%. do.

Further, when the estimation unit 133 acquires the first character string L3, the estimation unit 133 inputs the first character string L3 as the input information of the second learning model M2, thereby inputting the first character string L3 as the output information of the second learning model M2. The probability that the target indicated by is in each category is output for each category. For example, the estimation unit 133 outputs the probability that the object represented by the first character string L3 belongs to the fashion category C1 is 100%, and the probability that the object belongs to the other category is 0%.

Further, when the estimation unit 133 acquires the first character string L4, the estimation unit 133 inputs the first character string L4 as the input information of the second learning model M2, so that the first character string L4 is used as the output information of the second learning model M2. The probability that the target indicated by is in each category is output for each category. For example, the estimation unit 133 outputs the probability that the object indicated by the first character string L4 belongs to the fashion category C1 as 50%, the probability that the object belongs to the category C2 related to the person name as 50%, and the probability that the object belongs to the other category as 0%. ..

Further, when the estimation unit 133 acquires the first character string L5, the estimation unit 133 inputs the first character string L5 as the input information of the second learning model M2, so that the first character string L5 is used as the output information of the second learning model M2. The probability that the target indicated by is in each category is output for each category. For example, the estimation unit 133 outputs the probability that the object indicated by the first character string L5 belongs to the fashion category C1 as 100%, and the probability that the object belongs to another category is 0%.

Further, when the estimation unit 133 acquires the first character string L6, the estimation unit 133 inputs the first character string L6 as the input information of the second learning model M2, thereby inputting the first character string L6 as the output information of the second learning model M2. The probability that the target indicated by is in each category is output for each category. For example, the estimation unit 133 outputs the probability that the object indicated by the first character string L6 belongs to the fashion category C1 is 60%, the probability that the object belongs to the category C3 related to the brand name is 40%, and the probability that the object belongs to the other category is 0%. do.

(Extraction unit 134)
The extraction unit 134 extracts a target character string indicating an extraction target belonging to the target field from the character string based on the category estimated by the estimation unit 133. Specifically, the extraction unit 134 extracts a character string including the category indicating the target field from the categories estimated by the estimation unit 133 as the target character string. For example, in the extraction unit 134, when the category to which the target indicated by the first character string belongs is estimated by the estimation unit 133, whether or not the category indicating the target field is included in the categories estimated by the estimation unit 133. Is determined for each first character string. Subsequently, when the extraction unit 134 determines that the category indicating the target field is included in the category estimated by the estimation unit 133, the extraction unit 134 extracts the first character string. That is, the extraction unit 134 extracts the first character string indicating the target belonging to the category indicating the target field.

Further, the extraction unit 134 extracts a character string that does not include the unnecessary category registered as an unnecessary category from the categories estimated by the estimation unit 133 as the target character string. For example, the extraction unit 134 refers to the unnecessary category storage unit 123 to acquire a list of unnecessary categories registered as unnecessary categories. Subsequently, when the extraction unit 134 extracts the first character string including the category indicating the target field, the extraction unit 134 refers to the acquired list of unnecessary categories, and the unnecessary categories are included in the categories estimated by the estimation unit 133. It is determined for each extracted first character string whether or not the unnecessary category registered as is included. Subsequently, when the extraction unit 134 determines that the unnecessary category is not included in the categories estimated by the estimation unit 133, the extraction unit 134 extracts the first character string as the target character string. That is, the extraction unit 134 extracts the first character string other than the first character string indicating the target belonging to the unnecessary category as the target character string.

In the example shown on the right side of FIG. 1, the information processing apparatus 100 extracts the first character string "unknown term L1" (first character string L1) from the first character string as the target character string W1. Further, the information processing apparatus 100 extracts the first character string "Y-shirt" (first character string L3) from the first character string as the target character string W2. Further, the information processing apparatus 100 extracts the first character string "unknown term L2" (first character string L5) from the first character string as the target character string W3.

Specifically, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L1 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L1. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L1. Subsequently, the extraction unit 134 determines that the category estimated for the first character string L1 does not include an unnecessary category. Subsequently, since the extraction unit 134 determines that the unnecessary category is not included, the first character string L1 is extracted as the target character string W1.

Further, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L2 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L2. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L2. Subsequently, the extraction unit 134 determines that the category C3 relating to the store name, which is an unnecessary category, is included in the category estimated for the first character string L2. Subsequently, the extraction unit 134 determines that the first character string L2 is not extracted as the target character string because it is determined that the unnecessary category is included.

Further, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L3 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L3. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L3. Subsequently, the extraction unit 134 determines that the category estimated for the first character string L3 does not include an unnecessary category. Subsequently, since the extraction unit 134 determines that the unnecessary category is not included, the first character string L3 is extracted as the target character string W2.

Further, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L4 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L4. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L4. Subsequently, the extraction unit 134 determines that the category C2 relating to the personal name, which is an unnecessary category, is included in the category estimated for the first character string L4. Subsequently, the extraction unit 134 determines that the first character string L4 is not extracted as the target character string because it is determined that the unnecessary category is included.

Further, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L5 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L5. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L5. Subsequently, the extraction unit 134 determines that the category estimated for the first character string L5 does not include an unnecessary category. Subsequently, since the extraction unit 134 determines that the unnecessary category is not included, the first character string L5 is extracted as the target character string W3.

Further, the extraction unit 134 determines whether or not the fashion category C1 which is the target field is included in the category estimated for the first character string L6 by the estimation unit 133. The extraction unit 134 determines that the fashion category C1 which is the target field is included in the category estimated for the first character string L6. Subsequently, since the extraction unit 134 has determined that the fashion category C1 is included, it is determined whether or not the unnecessary category is included in the category estimated for the first character string L6. Subsequently, the extraction unit 134 determines that the category C3 relating to the brand name, which is an unnecessary category, is included in the category estimated for the first character string L6. Subsequently, the extraction unit 134 determines that the first character string L6 is not extracted as the target character string because it is determined that the unnecessary category is included.

Next, an example of information processing according to the embodiment will be described in more detail with reference to FIG. FIG. 8 is a diagram showing an example of information processing according to the embodiment. In the example shown in FIG. 8, the processing unit 132 refers to the query information storage unit 121 and acquires a character string input as a search query that has flowed into the fashion-related site (step S1').

For example, the processing unit 132 acquires the character string “site name N1 cosash” input as the search query Q1. Further, the processing unit 132 acquires the character string "site name N2 Kosash person name M1" input as the search query Q2. Further, the processing unit 132 acquires the character string "floral pattern thermal ladies" input as the search query Q3. Further, the processing unit 132 acquires the character string "floral pattern thermal popular brand name B1" input as the search query Q4. Further, the processing unit 132 acquires the character string "floral pattern thermal coordination store name T1" input as the search query Q5. Further, the processing unit 132 acquires the character string "mani-denim M size" input as the search query Q6. Further, the processing unit 132 acquires the character string "manidenim L size brand name B2" input as the search query Q7. In addition, the processing unit 132 acquires the character string "missing collar 20s" input as the search query Q8. Further, the processing unit 132 acquires the character string “missing collar 30s person name M2” input as the search query Q9.

Subsequently, when the processing unit 132 acquires the inflow search query, the processing unit 132 refers to the unnecessary character string storage unit 122, and an unnecessary character string from each character string input as each search query included in the acquired inflow search query group. The first character string from which the unnecessary character string registered as is removed is acquired. Specifically, the processing unit 132 temporarily stores the table 121A shown in the lower part of step S1'in FIG. 8 in the storage unit 120.

For example, the processing unit 132 removes the unnecessary character string "site name N1" (unnecessary character string UL11) from the character string "site name N1 cosash" input as the search query Q1, and removes the first character string "cosash" (first). The character string L11) is acquired. Further, the processing unit 132 removes the unnecessary character string "site name N2" (unnecessary character string UL12) from the character string "site name N2 Kosash person name M1" input as the search query Q2, and removes the first character string "Kosash person name M1". "(First character string L12) is acquired. Further, the processing unit 132 removes the unnecessary character string "ladies" (unnecessary character string UL21) from the character string "floral pattern thermal ladies" input as the search query Q3, and removes the first character string "floral pattern thermal" (first). The character string L13) is acquired. In addition, the processing unit 142 removes the unnecessary character string "popular" (unnecessary character string UL22) from the character string "flower pattern thermal popular brand name B1" input as the search query Q4, and removes the first character string "flower pattern thermal brand". First name B1 ”(first character string L14) is acquired. In addition, the processing unit 152 removes the unnecessary character string "coordination" (unnecessary character string UL23) from the character string "flower pattern thermal coordination store name T1" input as the search query Q5, and removes the first character string "flower pattern thermal coordination store". First name T1 ”(first character string L15) is acquired. Further, the processing unit 162 removes the unnecessary character string "M size" (unnecessary character string UL31) from the character string "mani denim M size" input as the search query Q6, and removes the first character string "mani denim" (first character string). L16) is acquired. Further, the processing unit 172 removes the unnecessary character string "L size" (unnecessary character string UL32) from the character string "mani denim L size brand name B2" input as the search query Q7, and removes the first character string "mani denim brand name B2". "(First character string L17) is acquired. Further, the processing unit 182 removes the unnecessary character string "20s" (unnecessary character string UL41) from the character string "missing collar 20s" input as the search query Q8, and removes the first character string "missing collar" (first). The character string L18) is acquired. In addition, the processing unit 192 removes the unnecessary character string "30s" (unnecessary character string UL42) from the character string "missing collar 30s person name M2" input as the search query Q9, and removes the first character string "missing collar person name M2". "(First character string L19) is acquired.

Subsequently, when the first character string is acquired by the processing unit 132, the estimation unit 133 refers to the model information storage unit 124 and acquires the second learning model (second learning model M2). Subsequently, when the estimation unit 133 acquires the second learning model, the estimation unit 133 outputs the probability that the object indicated by the first character string belongs to each category by using the second learning model. In the example shown in FIG. 8, the estimation unit 133 inputs the first character string as the input information of the second learning model M2, and the target indicated by the first character string as the output information of the second learning model M2 is each. The probability of belonging to a category is output for each category (step S2'). Specifically, the estimation unit 133 temporarily stores the information in the table 121B shown in step S2'in FIG. 8 in the storage unit 120.

For example, the estimation unit 133 inputs the first character string "Cosash" (first character string L11) as the input information of the second learning model M2, so that the first character string L11 is output information of the second learning model M2. The probability that the object indicated by the item belongs to the fashion category C1 is 100%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "Kosash person name M1" (first character string L12) as the input information of the second learning model M2, so that the first character is output information of the second learning model M2. The probability that the object represented by the column L12 belongs to the fashion category C1 is output as 50%, the probability that the object belongs to the category C2 related to the person name is 50%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "flower pattern thermal" (first character string L13) as the input information of the second learning model M2, so that the first character is output information of the second learning model M2. The probability that the object represented by the column L13 belongs to the fashion category C1 is 100%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "flower pattern thermal brand name B1" (first character string L14) as the input information of the second learning model M2, and thus as the output information of the second learning model M2. The probability that the object indicated by the first character string L14 belongs to the fashion category C1 is output as 60%, the probability that the object belongs to the category C3 related to the brand name is 40%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "flower pattern thermal store name T1" (first character string L15) as the input information of the second learning model M2, and thus as the output information of the second learning model M2. The probability that the object indicated by the first character string L15 belongs to the fashion category C1 is 70%, the probability that the object belongs to the category C4 related to the store name is 30%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "mani denim" (first character string L16) as the input information of the second learning model M2, so that the first character string L16 is output information of the second learning model M2. The probability that the object indicated by the item belongs to the fashion category C1 is 100%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "manidenim brand name B2" (first character string L17) as the input information of the second learning model M2, so that the first is the output information of the second learning model M2. The probability that the object indicated by the character string L17 belongs to the fashion category C1 is 50%, the probability that the object belongs to the category C3 related to the brand name is 50%, and the probability that the object belongs to the other category is 0%.

Further, the estimation unit 133 inputs the first character string "missing collar" (first character string L18) as the input information of the second learning model M2, so that the first character string is output information of the second learning model M2. The probability that the object indicated by L18 belongs to the fashion category C1 is output as 100%, and the probability that the object belongs to the other category is output as 0%.

Further, the estimation unit 133 inputs the first character string "missing collar person name M2" (first character string L19) as the input information of the second learning model M2, so that the first is the output information of the second learning model M2. The probability that the object indicated by the character string L19 belongs to the fashion category C1 is 80%, the probability that the object belongs to the category C2 related to the person name is 20%, and the probability that the object belongs to the other category is 0%.

Subsequently, when the extraction unit 134 outputs the probability that the object indicated by the first character string belongs to each category by the estimation unit 133 for each category, the category indicating the target field is included in the categories estimated by the estimation unit 133. The first character string containing is extracted. That is, the extraction unit 134 extracts the first character string indicating the target belonging to the category indicating the target field. Subsequently, when the extraction unit 134 extracts the first character string including the category indicating the target field, the unnecessary category is included in the categories estimated by the estimation unit 133 with reference to the unnecessary category storage unit 123. The first character string that cannot be extracted is extracted as the target character string. That is, the extraction unit 134 extracts the first character string other than the first character string indicating the target belonging to the unnecessary category as the target character string (also referred to as the second character string) (step S3'). Specifically, the estimation unit 133 stores the information of the table 121C shown in step S3'of FIG. 8 in the storage unit 120.

For example, the extraction unit 134 includes the fashion category C1 indicating the target field in the category estimated for the first character string “Cosash” (first character string L11), and also includes an unnecessary category. Since there is no such thing, the first character string "Cosash" (first character string L11) is extracted as the second character string W21.

Further, the extraction unit 134 includes the fashion category C1 in the category estimated for the first character string "Kosash person name M1" (first character string L12), but includes the category C2 for the person name which is an unnecessary category. Therefore, it is determined not to extract the first character string "Kosash person name M1" (first character string L12) as the second character string.

Further, the extraction unit 134 includes the fashion category C1 indicating the target field in the category estimated for the first character string "flower pattern thermal" (first character string L13), and the unnecessary category is included. Since it is not included, the first character string "flower pattern thermal" (first character string L13) is extracted as the second character string W22.

Further, the extraction unit 134 includes the fashion category C1 in the category estimated for the first character string "flower pattern thermal brand name B1" (first character string L14), but relates to a brand name which is an unnecessary category. Since the category C3 is included, it is determined not to extract the first character string "flower pattern thermal brand name B1" (first character string L14) as the second character string.

Further, the extraction unit 134 includes the fashion category C1 in the category estimated for the first character string "flower pattern thermal store name T1" (first character string L15), but relates to the store name which is an unnecessary category. Since the category C4 is included, it is determined not to extract the first character string "flower pattern thermal store name T1" (first character string L15) as the second character string.

Further, the extraction unit 134 includes the fashion category C1 indicating the target field in the category estimated for the first character string "mani denim" (first character string L16), and also includes an unnecessary category. Since there is no such thing, the first character string "mani denim" (first character string L16) is extracted as the second character string W23.

Further, the extraction unit 134 includes the fashion category C1 in the category estimated for the first character string "mani denim brand name B2" (first character string L17), but the category C3 regarding the brand name which is an unnecessary category. Is included, so it is determined not to extract the first character string "manidenim brand name B2" (first character string L17) as the second character string.

Further, the extraction unit 134 includes the fashion category C1 indicating the target field in the category estimated for the first character string "missing collar" (first character string L18), and also includes an unnecessary category. Therefore, the first character string "missing collar" (first character string L18) is extracted as the second character string W24.

Further, the extraction unit 134 includes the fashion category C1 in the category estimated for the first character string "missing collar person name M2" (first character string L19), but the category C2 relating to the person name which is an unnecessary category is included. Since it is included, it is determined not to extract the first character string "missing collar person name M2" (first character string L19) as the second character string.

[1-3. Information processing flow]
Next, the procedure of information processing according to the embodiment will be described with reference to FIG. FIG. 9 is a flowchart showing an information processing procedure according to the embodiment. In the example shown in FIG. 9, the information processing apparatus 100 acquires the search query that has flowed into the site related to the target field (step S101). Subsequently, when the information processing apparatus 100 acquires the search query that has flowed into the site related to the target field, it acquires the first character string obtained by removing the unnecessary character string from the character string input as the search query (step S102). Subsequently, when the information processing apparatus 100 acquires the first character string, the information processing apparatus 100 estimates the category to which the object indicated by the first character string belongs by using the second learning model (step S103). Subsequently, when the category is estimated, the information processing apparatus 100 extracts a second character string indicating an extraction target from the first character string based on the estimated category (step S104).

[2. Learning model generation process]
[2-1. First learning model generation process]
Next, the flow of the generation process of the first learning model will be described with reference to FIG. FIG. 10 is a diagram showing an example of a generation process of the first learning model according to the embodiment. In the example shown in FIG. 10, the generator 50 is a pair consisting of a search query Q11 "Roppongi pasta" and a search query Q12 "Roppongi Italian" continuously input by the same user U1 within a predetermined time. Extract the search query (step S11).

Subsequently, the generation device 50 inputs the extracted search query Q11 into the first model M1 and outputs the vector BQV11 which is a distributed representation of the search query Q11. Here, the vector BQV11 is a distributed expression of the search query Q11 just output from the output layer of the first model M1, and shows a distributed expression before giving feedback to the first model M1 (before learning). Further, the generation device 50 inputs the extracted search query Q12 into the first model M1 and outputs a vector BQV12 which is a distributed representation of the search query Q12. Here, the vector BQV12 is a distributed expression of the search query Q12 just output from the output layer of the first model M1, and shows a distributed expression before giving feedback to the first model M1 (before learning). In this way, the generation device 50 outputs the vector BQV11 which is the distributed representation of the search query Q11 and the vector BQV12 which is the distributed representation of the search query Q12 (step S12).

Subsequently, the generation device 50 is a pair of search queries composed of a search query Q11 (“Roppongi pasta”) and a search query Q12 (“Roppongi Italian”) continuously input by the same user U1 within a predetermined time. Is a search query entered with a predetermined search intention (for example, a search intention of "searching for a restaurant in a certain place"). The first model M1 is trained so that the distributed expression (vector QV11) and the distributed expression (vector QV12) of the search query Q12 paired with the search query Q11 are similar on the distributed expression space. For example, let Θ be the size of the angle formed by the vector BQV11 which is the distributed expression of the search query Q11 before giving feedback to the first model M1 (before learning) and the vector BQV12 which is the distributed expression of the search query Q12. Further, the size of the angle formed by the vector QV11 which is the distributed expression of the search query Q11 after giving feedback to the first model M1 (after learning) and the vector QV12 which is the distributed expression of the search query Q12 is Φ. At this time, the generator 50 trains the first model M1 so that Φ is smaller than Θ. For example, the generator 50 calculates the value of the cosine similarity between the vector BQV11 and the vector BQV12. Further, the generation device 50 calculates the value of the cosine similarity between the vector QV11 and the vector QV12. Subsequently, the generator 50 uses the first model so that the value of the cosine similarity between the vector QV11 and the vector QV12 is larger than the value of the cosine similarity between the vector BQV11 and the vector BQV12 (so that the value approaches 1). Learn M1. In this way, the generation device 50 trains the first model M1 so that the two vectors, which are a pair of distributed representations corresponding to the pair of search queries, are similar in the distributed representation space, so that the distributed representation is expressed from the search query. Is generated (step S13). The generator 50 is not limited to the cosine similarity, and may calculate the similarity between the distributed representations based on any index as long as it is an index applicable as a distance scale between vectors. Further, the generation device 50 may train the first model M1 based on any index as long as it is an index applicable as a distance scale between vectors. For example, the generation device 50 calculates the value of a predetermined distance function such as the Euclidean distance between distributed expressions, the distance in a non-Euclidean space such as a twin-curved space, the Manhattan distance, the Mahalanobis distance, and the like. Subsequently, the generation device 50 may train the first model M1 so that the value of a predetermined distance function between the distributed representations (that is, the distance in the distributed representation space) becomes small.

Next, the flow of the generation process of the first learning model will be described in more detail with reference to FIG. In the description of FIG. 11, a part that overlaps with the description of FIG. 10 will be omitted as appropriate. FIG. 11 is a diagram showing a generation process of the first learning model according to the embodiment. In the example shown in FIG. 11, the distributed representation output by the first model M1 generated by the generation device 50 is mapped to the distributed representation space. The generation device 50 trains the first model M1 so that the distributed representation of a predetermined search query and the distributed representation of another search query paired with the predetermined search query are closely mapped on the distributed representation space. ..

In the example shown in the upper part of FIG. 11, the generation device 50 is a search query Q11 (“Roppongi pasta”) and a search query Q12, which are four search queries continuously input by the same user U1 within a predetermined time. ("Roppongi Italian"), search query Q13 ("Akasaka pasta"), search query Q14 ("Azabu pasta") are extracted. The generation device 50 extracts four search queries in which the time interval in which each search query is input by the same user U1 is within a predetermined time. The generation device 50 extracts a plurality of search queries in which the time interval in which each search query pair described later is input by the same user U1 is within a predetermined time. When the generation device 50 arranges the search queries in the order in which they are input, the generation device 50 extracts four search queries input in the order of search query Q11, search query Q12, search query Q13, and search query Q14. When the generation device 50 extracts four search queries, it is a pair of three pairs of search queries (search query Q11, search query Q12), with two search queries adjacent in chronological order as a pair of search queries. (Search query Q12, search query Q13) and (search query Q13, search query Q14) are extracted (step S21-1). The generation device 50 may extract a plurality of search queries in which all the search queries are input within a predetermined time by the same user U1. Then, the generation device 50 selects two search queries from the extracted plurality of search queries regardless of whether they are adjacent to each other in chronological order, and the two selected search queries are used as a pair of search queries. It may be extracted.

Subsequently, the generation device 50 inputs the extracted search query Q1k (k = 1, 2, 3, 4) into the first model M1, and distributes the search query Q1k (k = 1, 2, 3, 4). The representation vector BQV1k (k = 1, 2, 3, 4) is output. Here, the vector BQV1k (k = 1, 2, 3, 4) is a distributed representation of the search query Q1k (k = 1, 2, 3, 4) just output from the output layer of the first model M1. , The distributed representation before giving feedback (before learning) to the first model M1 is shown (step S22-1).

Subsequently, in the generation device 50, a pair of search queries continuously input by the same user U1 within a predetermined time is a predetermined search intention (for example, "a restaurant at a certain place (near Minato-ku, Tokyo)). Since it is presumed that the search query was entered with the search intent of "searching for"), the distributed representation of the search query Q11 (vector QV11) and the search query Q11 are paired as having similar characteristics to each other. The first model M1 is trained so that the distributed expression (vector QV12) of the search query Q12 is similar on the distributed expression space. Further, in the generation device 50, the distributed representation of the search query Q12 (vector QV12) and the distributed representation of the search query Q13 paired with the search query Q12 (vector QV13) are similar to each other in the distributed representation space. Train model M1. Further, in the generation device 50, the distributed expression of the search query Q13 (vector QV13) and the distributed expression of the search query Q14 paired with the search query Q13 (vector QV14) are similar to each other in the distributed expression space. Train model M1. In this way, the generation device 50 trains the first model M1 so that the two vectors, which are a pair of distributed representations corresponding to the pair of search queries, are similar in the distributed representation space, so that the distributed representation is expressed from the search query. Is generated (step S23-1).

As a result of the information processing shown in the upper part of FIG. 11, the vector QV1k (k = 1, 2, 3, 4), which is a distributed representation of the search query Q1k (k = 1, 2, 3, 4), is close to the distributed representation space. It is shown that the position is mapped as the cluster CL11. For example, the search query Q1k (k = 1, 2, 3, 4) is a set of search queries searched by the user U1 with the search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)". Is presumed to be. That is, the search query Q1k (k = 1, 2, 3, 4) is a search query searched with the search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)". , Is presumed to be a search query with similar characteristics. Here, when the predetermined search query input with the search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)" is input to the first model, the generator 50 is located at the position of the cluster CL11. It is possible to output a distributed representation that is mapped. As a result, for example, the generator 50 has a search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)" by extracting a search query corresponding to the distributed expression mapped to the position of the cluster CL11. Search queries can be extracted according to. Therefore, the generation device 50 can appropriately interpret the meaning of the search query.

In the example shown in the lower part of FIG. 11, the generation device 50 is a search query Q21 (“refrigerator 400L”) and a search query Q22, which are three search queries continuously input by the same user U2 within a predetermined time. ("Refrigerator medium size"), search query Q23 ("Refrigerator medium size recommended") is extracted. When the generation device 50 arranges the search queries in the order in which they are input, the generation device 50 extracts three search queries input in the order of the search query Q21, the search query Q22, and the search query Q23. When the generation device 50 extracts three search queries, it is a pair of two search queries (search query Q21, search query Q22), with two search queries adjacent in chronological order as a pair of search queries. (Search query Q22, search query Q23) is extracted (step S21-2).

Subsequently, the generation device 50 inputs the extracted search query Q2m (m = 1, 2, 3) into the first model M1, and a vector that is a distributed representation of the search query Q2m (m = 1, 2, 3). BQV2m (m = 1, 2, 3) is output. Here, the vector BQV2m (m = 1, 2, 3) is a distributed representation of the search query Q2m (m = 1, 2, 3) just output from the output layer of the first model M1, and is the first model. A distributed expression before giving feedback to M1 (before learning) is shown (step S22-2).

Subsequently, the generation device 50 inputs a pair of search queries continuously input by the same user U2 within a predetermined time with a predetermined search intention (for example, a search intention of "examining a medium-sized refrigerator"). Since it is presumed that the search query was performed, the distributed representation of the search query Q21 (vector QV21) and the distributed representation of the search query Q22 paired with the search query Q21 (vector QV22) are assumed to have similar characteristics to each other. ) And the first model M1 are trained so as to be similar in the distributed expression space. Further, in the generation device 50, the first distributed representation of the search query Q22 (vector QV22) and the distributed representation of the search query Q23 paired with the search query Q22 (vector QV23) are similar in the distributed representation space. Train model M1. In this way, the generation device 50 trains the first model M1 so that the two vectors, which are a pair of distributed representations corresponding to the pair of search queries, are similar on the distributed representation space, so that the distributed representation is expressed from the search query. Is generated (step S23-2).

As a result of the information processing shown in the lower part of FIG. 11, the vector QV2m (m = 1, 2, 3), which is a distributed representation of the search query Q2m (m = 1, 2, 3), is located close to the distributed representation space in the cluster CL21. It is shown how it is mapped as. For example, the search query Q2m (m = 1, 2, 3) is presumed to be a set of search queries searched by the user U2 with the search intention of "checking a medium-sized refrigerator". That is, Q2m (m = 1, 2, 3) is a search query having similar characteristics in that it is a search query searched with the search intention of "searching for a medium-sized refrigerator". Presumed. Here, the generation device 50 outputs a distributed expression that is mapped to the position of the cluster CL21 when a predetermined search query input with the search intention of "checking a medium-sized refrigerator" is input to the first model. can do. Thereby, for example, the generation device 50 extracts the search query corresponding to the distributed expression mapped to the position of the cluster CL21, thereby extracting the search query according to the search intention of "examining the medium-sized refrigerator". Can be done. Therefore, the generation device 50 can appropriately interpret the meaning of the search query.

Further, the generation device 50 according to the present invention is a search query having different characteristics from each other in that a plurality of randomly extracted search queries are search queries searched under different search intentions. Assuming that, the first model M1 is trained. Specifically, the generation device 50 maps the distributed representation of a predetermined search query and the distributed representation of a search query randomly extracted independently of the predetermined search query far away in the distributed representation space. The first model M1 is trained. In the example shown in FIG. 11, it is assumed that the generation device 50 randomly extracts the search query regardless of the search query Q11, and the search query Q21 is extracted. In this case, in the generation device 50, the distributed representation of the search query Q11 (vector QV11) and the distributed representation of the search query Q21 randomly extracted regardless of the search query Q11 (vector QV21) are far apart on the distributed representation space. The first model M1 is trained so as to be mapped to. As a result, the vector QV1k, which is a distributed expression of the search query Q1k (k = 1, 2, 3, 4) searched under the search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)". It is a distributed representation of the cluster CL11 including (k = 1, 2, 3, 4) and the search query Q2m (m = 1, 2, 3) searched under the search intention of "searching for a medium-sized refrigerator". The cluster CL21 containing the vector QV2m (m = 1, 2, 3) is mapped far away on the distributed representation space. That is, the generation device 50 according to the present invention trains the first model M1 so that the distributed expressions of a plurality of randomly extracted search queries are different, thereby expressing the distributed expressions of the search queries having different search intentions. It is possible to output to a distant position in space.

As a result of mapping the distributed representation output by the first model M1 generated by the generation device 50 to the distributed representation space, in addition to the above-mentioned cluster CL11 and cluster CL21, the same user within a predetermined time. Cluster CL12 and cluster CL22, which are a set of distributed representations of a plurality of input search queries, are generated.

As described above, the generator 50 acquires the search query entered by the user. Further, the generation device 50 learns from the acquired search queries that a plurality of search queries input by the same user within a predetermined time have similar characteristics, so that the search queries can be determined from the predetermined search queries. Generate a first model that predicts the feature information of the search query. That is, the generation device 50 according to the present invention is similar to each other in that a plurality of search queries continuously input within a predetermined time are search queries searched under a predetermined search intention. The first model is trained by regarding it as a search query having characteristics. Specifically, the generation device 50 trains the first model so that the distributed representations of a plurality of search queries input by the same user within a predetermined time are similar, so that a predetermined search query can be used as a predetermined value. Generate a first model that outputs a distributed representation that includes the feature information of the search query. That is, the generation device 50 according to the present invention trains the first model M1 so that the distributed expressions of a plurality of search queries continuously input within a predetermined time are similar to each other, so that the first model M1 is trained under a predetermined search intention. The distributed representation of the search query searched in is able to be output to a close position on the distributed representation space. This makes it possible for the generation device 50 to output (interpret) the meaning (search intention) of the search query according to the context of the user who input the search query. Therefore, the generation device 50 can appropriately interpret the meaning of the search query. Further, the generation device 50 extracts a search query corresponding to the distributed expression mapped in the vicinity of the distributed expression including the feature information of the predetermined search query, so that the predetermined search query can be searched according to the search intention. You can extract search queries. That is, the generation device 50 makes it possible to analyze the search trend of the user in consideration of the search intention and the context of the user who input the search query. Therefore, the generation device 50 can improve the analysis accuracy of the user's search trend. Further, the first model M1 generated by the generation device 50 can be made to function as a part of the search system. Alternatively, the generation device 50 is a distributed representation of the search query output by the first model M1 as input information to another system (for example, a search engine) that uses the feature information of the search query predicted by the first model M1. Can also be provided. As a result, the search system can select the content output as the search result based on the feature information of the search query predicted by the first model M1. That is, the search system can select the content output as the search result in consideration of the search intention and context of the user who entered the search query. Further, the search system calculates the similarity between the distributed expression of the character string included in the content output as the search result and the distributed expression of the search query based on the characteristic information of the search query predicted by the first model M1. It will be possible. Then, the search system can determine the display order of the contents output as the search result based on the calculated similarity. That is, the search system can determine the display order of the contents output as the search result in consideration of the search intention and the context of the user who input the search query. Therefore, the generation device 50 can improve the usability in the search service.

[2-2. Second learning model generation process]
Next, the flow of the generation process of the second learning model will be described with reference to FIG. FIG. 12 is a diagram showing an example of a generation process of the second learning model according to the embodiment. In the following, the second learning model will be referred to as a second model (or a second model M2) as appropriate. In the example shown in the upper part of FIG. 12, the generation device 50 is a search query Q11 (“Roppongi pasta”) and a search query Q12, which are four search queries continuously input by the same user U1 within a predetermined time. ("Roppongi Italian"), search query Q13 ("Akasaka pasta"), search query Q14 ("Azabu pasta") are extracted. The generation device 50 extracts a plurality of search queries in which the time interval in which each search query is input by the same user U1 is within a predetermined time. Further, the generation device 50 extracts a plurality of search queries in which the time interval in which each search query pair is input by the same user U1 is within a predetermined time. Here, it is assumed that the four search queries are search queries in which each search query is input by the user U1 in the order of search query Q11, search query Q12, search query Q13, and search query Q14 within a predetermined time. When the generation device 50 extracts four search queries, it is a pair of three pairs of search queries (search query Q11, search query Q12), with two search queries adjacent in chronological order as a pair of search queries. (Search query Q12, search query Q13) and (search query Q13, search query Q14) are extracted. When the generation device 50 extracts three pairs of search queries, the extracted search queries Q1k (k = 1, 2, 3, 4) are input to the first model M1 (step S31). The generation device 50 may extract a plurality of search queries in which all the search queries are input within a predetermined time by the same user U1. Then, the generation device 50 selects two search queries from the plurality of extracted search queries regardless of whether they are adjacent to each other in chronological order, and the two selected search queries are used as a pair of search queries. It may be extracted.

Subsequently, the generation device 50 outputs the vector BQV1k (k = 1, 2, 3, 4), which is a distributed representation of the search query Q1k (k = 1, 2, 3, 4), as output data of the first model M1. (Step S32). Here, the vector BQV1k (k = 1, 2, 3, 4) is a distributed representation of the search query Q1k (k = 1, 2, 3, 4) just output from the output layer of the first model M1. , The distributed representation before giving feedback to the first model M1 (before learning) is shown.

Here, the search query Q1k (k = 1, 2, 3, 4) continuously input by the same user U1 within a predetermined time is, for example, "a certain place (near Minato-ku, Tokyo)" by the user U1. It is presumed to be a set of search queries searched under the search intention of "searching for restaurants in Tokyo". That is, the search query Q1k (k = 1, 2, 3, 4) is a search query searched with the search intention of "searching for a restaurant in a certain place (near Minato-ku, Tokyo)". , Is presumed to be a search query with similar characteristics. Therefore, the generation device 50 generates a first model that predicts the characteristic information of a predetermined search query from the predetermined search query by learning that the continuously input search queries have similar characteristics (the generation device 50). Step S33). Specifically, the generation device 50 predicts the distributed expression of a predetermined search query from a predetermined search query by learning that the distributed expressions of the continuously input search queries are similar to each other. To generate. For example, in the generation device 50, the first distributed representation of the search query Q11 (vector QV11) and the distributed representation of the search query Q12 paired with the search query Q11 (vector QV12) are similar in the distributed representation space. Train model M1. Further, in the generation device 50, the distributed representation of the search query Q12 (vector QV12) and the distributed representation of the search query Q13 paired with the search query Q12 (vector QV13) are similar to each other in the distributed representation space. Train model M1. Further, in the generation device 50, the distributed expression of the search query Q13 (vector QV13) and the distributed expression of the search query Q14 paired with the search query Q13 (vector QV14) are similar to each other in the distributed expression space. Train model M1.

On the right side of the upper part of FIG. 12, as the output result of the trained first model M1, the variance of the search query Q1k (k = 1, 2, 3, 4) input by the same user U1 within a predetermined time is distributed. It is shown that the representation vector QV1k (k = 1, 2, 3, 4) is mapped as the cluster CL11 in the distributed representation space. In this way, the generation device 50 generates the first learning model M1 that has learned the features of the plurality of search queries input by the same user within a predetermined time.

When the generation device 50 generates the first model M1, the generated first model M1 (model data MDT1 of the first model M1) is acquired. When the generation device 50 acquires the first model M1, the generation device 50 generates the second learning model M2 by using the acquired first model M1. Specifically, the generation device 50 relearns the first model M1 to generate the second model M2 having a connection coefficient different from that of the first model M1 which is the weight of the learning model. More specifically, the generation device 50 uses the first model M1 to generate a second learning model M2 that predicts a category to which a predetermined search query belongs from a predetermined search query (step S34).

In the example shown in the lower part of FIG. 12, when the search query is input to the second model M2, the generator 50 has CAT11 (“find a restaurant”), CAT12 (“find a product”), and CAT13 (“find a product”). A second model M2 that predicts which of the four categories of "reserving a restaurant") and CAT14 ("purchasing a product") belongs to is generated. Specifically, the generation device 50 generates the second model M2 that outputs the probability that the search query belongs to the category as the output information when the search query is input to the second model M2 as the input information. do. For example, the generation device 50 learns a set of a search query and a category (any of CAT11 to CAT14) to which the search query belongs as correct answer data of the second model M2.

The fact that the search query belongs to CAT11 (“search for a restaurant”) indicates that the search query is a search query entered with the intention of searching for a restaurant. Further, belonging to CAT12 (“searching for a product”) indicates that the search query is a search query input with the intention of searching for a product. Further, the fact that the search query belongs to CAT13 (“reserving a restaurant”) indicates that the search query is a search query input with the intention of reserving a restaurant. Further, the fact that the search query belongs to CAT14 (“purchase a product”) indicates that the search query is a search query input with the intention of purchasing a product.

Specifically, the generation device 50 determines that when the search query is input to the learning model, the classification result of the distributed expression output by the learning model is learned so as to correspond to the category to which the search query belongs. A second model M2 that predicts the category to which a predetermined search query belongs is generated from the search query of. Then, for example, when the search query is input to the second model M2 as input information, the generation device 50 outputs the probability that the search query belongs to the category as output information for each of the categories CAT11 to CAT14. To generate.

For example, the generation device 50 distributes the search query Q11 (“Roppongi pasta”) as output information when the search query Q11 (“Roppongi pasta”) is input to the second model M2 as input information (step S35). Is output as the vector BQV11. Here, the vector BQV11 is a distributed expression of the search query Q11 just output from the output layer of the second model M2, and shows a distributed expression before giving feedback to the second model M2 (before learning). Here, it is assumed that the correct answer category to which the search query Q11 (“Roppongi pasta”) belongs is CAT11 (“find a restaurant”). In this case, the generation device 50 sets the probability that the vector BQV11, which is a distributed representation of the output search query Q11 (“Roppongi pasta”), is classified into CAT11 (“searching for a restaurant”) exceeds a predetermined threshold value. The second model M2 is trained. The generation device 50 trains the second model using the correct answer data prepared in advance. The generation device 50 may generate correct answer data of the second model M2. Then, the generation device 50 may train the second model M2 using the generated correct answer data. Specifically, the generation device 50 determines the correct answer category to which the search query belongs based on the post-search behavior of the user who searched for the search query. More specifically, the generation device 50 corresponds to a predetermined action in which the ratio of users who have performed a predetermined action after the search exceeds a predetermined threshold with respect to the user who searched for the predetermined search query, according to the correct answer category. Determine as an action. For example, 90% of the users who searched for the search query Q11 (“Roppongi Pasta”) took a predetermined action after the search, and 90% of the users took the action of searching for a restaurant, and the action of searching for a product after the search. It is assumed that the percentage of users who have caused the problem is 0%, the percentage of users who have taken the action of reserving a restaurant after the search is 10%, and the percentage of users who have taken the action of purchasing the product after the search is 0%. In this case, since the percentage of users who have taken the action of searching for a restaurant exceeds a predetermined threshold value (for example, 90%), the generation device 50 searches for the action of searching for a restaurant with the correct answer of the search query Q11 (“Roppongi pasta”). Determined as an action corresponding to the category. Then, since the generation device 50 determines that the action corresponding to the correct answer category is the action of searching for a restaurant, the correct answer category to which the search query Q11 (“Roppongi pasta”) belongs is set to CAT11 (“find a restaurant”). decide.

For example, in the generation device 50, when the search query Q11 (“Roppongi pasta”) is input to the second model M2 before learning, the vector BQV11 which is a distributed expression is classified into CAT11 (“find a restaurant”). 80% probability of being classified as CAT12 ("find a product") 0%, 20% probability of being classified as CAT13 ("book a restaurant"), CAT14 ("buy a product") It is assumed that the probability of being classified as 0% is output. In this case, the generation device 50 trains the second model M2 so that the probability that the vector BQV11, which is a distributed representation, is classified into CAT11 (“find a restaurant”) exceeds a predetermined threshold value (for example, 90%). .. Further, the generator 50 is distributed so that the probability that the vector BQV11, which is a distributed expression, is classified into CAT11 (“find a restaurant”) exceeds a predetermined threshold value (for example, 90%). The second model M2 is trained so that the probability that the representation vector BQV11 is classified into another category CAT13 (“book a restaurant”) is reduced to 10%.

As described above, when the predetermined search query is input as the input information, the generation device 50 makes the second so that the probability that the distributed expression of the predetermined search query is classified into the correct answer category as the output information exceeds the predetermined threshold value. Train the model. Then, when a predetermined search query is input as input information, the generation device 50 sets a category in which the probability that the distributed expression of the predetermined search query belongs to the category exceeds a predetermined threshold value as a predetermined search query category. Output. For example, in the generation device 50, when the search query Q11 (“Roppongi pasta”) is input as input information to the trained second model M2, the vector BQV11 which is a distributed representation of the search query Q11 (“Roppongi pasta”) is generated. Since the probability of belonging to the category CAT11 ("find a restaurant") exceeds 90%, the category to which the search query belongs is output as CAT11 ("find a restaurant") as output information (step S36). In this way, the generation device 50 generates a second model that predicts a predetermined search query category from a predetermined search query by learning a set of a search query and a correct answer category of the search query (step S37). ..

In general, it is considered that the user performs a search a plurality of times with a certain intention, so that it is assumed that the search queries that are continuously input within a predetermined time have similar search intentions. Therefore, the generation device 50 according to the present invention is similar to each other in that a plurality of search queries continuously input within a predetermined time are search queries searched under a predetermined search intention. The first model M1 is trained by regarding it as a search query having characteristics. As a result, the generation device 50 can make the first model M1 learn the characteristics of the search query in consideration of the search intention. Then, the generation device 50 efficiently generates a second model that predicts a predetermined search query category from a predetermined search query by utilizing the first model M1 that has learned the characteristics of the search query in consideration of the search intention. can do. This allows the generator 50 to classify the search query into categories that take into account the search intent of the user who entered the search query. Further, conventionally, in order to classify search queries into categories and obtain high classification accuracy, it has been necessary to prepare a sufficient amount of correct answer data. However, because the search queries themselves are diverse and have long-tailed properties, it is very laborious and difficult to label the correct answer categories that correspond to a large number of search queries. Here, instead of labeling the correct answer category, the generation device 50 trains a second model that predicts the category of the search query by using the user's search intention (the context of the user who entered the search query) as a kind of correct answer. be able to. As a result, the generation device 50 can train the second model without manually labeling the correct answer category of the search query. That is, the generation device 50 can obtain sufficient classification accuracy even when the number of correct answer data is small. Further, the generation device 50 can obtain even higher classification accuracy when there are many correct answer data. Therefore, the generation device 50 can improve the classification accuracy of the search query.

[2-3. Information processing device configuration]
Next, the configuration of the generation device 50 according to the embodiment will be described with reference to FIG. FIG. 13 is a diagram showing a configuration example of the generation device 50 according to the embodiment. As shown in FIG. 13, the generation device 50 includes a communication unit 51, a storage unit 53, and a control unit 52. The generation device 50 has an input unit (for example, a keyboard, a mouse, etc.) that receives various operations from the administrator of the generation device 50, and a display unit (for example, a liquid crystal display, etc.) for displaying various information. You may.

(Communication unit 51)
The communication unit 51 is realized by, for example, a NIC (Network Interface Card) or the like. Then, the communication unit 51 is connected to the network by wire or wirelessly, and for example, information is transmitted / received between the user terminal 10 and the search server 20.

(Memory unit 53)
The storage unit 53 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. As shown in FIG. 13, the storage unit 53 includes a query information storage unit 531, a vector information storage unit 532, a classification definition storage unit 533, a category information storage unit 534, and a model information storage unit 535.

(Query information storage unit 531)
The query information storage unit 531 stores various information related to the search query input by the user. FIG. 14 shows an example of the query information storage unit according to the embodiment. In the example shown in FIG. 14, the query information storage unit 531 has items such as "user ID", "date and time", "search query", and "search query ID".

The "user ID" indicates identification information for identifying the user who entered the search query. "Date and time" indicates the date and time when the search server received the search query from the user. A "search query" indicates a search query entered by the user. The "search query ID" indicates identification information for identifying the search query entered by the user.

In the example shown in the first record of FIG. 14, the search query (search query Q11) identified by the search query ID "Q11" corresponds to the search query Q11 shown in FIG. Further, the user ID "U1" indicates that the user who has input the search query Q11 is the user (user U1) identified by the user ID "U1". Further, the date and time "2018/9/1 PM 17:00" indicates that the date and time when the search server receives the search query Q11 from the user U1 is 17:00 pm on September 1, 2018. Further, the search query "Roppongi pasta" indicates the search query Q11 input by the user U1. Specifically, the search query "Roppongi pasta" indicates that the characters "Roppongi" indicating the place name and the characters "pasta" indicating the type of food are separated by a space which is a delimiter.

(Vector information storage unit 532)
The vector information storage unit 532 stores various information related to the vector, which is a distributed representation of the search query. FIG. 15 shows an example of the vector information storage unit according to the embodiment. In the example shown in FIG. 15, the vector information storage unit 532 has items such as "vector ID", "search query ID", and "vector information".

The "vector ID" indicates identification information for identifying a vector which is a distributed representation of a search query. The "search query ID" indicates identification information for identifying the search query corresponding to the vector. "Vector information" indicates an N-dimensional vector which is a distributed representation of a search query. The vector that is the distributed representation of the search query is, for example, a 128-dimensional vector.

In the example shown in the first record of FIG. 15, the vector (vector QV11) identified by the vector ID “QV11” corresponds to the vector QV11 which is the distributed representation of the search query Q11 shown in FIG. Further, the search query (search query Q11) identified by the search query ID "Q11" indicates that the search query corresponding to the vector QV11 is the search query Q11. Further, the vector information "QVDT11" indicates an N-dimensional vector which is a distributed representation of the search query Q11.

(Classification definition storage unit 533)
The classification definition storage unit 533 stores various information regarding the definition of the category in which the search query is classified. FIG. 16 shows an example of the classification definition storage unit according to the embodiment. In the example shown in FIG. 16, the classification definition storage unit 533 has items such as "major classification ID", "major classification", "minor classification ID", and "minor classification".

"Major classification" indicates a major classification of the category in which the search query is classified. The "major classification ID" indicates identification information for identifying the major classification. In the example shown in FIG. 16, the major classification “purchasing behavior system” corresponds to the major classification described in the example shown in the lower part of FIG. Major classification "Purchasing behavior system" indicates a major classification of categories that classify search queries based on the user's purchasing behavior. In the example shown in FIG. 16, the major classification "purchasing behavior system" has four further minor classifications. The major classification ID "CAT1" indicates identification information for identifying the major classification "purchasing behavior system".

"Minor classification", indicates the subclassification of the category in which the search query is classified. The "minor classification ID" indicates identification information for identifying the minor classification. In the example shown in FIG. 16, the minor classification "find a restaurant" is a classification belonging to the major classification "purchasing behavior system", and the search query classified into the minor classification is input with the intention of searching for a restaurant by the user. Indicates that the search query was made. The sub-category ID "CAT11" indicates identification information for identifying the sub-category "find a restaurant".

The sub-category "search for products" is a category that belongs to the major category "purchasing behavior system", and indicates that the search query classified in the sub-category is a search query entered by the user with the intention of searching for products. .. The sub-classification ID "CAT12" indicates identification information for identifying the sub-classification "search for a product".

The sub-category "reserve restaurant" is a classification belonging to the major category "purchasing behavior system", and the search query classified into the sub-category is a search query input by the user with the intention of reserving the restaurant. Show that. The sub-category ID "CAT13" indicates identification information for identifying the sub-category "reserve a restaurant".

The sub-category "Purchase a product" is a category that belongs to the major category "Purchase behavior system", and the search query classified in the sub-category is a search query entered by the user with the intention of purchasing the product. show. The sub-category ID "CAT14" indicates identification information for identifying the sub-category "purchase a product".

(Category information storage unit 534)
The category information storage unit 534 stores various information about the category to which the search query belongs. Specifically, the category information storage unit 534 stores various information related to the category output by the second learning model when a search query is input to the trained second learning model. FIG. 17 shows an example of the category information storage unit according to the embodiment. In the example shown in FIG. 17, the category information storage unit 534 has items such as "search query ID", "major classification ID", "minor classification ID", and "probability (%)".

The "search query ID" indicates identification information for identifying the search query entered by the user. In the example shown in FIG. 17, the search query (search query Q11) identified by the search query ID "Q11" corresponds to the search query Q11 shown in FIG.

The "major classification ID" indicates identification information for identifying the major classification. The "minor classification ID" indicates identification information for identifying the minor classification. The "probability (%)" indicates the probability for each subclass output by the second learning model when a search query is input to the trained second learning model. In the example shown in FIG. 17, the probability (%) "90" indicates that the probability that the search query Q11 is classified into the category CAT11 is 90%.

(Model information storage unit 535)
The model information storage unit 535 stores various information about the learning model generated by the generation device 50. FIG. 18 shows an example of the model information storage unit according to the embodiment. In the example shown in FIG. 18, the model information storage unit 535 has items such as "model ID" and "model data".

The "model ID" indicates identification information for identifying the learning model generated by the generation device 50. The "model data" indicates the model data of the learning model generated by the generation device 50. For example, "model data" stores data for converting a search query into a distributed representation.

In the example shown in the first record of FIG. 18, the learning model identified by the model ID “M1” corresponds to the first model M1 shown in FIG. Further, the model data "MDT1" indicates model data (model data MDT1) of the first model M1 generated by the generation device 50.

The model data MDT1 includes an input layer into which a search query is input, an output layer, a first element which is any layer from the input layer to the output layer and belongs to a layer other than the output layer, and the first element and the first element. The output layer outputs a distributed representation of the search query input to the input layer according to the search query input to the input layer, including the second element whose value is calculated based on the weight of one element. As such, the generator 50 may be made to function.

Here, it is assumed that the model data MDT1 is realized by the regression model shown by "y = a1 * x1 + a2 * x2 + ... + ai * xi". In this case, the first element included in the model data MDT1 corresponds to input data (xi) such as x1 and x2. Further, the weight of the first element corresponds to the coefficient ai corresponding to xi. Here, the regression model can be regarded as a simple perceptron having an input layer and an output layer. When each model is regarded as a simple perceptron, the first element corresponds to any node of the input layer, and the second element can be regarded as the node of the output layer.

Further, it is assumed that the model data MDT1 is realized by a neural network having one or a plurality of intermediate layers such as DNN (Deep Neural Network). In this case, the first element included in the model data MDT1 corresponds to either the node of the input layer or the intermediate layer. Further, the second element corresponds to the node of the next stage, which is the node to which the value is transmitted from the node corresponding to the first element. Further, the weight of the first element corresponds to a connection coefficient which is a weight considered for the value transmitted from the node corresponding to the first element to the node corresponding to the second element.

The generation device 50 calculates the distributed representation using a model having an arbitrary structure such as the regression model and the neural network described above. Specifically, the model data MDT1 is set with a coefficient so as to output a distributed representation when a search query is input. The generation device 50 calculates the distributed representation using such model data MDT1.

In the above example, the model data MDT1 is a model (hereinafter referred to as model X1) that outputs a distributed representation of the search query when the search query is input. However, the model data MDT1 according to the embodiment may be a model generated based on the result obtained by repeating the input / output of data to the model X1. For example, the model data MDT1 may be a model (hereinafter referred to as model Y1) learned by inputting the distributed representation output by the model X1 when the search query is input. Alternatively, the model data MDT1 may be a model trained to input a search query and output the output value of the model Y1.

Further, when the generation device 50 performs estimation processing using GAN (Generative Adversarial Networks), the model data MDT1 may be a model constituting a part of GAN.

In the example shown in the second record of FIG. 18, the learning model identified by the model ID “M2” corresponds to the second model M2 shown in FIG. Further, the model data "MDT2" indicates model data (model data MDT2) of the second model M2 generated by the generation device 50.

The model data MDT2 includes an input layer into which a search query is input, an output layer, a first element which is any layer from the input layer to the output layer and belongs to a layer other than the output layer, and the first element and the first element. An output layer that includes a second element whose value is calculated based on the weight of one element, and the probability that the search query input to the input layer belongs to each category according to the search query input to the input layer. The generator 50 may function to output from.

Here, it is assumed that the model data MDT2 is realized by the regression model shown by "y = a1 * x1 + a2 * x2 + ... + ai * xi". In this case, the first element included in the model data MDT2 corresponds to input data (xi) such as x1 and x2. Further, the weight of the first element corresponds to the coefficient ai corresponding to xi. Here, the regression model can be regarded as a simple perceptron having an input layer and an output layer. When each model is regarded as a simple perceptron, the first element corresponds to any node of the input layer, and the second element can be regarded as the node of the output layer.

Further, it is assumed that the model data MDT2 is realized by a neural network having one or a plurality of intermediate layers such as DNN (Deep Neural Network). In this case, the first element included in the model data MDT2 corresponds to either the node of the input layer or the intermediate layer. Further, the second element corresponds to the node of the next stage, which is the node to which the value is transmitted from the node corresponding to the first element. Further, the weight of the first element corresponds to a connection coefficient which is a weight considered for the value transmitted from the node corresponding to the first element to the node corresponding to the second element.

The generation device 50 calculates the probability that the search query belongs to each category by using a model having an arbitrary structure such as the regression model and the neural network described above. Specifically, in the model data MDT2, when a search query is input, a coefficient is set so as to output the probability that the search query belongs to each category. The generation device 50 uses such model data MDT2 to calculate the probability that the search query belongs to each category.

In the above example, the model data MDT2 is a model (hereinafter referred to as model X2) that outputs a distributed representation of the search query when the search query is input. However, the model data MDT2 according to the embodiment may be a model generated based on the result obtained by repeating the input / output of data to the model X2. For example, the model data MDT2 may be a model (hereinafter referred to as model Y2) learned by inputting the distributed representation output by the model X2 when the search query is input. Alternatively, the model data MDT2 may be a model trained to input a search query and output the output value of the model Y2.

Further, when the generation device 50 performs estimation processing using GAN (Generative Adversarial Networks), the model data MDT2 may be a model constituting a part of GAN.

(Control unit 52)
Returning to the description of FIG. 13, the control unit 52 is a controller, and is stored in a storage device inside the generation device 50 by, for example, a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). It is realized by executing various programs (corresponding to an example of a generation program) using the RAM as a work area. Further, the control unit 52 is a controller, and is realized by, for example, an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

Further, the control unit 52 sends the search query input to the input layer to each layer other than the output layer by information processing according to the first model M1 (model data MDT1) stored in the model information storage unit 535. The computer is made to function so that the distributed representation is output from the output layer by performing an operation based on the first element and the weight of the first element with each element to which it belongs as the first element.

Further, the control unit 52 sends the search query input to the input layer to each layer other than the output layer by information processing according to the second model M2 (model data MDT2) stored in the model information storage unit 535. By performing an operation based on the first element and the weight of the first element with each element to which it belongs as the first element, the computer is made to function so that the probability that the search query belongs to each category is output from the output layer.

As shown in FIG. 13, the control unit 52 includes an acquisition unit 521, an extraction unit 522, and a generation unit 523, and realizes or executes the information processing operation described below. The internal configuration of the control unit 52 is not limited to the configuration shown in FIG. 13, and may be any other configuration as long as it is configured to perform information processing described later.

(Acquisition unit 521)
The acquisition unit 521 acquires various information. Specifically, the acquisition unit 521 acquires the search query input by the user from the search server 20. When the acquisition unit 521 acquires the search query input by the user, the acquisition unit 521 stores the acquired search query in the query information storage unit 531. Further, the acquisition unit 521 acquires vector information regarding a vector which is a distributed representation of the search query. When the acquisition unit 521 acquires the vector information, the acquisition unit 521 stores the acquired vector information in the vector information storage unit 532. In addition, the acquisition unit 521 acquires information that defines the search query and the classification of the category to which the search query belongs. When the acquisition unit 521 acquires the search query and the classification definition information that defines the classification of the category to which the search query belongs, the acquisition unit 521 stores the acquired classification definition information in the classification definition storage unit 533. Further, the acquisition unit 521 acquires the category information regarding the category to which the search query belongs. When the acquisition unit 521 acquires the category information, the acquisition unit 521 stores the acquired category information in the category information storage unit 534.

(Extraction unit 522)
The extraction unit 522 extracts various information. Specifically, the extraction unit 522 extracts a plurality of search queries input by the same user within a predetermined time from the search queries acquired by the acquisition unit 521. For example, the extraction unit 522 extracts a plurality of search queries in which the time interval in which each search query is input by the same user is within a predetermined time. Subsequently, the extraction unit 522 extracts a pair of search queries continuously input by the same user within a predetermined time from among a plurality of search queries input by the same user within a predetermined time. For example, the extraction unit 522 extracts a plurality of search queries in which the time interval in which each search query pair is input by the same user is within a predetermined time. For example, the extraction unit 522 is a search query Q11 (“Roppongi pasta”” which is four search queries continuously input by the same user U1 within a predetermined time among the search queries acquired by the acquisition unit 521. ), Search query Q12 (“Roppongi Italian”), search query Q13 (“Akasaka pasta”), and search query Q14 (“Azabu pasta”). When the search queries are arranged in the order in which the search queries are input, the extraction unit 522 extracts four search queries input in the order of search query Q11, search query Q12, search query Q13, and search query Q14. Subsequently, when the extraction unit 522 extracts four search queries, it is a pair of three pairs of search queries (search query Q11, search query), with two search queries adjacent in chronological order as a pair of search queries. Q12), (search query Q12, search query Q13), (search query Q13, search query Q14) are extracted. The extraction unit 522 may extract a plurality of search queries in which all the search queries are input by the same user within a predetermined time. Then, the extraction unit 522 selects two search queries from the plurality of extracted search queries regardless of whether they are adjacent to each other in chronological order, and the two selected search queries are used as a pair of search queries. It may be extracted.

Further, the extraction unit 522 extracts a predetermined search query and other search queries unrelated to the predetermined search query from the search queries acquired by the acquisition unit 521. For example, the extraction unit 522 extracts a predetermined search query from the search queries acquired by the acquisition unit 521. Subsequently, the extraction unit 522 randomly extracts other search queries from the search queries acquired by the acquisition unit 521, regardless of the predetermined search query.

(Generation unit 523)
The generation unit 523 generates various information. Specifically, the generation unit 523 learns that among the search queries acquired by the acquisition unit 521, a plurality of search queries input by the same user within a predetermined time have similar characteristics. , Generate a learning model that predicts the feature information of a given search query from a given search query. Specifically, the generation unit 523 trains a learning model so that the distributed expressions of a plurality of search queries input by the same user within a predetermined time are similar to each other, thereby performing a predetermined search from a predetermined search query. Generate a learning model that predicts query feature information. For example, the generation unit 523 generates a learning model by learning so that the distributed representations of a pair of search queries that are continuously input within a predetermined time are similar. For example, the generation unit 523 calculates the value of the similarity of the distributed representation before learning the pair of search queries. In addition, the generation unit 523 calculates the value of the similarity of the distributed representation after learning the pair of search queries. Subsequently, the generation unit 523 trains the learning model so that the value of the similarity of the distributed expression after learning is larger than the value of the similarity of the distributed expression before learning. In this way, the generation unit 523 outputs the distributed expression from the search query by training the learning model so that the two vectors, which are the pair of distributed expressions corresponding to the pair of search queries, are similar on the distributed expression space. Generate a learning model to do. More specifically, the generation unit 523 generates a learning model that outputs a distributed expression from a search query by using the DSSM technology that uses LSTM, which is a kind of RNN, for the distributed expression generation. For example, the generation unit 523 assumes that the pair of search queries input by the same user within a predetermined time has similar characteristics as the correct answer data of the learning model, and the distributed representation of the predetermined search query and the predetermined search query are predetermined. Learn so that the distributed representations of other search queries that are paired with the search query are close together in the distributed representation space. Further, when the first learning model is generated, the generation unit 523 stores the generated first learning model (model data MDT1) in the model information storage unit 535 in association with the identification information that identifies the first learning model.

[2-4. An example of the first learning model]
Here, an example of the first learning model generated by the generation device 50 will be described with reference to FIG. FIG. 19 is a diagram showing an example of the first learning model according to the embodiment. In the example shown in FIG. 19, the first learning model M1 generated by the generation device 50 is composed of three layers of LSTM RNNs. In the example shown in FIG. 19, the extraction unit 522 is a pair consisting of a search query Q11 "Roppongi pasta" and a search query Q12 "Roppongi Italian" continuously input by the same user U1 within a predetermined time. Extract search queries. The generation unit 523 inputs the search query Q11 extracted by the extraction unit 522 to the input layer of the first learning model M1 (step S41).

Subsequently, the generation unit 523 outputs a 256-dimensional vector BQV11 which is a distributed representation of the search query Q11 from the output layer of the first learning model M1. Further, the generation unit 523 inputs the search query Q12 extracted by the extraction unit 522 to the input layer of the first learning model M1. Subsequently, the generation unit 523 outputs the 256-dimensional vector BQV12, which is a distributed representation of the search query Q12, from the output layer of the first learning model M1 (step S42).

Subsequently, the generation unit 523 generates the first learning model M1 that outputs the distributed expression from the search query by learning so that the distributed expressions of the two consecutively input search queries are similar (step S43). ). For example, let Θ be the size of the angle formed by the vector BQV11 which is the distributed expression of the search query Q11 before giving feedback to the first learning model M1 (before learning) and the vector BQV12 which is the distributed expression of the search query Q12. Further, let Φ be the size of the angle formed by the vector QV11 which is the distributed expression of the search query Q11 after giving feedback to the first learning model M1 (after learning) and the vector QV12 which is the distributed expression of the search query Q12. .. At this time, the generation unit 523 trains the first learning model M1 so that Φ is smaller than Θ. For example, the generation unit 523 calculates the value of the cosine similarity between the vector BQV11 and the vector BQV12. Further, the generation unit 523 calculates the value of the cosine similarity between the vector QV11 and the vector QV12. Subsequently, the generation unit 523 prepares the learning model M1 so that the value of the cosine similarity between the vector QV11 and the vector QV12 is larger than the value of the cosine similarity between the vector BQV11 and the vector BQV12 (so that the value approaches 1). To learn. In this way, the generation unit 523 distributes from the search query by training the first learning model M1 so that the two vectors, which are a pair of distributed expressions corresponding to the pair of search queries, are similar on the distributed expression space. Generate the first learning model M1 that outputs the expression. The generation unit 523 is not limited to the cosine similarity, and may calculate the similarity between the distributed representations based on any index as long as it is an index applicable as a distance scale between vectors. Further, the generation unit 523 may train the learning model M1 based on any index as long as it is an index applicable as a distance scale between vectors. For example, the generation unit 523 calculates the value of a predetermined distance function such as the Euclidean distance between distributed expressions, the distance in a non-Euclidean space such as a twin-curved space, the Manhattan distance, and the Mahalanobis distance. Subsequently, the generation unit 523 may train the learning model M1 so that the value of a predetermined distance function between the distributed expressions (that is, the distance in the distributed expression space) becomes small.

Further, the generation unit 523 learns that a plurality of search queries including a character string separated by a predetermined delimiter have similar characteristics as a plurality of search queries input by the same user within a predetermined time. By doing so, the first learning model is generated. For example, the generation unit 523 uses a search query "Roppongi pasta" in which the characters "Roppongi" indicating a place name and "pasta" indicating a food type are separated by a space as a delimiter, and "Roppongi" indicating a place name. The first learning model is generated by learning as having similar characteristics to the search query "Roppongi Italian" separated by a space in which the character "Italian" indicating the type of food is separated.

Further, the generation unit 523 generates the first learning model by learning that a plurality of randomly extracted search queries among the search queries acquired by the acquisition unit 521 have different characteristics. Specifically, the generation unit 523 generates the first learning model by learning so that the distributed expressions of the pair of randomly extracted search queries among the search queries acquired by the acquisition unit 521 are different. do. For example, in the generation unit 523, the distributed representation of the predetermined search query extracted by the extraction unit 522 and the distributed representation of the search query randomly extracted regardless of the predetermined search query are far apart on the distributed representation space. The first learning model M1 is trained so as to be mapped.

In addition, the generation unit 523 generates the second learning model. Specifically, the generation unit 523 refers to the model information storage unit 535 and acquires the first learning model (model data MDT1 of the first learning model M1) generated by the generation unit 523. Subsequently, the generation unit 523 uses the acquired first learning model to generate a second learning model that predicts the category to which the predetermined search query belongs from the predetermined search query. When the generation unit 523 acquires the first model M1, the generation unit 523 generates the second learning model M2 by using the acquired first model M1. By retraining the first model M1, the generation unit 523 generates the second model M2 having a connection coefficient different from that of the first model M1 which is the weight of the learning model. Specifically, the generation unit 523 determines that when the search query is input to the learning model, the classification result of the distributed expression output by the learning model is learned so as to correspond to the category to which the search query belongs. A second model M2 that predicts the category to which a predetermined search query belongs is generated from the search query of.

Specifically, the generation unit 523 determines that when the search query is input to the learning model, the classification result of the distributed expression output by the learning model is learned so as to correspond to the category to which the search query belongs. A second learning model that predicts the category to which a predetermined search query belongs is generated from the search query of. The generation unit 523 generates a second learning model that outputs the probability of each category to which the search query belongs as output information when the search query is input to the learning model as input information. For example, the generation unit 523 uses the first model M1 to classify the probability that the distributed expression of the search query is classified into the category as the output information when the predetermined search query is input to the learning model as the input information. A second model M2 to be output is generated every time. When a predetermined search query is input as input information, the generation unit 523 trains the second model so that the probability that the distributed expression of the predetermined search query is classified into the correct answer category as output information exceeds a predetermined threshold. .. Then, when a predetermined search query is input as input information, the generation unit 523 sets a category in which the probability that the distributed expression of the predetermined search query belongs to the category exceeds a predetermined threshold as a predetermined search query category. Generate the second model M2 to be output. Further, when the second learning model is generated, the generation unit 523 stores the generated second learning model (model data MDT2) in the model information storage unit 535 in association with the identification information that identifies the second learning model.

For example, the generation unit 523 acquires the first model M1 (model data MDT1 of the first model M1) with reference to the model information storage unit 535 shown in FIG. Subsequently, the generation unit 523 selects a major classification of the category for classifying the search query with reference to the classification definition storage unit 533 shown in FIG. Subsequently, when the major classification is selected, the generation unit 523 learns a set of the search query and the minor classification to which the search query belongs as the learning data of the second model M2.

For example, assume that the correct answer category to which the search query Q11 (“Roppongi pasta”) belongs is CAT11 (“find a restaurant”). When the search query Q11 (“Roppongi pasta”) is input to the second model M2 as input information, the generation unit 523 uses a distributed representation of the search query Q11 (“Roppongi pasta”) from the output layer of the second model M2. A certain vector BQV11 is output. Here, the vector BQV11 is a distributed expression of the search query Q11 just output from the output layer of the second model M2, and shows a distributed expression before giving feedback to the second model M2 (before learning). In this case, the generation unit 523 has a probability that the vector BQV11, which is a distributed expression of the output search query Q11 (“Roppongi pasta”), is classified into the correct answer category CAT11 (“find a restaurant”) exceeds a predetermined threshold value. The second model M2 is trained in this way.

For example, in the generation unit 523, when the search query Q11 (“Roppongi pasta”) is input to the second model M2 before learning, the vector BQV11 which is a distributed expression is classified into CAT11 (“find a restaurant”). 80% probability of being classified as CAT12 ("find a product") 0%, 20% probability of being classified as CAT13 ("book a restaurant"), CAT14 ("buy a product") It is assumed that the probability of being classified as 0% is output. In this case, the generation unit 523 trains the second model M2 so that the probability that the vector BQV11, which is a distributed representation, is classified into CAT11 (“find a restaurant”) exceeds a predetermined threshold value (for example, 90%). .. Further, the generation unit 523 distributes the vector BQV11, which is a distributed expression, so that the probability of being classified into CAT11 (“searching for a restaurant”) exceeds a predetermined threshold value (for example, 90%). The second model M2 is trained so that the probability that the representation vector BQV11 is classified into another category CAT13 (“book a restaurant”) is reduced to 10%. Subsequently, the generation unit 523 receives the search query Q11 (“Roppongi pasta”) as input information in the trained second model M2, and the vector BQV11 is a distributed representation of the search query Q11 (“Roppongi pasta”). Since the probability of belonging to the category CAT11 ("find a restaurant") exceeds 90%, the category to which the search query belongs is output as CAT11 ("find a restaurant") as output information.

The generation unit 523 may select any number of major classifications as the major classification. Then, when the search query is input to the second model M2 as input information, the generation unit 523 determines the probability of belonging to each minor classification belonging to any number of major classifications selected by the search query as output information for each minor classification. The second model M2 to be output to may be generated. Further, the generation unit 523 may select all major classifications as the major classification. Then, the generation unit 523 may generate the second model M2 that outputs the probability of belonging to each subclass for each subclass when the search query is input to the second model M2.

[2-5. An example of the second learning model]
Here, an example of the second learning model generated by the generation device 50 will be described with reference to FIG. 20. FIG. 20 is a diagram showing an example of the second learning model according to the embodiment. In the example shown in FIG. 20, the second learning model M2 generated by the generation device 50 is generated using the first learning model M1. That is, the generation device 50 relearns the first learning model M1 to generate the second learning model M2 having a connection coefficient different from that of the first learning model M1.

More specifically, the second learning model M2 generated by the generation device 50 is composed of three layers of LSTM RNNs, like the first learning model M1. In the example shown in FIG. 20, the extraction unit 522 inputs the search query Q11 "Roppongi pasta" input by the user U1 to the input layer of the second learning model M2 (step S51).

Subsequently, the generation unit 523 outputs a 256-dimensional vector BQV11 which is a distributed representation of the search query Q11 from the output layer of the second learning model M2 (step S52).

Subsequently, the generation unit 523 outputs the probability that the vector BQV11, which is the distributed representation of the search query Q11, is classified into each category (step S53).

Subsequently, the generation unit 523 learns the second learning model M2 so as to increase the probability that the vector BQV11, which is the distributed representation of the search query Q11, is classified into the correct answer category, so that the search query category is selected from the search query. A second model to be predicted is generated (step S54).

[2-6. Flow of generation process of the first learning model]
Next, the procedure for generating the first learning model according to the embodiment will be described with reference to FIG. 21. FIG. 21 is a flowchart showing a generation processing procedure of the first learning model according to the embodiment.

In the example shown in FIG. 21, the generation device 50 acquires the search query input by the user (step S1001).

Subsequently, the generation device 50 extracts a plurality of search queries input by the same user within a predetermined time (step S1002).

Subsequently, the generation device 50 generates a first learning model that predicts the characteristic information of a predetermined search query from the predetermined search query by learning that the extracted plurality of search queries have similar characteristics (the following). Step S1003).

[2-7. Flow of generation process of the second learning model]
Next, the procedure for generating the second learning model according to the embodiment will be described with reference to FIG. 22. FIG. 22 is a flowchart showing the procedure of the generation process of the second learning model according to the embodiment.

In the example shown in FIG. 22, the generation device 50 acquires the first learning model (model data MDT1 of the first learning model M1) (step S2001).

Subsequently, the generation device 50 uses the first learning model to generate a second learning model that predicts a predetermined search query category from a predetermined search query (step S2002).

[3. effect〕
As described above, the information processing apparatus 100 according to the embodiment includes an acquisition unit 131, an estimation unit 133, and an extraction unit 134. The acquisition unit 131 is generated by using a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. A second learning model that predicts a category to which a predetermined search query belongs is acquired from a predetermined search query, which is a two-learning model. The estimation unit 133 uses the second learning model acquired by the acquisition unit 131 to estimate the category to which the target indicated by the character string input as the search query belongs. The extraction unit 134 extracts a target character string indicating an extraction target belonging to the target field from the character string based on the category estimated by the estimation unit 133.

As a result, the information processing apparatus 100 can estimate the category to which the target indicated by the unknown term belongs even when the character string input as the search query is an unknown term, and therefore belongs to a predetermined category. The character string indicating the target can be extracted with high accuracy. For example, the information processing apparatus 100 can remove an unknown term as an unnecessary character string when a category indicating a non-target field different from the target field is estimated for the unknown term. Further, even if the character string input as the inflow search query is an unknown term, the information processing apparatus 100 uses the unknown term when a category indicating a target field is estimated for the unknown term. It can be extracted as a target character string belonging to the target field. Therefore, the information processing apparatus 100 according to the present invention can appropriately extract a character string indicating an object belonging to a predetermined category.

Further, since it is generally considered that the user performs a search a plurality of times with a certain intention, it is assumed that the search queries continuously input within a predetermined time have similar search intentions. Therefore, the generation device 50 according to the present invention is similar to each other in that a plurality of search queries continuously input within a predetermined time are search queries searched under a predetermined search intention. The first model M1 is trained by regarding it as a search query having characteristics. As a result, the generation device 50 can make the first model M1 learn the characteristics of the search query in consideration of the search intention. Then, the generation device 50 efficiently generates a second model that predicts a predetermined search query category from a predetermined search query by utilizing the first model M1 that has learned the characteristics of the search query in consideration of the search intention. can do. This allows the generator 50 to classify the search query into categories that take into account the search intent of the user who entered the search query. Further, conventionally, in order to classify search queries into categories and obtain high classification accuracy, it has been necessary to prepare a sufficient amount of correct answer data. However, because the search queries themselves are diverse and have long-tailed properties, it is very laborious and difficult to label the correct answer categories that correspond to a large number of search queries. Here, the generation device 50 learns the second model starting from the first model in which the characteristics of the search query considering the search intention are learned, so that the user's search intention is used instead of labeling the correct answer category. (The context of the user who entered the search query) can be used as a kind of correct answer to train the second model that predicts the category of the search query. As a result, the generation device 50 can train the second model without manually labeling the correct answer category of the search query. That is, the second model can obtain sufficient classification accuracy even when the number of correct answer data is small. In addition, the second model can obtain even higher classification accuracy when there are many correct answer data. Therefore, the information processing apparatus 100 can improve the classification accuracy of the search query.

Further, the estimation unit 133 estimates a plurality of categories to which the object indicated by the character string belongs. Further, the estimation unit 133 outputs the probability that the object indicated by the character string belongs to each category for each category.

Thereby, the information processing apparatus 100 can simultaneously estimate the category indicating the target field and the category indicating the plurality of non-target fields as the category to which the target indicated by the character string belongs.

Further, the extraction unit 134 extracts a character string including the category indicating the target field from the categories estimated by the estimation unit 133 as the target character string.

As a result, the information processing apparatus 100 can appropriately extract a character string indicating a target belonging to the target field.

Further, the extraction unit 134 extracts a character string that does not include the unnecessary category registered as an unnecessary category from the categories estimated by the estimation unit 133 as the target character string.

As a result, the information processing apparatus 100 can appropriately remove the character string indicating the target belonging to the non-target field.

Further, the information processing apparatus 100 according to the embodiment further includes a processing unit 132. The processing unit 132 acquires the inflow search query that has flowed into the site related to the target field, and acquires the first character string obtained by removing the unnecessary character string registered as an unnecessary character string from the character string input as the inflow search query. do. The estimation unit 133 estimates the category to which the object indicated by the first character string acquired by the processing unit 132 belongs.

As a result, the information processing apparatus 100 can improve the accuracy of the input information input to the second learning model by removing the character strings that can be removed on a dictionary basis in advance. Therefore, the information processing apparatus 100 can improve the estimation accuracy of the category estimated by the second learning model.

Further, the acquisition unit 131 uses a first learning model in which the characteristics of a plurality of search queries are learned so that the distributed expressions of a pair of search queries input by the same user within a predetermined time have similar characteristics. A second learning model that predicts a category to which a predetermined search query belongs is acquired from a predetermined search query.

In general, two search queries entered by the same user in succession in a short period of time are considered to have the same search intent or similar search intents even if they are not the same. That is, it is considered that the pair of search queries that are continuously input within a predetermined time have the same search intent, or the search intents are close even if they are not the same. That is, the generation device 50 can improve the learning accuracy of the first model by learning so that the distributed expressions of the pair of search queries continuously input within a predetermined time are similar. Therefore, since the generation device 50 can generate the second model using the first model with improved learning accuracy, the learning accuracy of the second model can be improved.

Further, the acquisition unit 131 learns that a plurality of search queries including a character string delimited by a predetermined delimiter have similar characteristics as a plurality of search queries input by the same user within a predetermined time. By doing so, it is assumed that a plurality of search queries have similar characteristics, and it is a second learning model generated by using the first learning model in which the characteristics possessed by the plurality of search queries are learned, from a predetermined search query. Acquire a second training model that predicts the category to which a given search query belongs.

In general, a search query containing a plurality of character strings is considered to have a clearer search intent than a search query consisting of a single character string. That is, the generation device 50 can improve the learning accuracy of the first model by training using a search query including a character string delimited by a predetermined delimiter. Therefore, since the generation device 50 can generate the second model using the first model with improved learning accuracy, the learning accuracy of the second model can be improved.

Further, the acquisition unit 131 is a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned, and is random. It is a second learning model generated by using the first learning model that learned the characteristics of the plurality of search queries by learning that the plurality of search queries extracted from the above have different characteristics. A second learning model that predicts the category to which a predetermined search query belongs is acquired from the search query of. Further, the acquisition unit 131 is a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned, and is random. It is a second learning model generated by using the first learning model that learned the characteristics of a plurality of search queries by learning so that the distributed expressions of the pair of search queries extracted in the above are different. A second learning model that predicts the category to which a predetermined search query belongs is acquired from the search query of.

In general, a plurality of randomly extracted search queries are search queries that are input independently of each other, so it is considered that the search intents are different or the search intents are distant. Therefore, the generation device 50 according to the present invention is a search query having different characteristics from each other in that a plurality of randomly extracted search queries are search queries searched under different search intentions. Assuming that, the learning model M1 is trained. As a result, the learning model can learn correct answer data, which is a pair of search queries with similar search intentions, and incorrect answer data, which is a pair of search queries with distant search intentions. That is, the generation device 50 can improve the learning accuracy of the first model. Therefore, since the generation device 50 can generate the second model using the first model with improved learning accuracy, the learning accuracy of the second model can be improved.

Further, the acquisition unit 131 is generated by using a first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned. In the second learning model, when the search query is input to the second learning model, the classification result of the distributed expression output by the second learning model is learned so as to correspond to the category to which the search query belongs. Then, the second learning model that predicts the category to which the predetermined search query belongs is acquired from the predetermined search query. Further, the acquisition unit 131 acquires a second learning model generated by using a first learning model that outputs a distributed representation of the predetermined search query as output information when a predetermined search query is input as input information. do. Further, the acquisition unit 131 acquires a second learning model that outputs the probability that the search query belongs to a category as output information when the search query is input to the second learning model as input information.

As a result, the generation device 50 utilizes a distributed expression including the characteristics of the search query considering the search intention, and classifies the search query into a category considering the search intention of the user who input the search query. It can be generated efficiently. That is, the generation device 50 makes it possible to classify the search query into a category considering the search intention of the user who has input the search query. Therefore, the generation device 50 can improve the classification accuracy of the search query.

[4. Hardware configuration]
Further, the information processing apparatus 100 according to the above-described embodiment and the generation apparatus 50 according to the embodiment are realized by, for example, a computer 1000 having a configuration as shown in FIG. 23. FIG. 23 is a hardware configuration diagram showing an example of a computer that realizes the functions of the information processing device 100 and the generation device 50. The computer 1000 includes a CPU 1100, a RAM 1200, a ROM 1300, an HDD 1400, a communication interface (I / F) 1500, an input / output interface (I / F) 1600, and a media interface (I / F) 1700.

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

The HDD 1400 stores a program executed by the CPU 1100, data used by such a program, and the like. The communication interface 1500 receives data from another device via a predetermined communication network and sends the data to the CPU 1100, and transmits the data generated by the CPU 1100 to the other device via the predetermined communication network.

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

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

For example, when the computer 1000 functions as the information processing device 100 or the generation device 50, the CPU 1100 of the computer 1000 realizes the functions of the control unit 130 or the control unit 52 by executing the program loaded on the RAM 1200. The CPU 1100 of the computer 1000 reads these programs from the recording medium 1800 and executes them, but as another example, these programs may be acquired from another device via a predetermined communication network.

Although some of the embodiments of the present application have been described in detail with reference to the drawings, these are examples, and various modifications are made based on the knowledge of those skilled in the art, including the embodiments described in the disclosure column of the invention. It is possible to carry out the present invention in other modified forms.

[5. others〕
Further, among the processes described in the above-described embodiments and modifications, all or part of the processes described as being automatically performed can be manually performed, or are described as being manually performed. It is also possible to automatically perform all or part of the performed processing by a known method. In addition, information including processing procedures, specific names, various data and parameters shown in the above documents and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the information shown in the figure.

Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically distributed in any unit according to various loads and usage conditions. Can be integrated and configured.

Further, the above-described embodiments and modifications can be appropriately combined as long as the processing contents do not contradict each other.

Further, the above-mentioned "section, module, unit" can be read as "means" or "circuit". For example, the extraction unit can be read as an extraction means or an extraction circuit.

1 Information processing system 10 User terminal 20 Search server 50 Generation device 100 Information processing device 110 Communication unit 120 Storage unit 121 Query information storage unit 122 Unnecessary character string storage unit 123 Unnecessary category storage unit 124 Model information storage unit 130 Control unit 131 Acquisition unit 132 Processing unit 133 Estimating unit 134 Extracting unit

Claims (15)

In the second learning model generated by using the first learning model in which the characteristics of the plurality of search queries are trained assuming that the plurality of search queries input by the same user within a predetermined time have similar characteristics. There is an acquisition unit that acquires the second learning model that predicts the category to which the predetermined search query belongs from the predetermined search query.
Using the second learning model acquired by the acquisition unit, an estimation unit that estimates the category to which the target indicated by the character string input as a search query belongs, and an estimation unit.
An information processing apparatus including an information processing unit that extracts a target character string indicating an extraction target belonging to a target field corresponding to the category from the character string based on a category estimated by the estimation unit. .. The estimation unit
The information processing apparatus according to claim 1, wherein a plurality of categories to which the object indicated by the character string belongs are estimated. The estimation unit
The information processing apparatus according to claim 2, wherein the probability that the object indicated by the character string belongs to each category is output for each category. The extraction unit
The information processing according to any one of claims 1 to 3, wherein a character string including a category indicating the target field is extracted as the target character string from the categories estimated by the estimation unit. Device. The extraction unit
One of claims 1 to 4, wherein a character string that does not include an unnecessary category registered as an unnecessary category is extracted as the target character string from the categories estimated by the estimation unit. The information processing device described. A processing unit that acquires an inflow search query that has flowed into a site related to the target field, and acquires a first character string obtained by removing an unnecessary character string registered as an unnecessary character string from the character string input as the inflow search query. Further prepared,
The estimation unit
The information processing apparatus according to any one of claims 1 to 5, wherein the category to which the object indicated by the first character string acquired by the processing unit belongs is estimated. The acquisition unit
The second learning generated by using the first learning model in which the characteristics of a plurality of search queries are learned as the distributed representations of a pair of search queries input by the same user within a predetermined time have similar characteristics. The information processing apparatus according to any one of claims 1 to 6, further comprising acquiring the second learning model, which is a model and predicts a category to which the predetermined search query belongs from a predetermined search query. .. The acquisition unit
As a plurality of search queries input by the same user within a predetermined time, the plurality of search queries including a character string separated by a predetermined delimiter are learned as having similar characteristics. Assuming that the search queries have similar characteristics, it is a second learning model generated by using the first learning model that has learned the characteristics of the plurality of search queries, and is the predetermined search query from the predetermined search query. The information processing apparatus according to any one of claims 1 to 7, wherein the second learning model for predicting the category to which the second learning model belongs is acquired. The acquisition unit
A first learning model in which a plurality of search queries input by the same user within a predetermined time have similar characteristics and the characteristics of the plurality of search queries are learned, and a plurality of randomly extracted search queries. A second learning model generated by using the first learning model in which the characteristics of the plurality of search queries are learned by learning the search queries as having different characteristics, from a predetermined search query. The information processing apparatus according to any one of claims 1 to 8, wherein the second learning model for predicting the category to which the predetermined search query belongs is acquired. The acquisition unit
A first learning model in which a plurality of search queries entered by the same user within a predetermined time learn the characteristics of the plurality of search queries as having similar characteristics, and a pair of randomly extracted search queries. It is a second learning model generated by using the first learning model that learned the characteristics of the plurality of search queries by learning so that the distributed expressions of the search queries are different, and is from a predetermined search query. The information processing apparatus according to any one of claims 1 to 9, wherein the second learning model for predicting a category to which the predetermined search query belongs is acquired. The acquisition unit
In the second learning model generated by using the first learning model in which the characteristics of the plurality of search queries are trained assuming that the plurality of search queries input by the same user within a predetermined time have similar characteristics. Therefore, when the search query is input to the second learning model, the classification result of the distributed expression output by the second learning model is learned so as to correspond to the category to which the search query belongs. The information processing apparatus according to any one of claims 1 to 10, wherein the second learning model that predicts the category to which the predetermined search query belongs is acquired from the search query of. The acquisition unit
When a predetermined search query is input as input information, the second learning model generated by using the first learning model that outputs a distributed representation of the predetermined search query as output information is acquired. The information processing apparatus according to any one of claims 1 to 11. The acquisition unit
When a search query is input to the second learning model as input information, a request characterized by acquiring the second learning model that outputs the probability that the search query belongs to the category as output information for each category. Item 6. The information processing apparatus according to any one of Items 1 to 12. It is an information processing method executed by a computer.
In the second learning model generated by using the first learning model in which the characteristics of the plurality of search queries are trained assuming that the plurality of search queries input by the same user within a predetermined time have similar characteristics. Therefore, the acquisition process of acquiring the second learning model that predicts the category to which the predetermined search query belongs from the predetermined search query, and
Using the second learning model acquired by the acquisition process, an estimation process for estimating the category to which the target indicated by the character string input as a search query belongs, and an estimation process.
An information processing method comprising an extraction step of extracting a target character string indicating an extraction target belonging to a target field corresponding to the category from the character string based on a category estimated by the estimation step. .. In the second learning model generated by using the first learning model in which the characteristics of the plurality of search queries are trained assuming that the plurality of search queries input by the same user within a predetermined time have similar characteristics. Therefore, the acquisition procedure for acquiring the second learning model that predicts the category to which the predetermined search query belongs from the predetermined search query, and the acquisition procedure.
Using the second learning model acquired by the acquisition procedure, an estimation procedure for estimating the category to which the target indicated by the character string input as a search query belongs, and an estimation procedure.
Based on the category estimated by the estimation procedure, the computer is made to execute an extraction procedure for extracting a target character string indicating an extraction target belonging to the target field corresponding to the category from the character string. Information processing program.

Images