JP2013161136A - Teacher data generation device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a teacher data generation device, method and program that can efficiently generate teacher data.SOLUTION: A teacher data generation server 3 includes: a teacher data DB 31 for storing teacher data; a query reception section 32 for receiving a first query and a second query different from the first query; a category DB 33 for storing a category and information showing a parental relation between the category and other categories in association with each other; a degree of association calculation section 34 for calculating a degree of association between a first category corresponding to the first query received in the query reception section 32 and a second category corresponding to the second query on the basis of the category DB 33; an identity information acquisition section 35 for acquiring identity information showing each character of the first query and the second query; and a storage control section 36 for storing the first query, the second query, the degree of association and the identity information in association with each other as teacher data in the teacher data DB 31.

Description

  The present invention relates to a teacher data generation apparatus, method, and program for generating teacher data used for machine learning.

  Conventionally, classification using machine learning is performed when a computer classifies a large amount of data. Among the machine learning methods, in a method called supervised learning, a computer generates, for sample data, teacher data in which feature information that is information representing data characteristics and classification results are predetermined. After the computer learns the classification pattern based on the teacher data, the computer classifies the unclassified data based on the previous learning (see, for example, Patent Document 1).

JP 2005-181928 A

  By the way, in supervised learning disclosed in Patent Document 1, generation of teacher data from sample data is performed manually. For this reason, since the quality of the teacher data is guaranteed, the supervised learning disclosed in Patent Document 1 has high data classification accuracy.

  However, in the supervised learning disclosed in Patent Document 1, since teacher data is generated manually, there is a problem that it takes time to generate teacher data. Thus, a method of reducing the number of teacher data is conceivable. However, if the number of teacher data is reduced, the accuracy of machine learning is lowered.

  An object of the present invention is to provide a teacher data generation apparatus, method, and program capable of efficiently generating teacher data.

  (1) A teacher data generation device that generates teacher data used for machine learning, a query acquisition unit that acquires a first query and a second query different from the first query, a category, and the category A first category corresponding to the first query acquired by the acquisition unit based on a category information storage unit that associates and stores information indicating a parent-child relationship with another category, and acquired by the acquisition unit Relevance calculating means for calculating relevance with the second category corresponding to the second query, information representing the properties of the first query and the second query, the first query, and the second query. Feature information acquisition means for acquiring at least one of information representing the relationship between the first query, the second query, the degree of association, and the feature information as feature information. Tutor data generating apparatus and a storage control means for storing in a predetermined storage means as teacher data put communicating.

  (2) Based on the category information storage unit, the degree-of-relevance calculation unit includes a path between the first category and a category that is the highest level with respect to the first category, and the second category. The teacher data generation device according to (1), wherein a path with a category that is the highest level with respect to the second category is specified, and the degree of association is calculated based on the specified path.

  (3) The teacher data generation device according to (2), wherein the relevance calculation unit calculates the relevance based on a common path length from the highest category among the identified paths.

  (4) The teacher data generation device according to (2), wherein the relevance calculation unit calculates the relevance based on a common path length in the identified paths.

  (5) The relevance calculation means calculates the relevance based on a common path length from the highest category among the identified paths and a common path length in the identified paths. The teacher data generation device according to 2).

  (6) A teacher data generation method in which a computer executes generation of teacher data used for machine learning, a query acquisition step of acquiring a first query and a second query different from the first query; A first category corresponding to the first query acquired in the query acquisition step based on category information storage means for storing the category and information indicating a parent-child relationship between the category and another category in association with each other; A degree-of-association calculating step for calculating a degree of association with the second category corresponding to the second query acquired in the query acquisition step, feature information representing the properties of the first query and the second query, A feature information acquisition step of acquiring at least one of information representing the relationship between the first query and the second query as feature information; The first query, the second query, the relevance and the teacher data generating method and a storage control step, the computer executes the feature information that are correlated is stored in a predetermined storage means as teacher data.

  (7) A teacher data generation program for causing a computer to generate teacher data used for machine learning, a query acquisition step for acquiring a first query and a second query different from the first query, a category A first category corresponding to the first query acquired in the query acquisition step based on category information storage means that stores information indicating a parent-child relationship between the category and another category in association with each other, and A degree-of-association calculating step of calculating a degree of association with the second category corresponding to the second query acquired in the query acquisition step, feature information representing the properties of the first query and the second query, and the first A feature information acquisition step for acquiring, as feature information, at least one of information representing the relevance of one query and the second query. Flop and the first query, the second query, the relevance and teacher data generating program to be executed by the storage control step, to the computer to be stored in a predetermined storage means as teacher data in association with the feature information.

  ADVANTAGE OF THE INVENTION According to this invention, it becomes possible to provide the teacher data generation apparatus, method, and program which can generate teacher data efficiently.

It is a figure which shows a similarity determination system. It is a block diagram which shows the function structure of the query extraction server which comprises a similarity determination system, a teacher data generation server, and a similarity determination server. It is a figure which shows search log DB. It is a figure which shows teacher data DB. It is a figure which shows category DB. It is a figure which shows query DB. It is a flowchart which shows the flow of the teacher data generation process which a query extraction server and a teacher data generation server perform. It is a flowchart which shows the flow of the similarity determination process which a similarity determination server performs.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Outline of similarity determination system]
FIG. 1 is a diagram showing a similarity determination system 1 according to the present embodiment.
The similarity determination system 1 includes a query extraction server 2, a teacher data generation server 3 as a teacher data generation device, and a similarity determination server 4. The query extraction server 2, the teacher data generation server 3, and the similarity determination server 4 are configured to be communicable via a communication network N such as the Internet.

  In the similarity determination system 1, the query extraction server 2 extracts a combination of the first query and the second query from the click log used in the search engine. The teacher data generation server 3 calculates the degree of association based on the category information for the first query and the second query extracted by the query extraction server 2, and the properties of the first query and the second query, respectively. And at least one of information representing the relationship between the first query and the second query is acquired as feature information. Then, the teacher data generation server 3 generates teacher data by associating the first query, the second query, the degree of association, and the feature information. The similarity determination server 4 performs machine learning based on the teacher data generated by the teacher data generation server 3 and generates an identification model. And the similarity determination server 4 determines a similarity based on an identification model about the combination of the query which exists separately.

Next, the functional configuration of each server constituting the similarity determination system 1 will be described.
FIG. 2 is a block diagram showing a functional configuration of the query extraction server 2, the teacher data generation server 3, and the similarity determination server 4 constituting the similarity determination system 1 according to the present embodiment.

[Configuration of query extraction server 2]
The query extraction server 2 is configured by one or a plurality of general computers. A typical computer includes a central processing unit, a storage device, a communication device, an input device, a display device, and a bus line to which these devices are connected. The central processing unit is configured by a CPU or the like and functions as a query extraction unit 22 described later. The storage device includes a memory (RAM, ROM), a hard disk (HDD), an optical disk (CD, DVD, etc.), and functions as a search log DB 21 described later. The communication device includes various wired and wireless LAN devices. The display device includes various displays such as a liquid crystal display and a plasma display. The input device includes a touch panel or a keyboard and a pointing device (mouse, tracking ball, etc.). In such a general computer, the CPU comprehensively controls the query extraction server 2 and appropriately reads and executes various programs such as a program for teacher data generation processing, thereby cooperating with the hardware described above. The various functions according to the present invention are realized.

  As shown in FIG. 2, the query extraction server 2 includes a search log DB 21 and a query extraction unit 22.

  FIG. 3 is a diagram showing the search log DB 21 according to the present embodiment. Here, the search log is information indicating a history of selecting one URL on the search result page when a user of a user terminal (not shown) performs a search in the search engine via the user terminal. . The search log DB 21 stores the terminal IP address, the search date and time, the session ID, the rank, the search query, and the URL in association with each other as a search log.

  The terminal IP address is the IP address of the user terminal that performed the search. The search date and time is the date and time when the search was performed on the user terminal. The session ID is a session ID between the user terminal and the search engine when a search is performed on the user terminal. The rank is the page rank of the Web page corresponding to the URL selected from the search result in the user terminal. The search query is one or more search queries received by the search engine from the user terminal in one search. The URL is a URL selected by the user terminal on the search result page.

  The query extraction unit 22 refers to the search log stored in the search log DB 21 and extracts a combination of the two search queries when a search is performed using the two search queries. The query extraction unit 22 transmits the combination of the extracted two search queries to the teacher data generation server 3.

[Configuration of Teacher Data Generation Server 3]
The teacher data generation server 3 is configured by one or a plurality of general computers, like the query extraction server 2. Description of the same configuration as the query extraction server 2 in the teacher data generation server 3 is omitted. The central processing unit of the teacher data generation server 3 functions as a query receiving unit 32, a relevance calculation unit 34, a feature information acquisition unit 35, and a storage control unit 36, which will be described later. The storage device of the teacher data generation server 3 functions as a teacher data DB 31 and a category DB 33 described later.

  As shown in FIG. 2, the teacher data generation server 3 includes a teacher data DB 31, a query reception unit 32, a category DB 33, a relevance calculation unit 34, a feature information acquisition unit 35, and a storage control unit 36. I have.

FIG. 4 is a diagram showing the teacher data DB 31 according to the present embodiment.
As shown in FIG. 4, the teacher data DB 31 stores teacher data. The teacher data includes a first query, a second query, a degree of association, and feature information.
The degree of association indicates the degree of association between the first query and the second query. The feature information is information representing at least one of the properties of the first query and the second query and information representing the relationship between the first query and the second query, and is composed of a plurality of elements. .

  The query receiving unit 32 acquires two queries by receiving a combination of two queries from the query extraction server 2. Of these two queries, one query is the first query, and the other query different from the first query is the second query. That is, the query receiving unit 32 automatically receives the first query and the second query from the query extraction server 2 without receiving a manual input operation.

The category DB 33 stores information indicating a parent-child relationship between a category and another category in association with each other.
FIG. 5 is a diagram showing the category DB 33 according to the present embodiment. As shown in FIG. 5, the category DB 33 stores the category and the other category directly connected to the category in association with each other.

The category is, for example, the name of each directory in the directory type search service. In the category DB 33, a plurality of other categories are associated below the category. Thereby, in the category DB 33, a certain category may exist under a plurality of categories.
In the present embodiment, the category DB 33 includes categories and other categories. However, a category code for identifying a category and a category code for identifying another category may also be stored. .

  Based on the category DB 33, the relevance calculating unit 34 includes a first category corresponding to the first query received by the query receiving unit 32, and a second category corresponding to the second query received by the query receiving unit 32. The relevance of is calculated.

  Specifically, when the query receiving unit 32 receives the first query and the second query in the query receiving unit 32, the relevance calculation unit 34 refers to the category DB 33, and determines the category corresponding to the first query and the second query. Identify the corresponding category. In the present embodiment, a category corresponding to a query refers to a category having a character string that matches a character string constituting the query. Here, let the category corresponding to the first query be the first category. The category corresponding to the second query is set as the second category.

  Subsequently, based on the category DB 33, the degree-of-association calculation unit 34 specifies a category that is the highest category of the first category and that is the highest category. For example, the relevance calculation unit 34 repeats the process of extracting the upper category until there is no higher category that is directly associated with the extracted category. When there is no higher category than the extracted category, the relevance calculation unit 34 specifies the extracted category as the highest category. Subsequently, the degree-of-association calculation unit 34 specifies a path between the first category and the highest category. This identified path is referred to as a first path.

  For example, when the first query is “Spain” and the data shown in FIG. 5 is stored in the category DB 33, the relevance calculation unit 34 selects “ Extract "region". Then, the relevance calculation unit 34 identifies the first path as “region / country / Spain”. In this example, one example is shown as the first path, but a plurality of first paths are specified. This is because an upper category (for example, “country”) of “Spain” exists under a plurality of categories.

  Subsequently, the relevance calculation unit 34 identifies the highest category for the second category. The method for specifying the highest category is the same as the method for specifying the highest category for the first category. Subsequently, the relevance calculation unit 34 identifies the path between the second category and the highest category. This identified path is referred to as a second path.

  For example, when the second query is “Barcelona” and the data shown in FIG. 5 is stored in the category DB 33, the relevance calculation unit 34 selects “ Extract "region". Then, the relevance calculation unit 34 identifies the second path as “region / country / Spain / local government / Catalogonia / city / Barcelona”. In this example, a plurality of second paths are specified in the same manner as the first path.

Subsequently, the degree-of-association calculation unit 34 calculates the degree of association between the first category and the second category by performing the following processing based on the first path and the second path.
In other words, the relevance calculation unit 34 calculates the length of a common path from the highest category when the highest category matches the first path and the second path. Then, the relevance calculation unit 34 calculates the relevance for the length of the path based on the following equation (1).

ここで、ｍａｘ｛｜Ｄ｜，｜Ｄ’｜｝は、特定された複数の第１のパスのうち最も長い距離と、特定された複数の第２のパスのうち最も長い距離とによって求められる。
また、｜Ｐ（Ｄ，Ｄ’）｜は、特定された第１のパスと第２のパスとの組み合わせのうち、最上位のカテゴリから共通するパスが最も長い組み合わせにおける、共通するパスの長さである。例えば、第１のパスが、「地域／国／スペイン」であり、第２のパスが、「地域／国／スペイン／自治体／カタロニア／市／バルセロナ」であり、第１のパスと第２のパスとは、「地域／国／スペイン」で共通する。そして、この第１のパスと第２のパスとの組み合わせが、共通するパスの最も長い組み合わせである場合、｜Ｐ（Ｄ，Ｄ’）｜は、「３」となる。

Here, max {| D |, | D ′ |} is determined by the longest distance among the plurality of identified first paths and the longest distance among the plurality of identified second paths. .
| P (D, D ′) | is the length of the common path in the combination having the longest common path from the highest category among the identified combinations of the first path and the second path. That's it. For example, the first pass is “Region / Country / Spain” and the second pass is “Region / Country / Spain / Local Government / Catalania / City / Barcelona”. The first pass and the second pass A pass is common to "region / country / spain". When the combination of the first path and the second path is the longest combination of the common paths, | P (D, D ′) | is “3”.

In addition, instead of calculating the degree of association as described above, the degree-of-association calculation unit 34 specifies a common path for the first path and the second path, and determines the degree of association based on the common path. You may make it calculate. In this case, the relevance calculation unit 34 calculates the relevance for the length of this path based on the following equation (2).

  Here, C (D, D ′) is the length of the common path in the combination having the longest common path among the identified combinations of the first path and the second path. This common path is not based on the highest category as in the formula (1), and the first path and the second path may be common from the middle of the path. Good.

  In addition, the relevance calculation unit 34 may calculate the relevance based on the values calculated in both the formulas (1) and (2) described above. For example, the degree-of-relevance calculation unit 34 weights the value calculated by Expression (1) and the value calculated by Expression (2), and then adds these values to add the degree of association. Is calculated.

  The relevance calculating unit 34 calculates the relevance between the first query and the second query when one of the category corresponding to the first query and the category corresponding to the second query cannot be specified. Suppose that it failed. When the calculation of the degree of association fails, the degree-of-association calculation unit 34 omits the processing of the feature information acquisition unit 35 and the processing of the storage control unit 36 that are performed following the calculation of the degree of association. Thereby, the teacher data generation server 3 does not employ the first query and the second query for which the degree of association could not be calculated as teacher data. Thereby, the teacher data generation server 3 can prevent information with unknown relevance from being included in the teacher data, and can improve the accuracy of the teacher data.

The feature information acquisition unit 35 acquires feature information representing the properties of the first query and the second query.
Specifically, the feature information acquisition unit 35 obtains feature information including a facet extraction characteristic, a query text characteristic, a result click characteristic, and a query session co-occurrence characteristic for each of the first query and the second query. Get from.

  The facet extraction characteristics are, for example, the probability that the first query and the second query in the search log DB 21 are used at once, the probability that is used at the top of the search keyword, the probability that is used at other than the top of the search keyword, This is information indicating the probability of being present in the same session, the frequency of clicking, and the like.

  The query text characteristic is information indicating, for example, the length of characters, the amount of words configured, the Levenshtein distance based on a multibyte character criterion, and the like for each of the first query and the second query.

  The result click characteristic is information calculated based on the amount of information in the search results of each of the first query and the second query, for example.

  In addition, although the feature information acquisition part 35 decided to acquire feature information from query DB42, it is not restricted to this. For example, the feature information acquisition unit 35 may calculate the query text from each of the first query and the second query.

  The storage control unit 36 associates the first query and the second query received by the query receiving unit 32, the relevance calculated by the relevance calculation unit 34, and the feature information acquired by the feature information acquisition unit 35 with teacher data. The teacher data is stored in the teacher data DB 31.

[Configuration of similarity determination server 4]
Similarity determination server 4 is configured by one or a plurality of general computers, similar to query extraction server 2 and teacher data generation server 3. Description of the configuration similar to that of the query extraction server 2 in the similarity determination server 4 is omitted. The central processing unit of the similarity determination server 4 functions as a model generation unit 41 and a similarity determination unit 43 described later. The storage device of the similarity determination server 4 functions as a query DB 42 described later.

  As shown in FIG. 2, the similarity determination server 4 includes a model generation unit 41, a query DB 42, and a similarity determination unit 43.

  The model generation unit 41 performs machine learning with reference to the teacher data DB 31 to generate an identification model between the first query and the second query. Specifically, the model generation unit 41 determines the relationship between the values of the first query, the second query, and the feature data and the relevance (similarity) based on the teacher data stored in the teacher data DB 31. Machine learning is performed to generate an identification model. The model generation unit 41 generates an identification model based on all data stored in the teacher data DB 31.

  FIG. 6 is a diagram showing the query DB 42 according to the present embodiment. As shown in FIG. 6, the query DB 42 stores the first query, the second query, and the feature information in association with each other. That is, in the present embodiment, it is assumed that the similarity between the first query and the second query stored in the query DB 42 is unknown although the feature information is calculated in advance.

  The similarity determination unit 43 uses the identification model generated by the model generation unit 41 and the feature information stored in the query DB 42 to use the relevance (similarity of the first query and the second query stored in the query DB 42). Degree). The similarity determination unit 43 may receive the first query and the second query from another server and determine the relevance of the received first query and second query.

[Operation]
Next, the operation of the similarity determination system 1 will be described with reference to FIGS.
FIG. 7 is a flowchart showing the flow of teacher data generation processing executed by the query extraction server 2 and the teacher data generation server 3.

[Operations of Query Extraction Server 2 and Teacher Data Generation Server 3]
The query extraction unit 22 of the query extraction server 2 refers to the search log stored in the search log DB 21 and extracts two search queries (step S1).
Subsequently, the query extraction unit 22 of the query extraction server 2 transmits the two search queries extracted in step S1 to the teacher data generation server 3 (step S2).

Subsequently, the query receiving unit 32 of the teacher data generation server 3 receives two queries from the query extraction server 2 (step S3).
Subsequently, the relevance calculation unit 34 of the teacher data generation server 3 corresponds to the first category corresponding to the first query accepted in step S3 and the second query accepted in step S3 based on the category DB 33. The degree of association with the second category is calculated (step S4).

  Subsequently, the feature information acquisition unit 35 of the teacher data generation server 3 acquires feature information representing the properties of the first query and the second query from the query DB 42 (step S5). Note that the order of the process of step S4 and the process of step S5 may be the reverse order of the order described.

  Subsequently, the storage control unit 36 of the teacher data generation server 3 associates the first query and the second query received in step S3, the relevance calculated in step S4, and the feature information acquired in step S5 with the teacher. This teacher data is stored in the teacher data DB 31 (step S6).

[Operation of similarity determination server 4]
FIG. 8 is a flowchart showing the flow of similarity determination processing executed by the similarity determination server 4.
The model generation unit 41 of the similarity determination server 4 performs machine learning on the relationship between the values of the first query, the second query, and the feature data and the similarity based on the teacher data stored in the teacher data DB 31. To generate an identification model (step S11).

  The similarity determination unit 43 of the similarity determination server 4 uses the identification model generated by the model generation unit 41 and the feature information stored in the query DB 42 to use the first query and the second query stored in the query DB 42. The similarity (relevance) of the query is determined (step S12).

  As described above, in the similarity determination system 1 according to the present embodiment, the teacher data generation server 3 corresponds to the first query received by the query reception unit 32 based on the category DB 33 by the relevance calculation unit 34. The degree of association between the first category and the second category corresponding to the second query received by the query receiving unit 32 is calculated. Then, the teacher data generation server 3 acquires feature information representing the properties of the first query and the second query by the feature information acquisition unit 35, and the first query, the second query, and the relevance degree by the storage control unit 36. And the feature information are associated with each other and stored in the teacher data DB 31 as teacher data.

  As described above, the teacher data generation server 3 can automatically generate teacher data without manually generating teacher data, so that teacher data can be generated efficiently.

  In addition, the teacher data generation server 3 causes the relevance calculation unit 34 to determine, based on the category DB 33, the path of the first category, the highest category with respect to the first category, the second category, The path with the category that is the highest level with respect to the second category is specified, and the degree of association is calculated based on the specified path. Therefore, the teacher data generation server 3 can calculate the degree of association using the hierarchical structure of categories stored in the category DB 33.

  In addition, the teacher data generation server 3 calculates the relevance degree based on the common path length from the highest category among the identified paths by the relevance degree calculation unit 34. When the path from the highest category is common, the first category and the second category corresponding to these paths are in a vertical relationship, so compared to the case where the path from the highest category is not common. It can be said that they are similar. Therefore, the teacher data generation server 3 can accurately calculate the degree of association between the first query and the second query.

  In addition, the teacher data generation server 3 calculates the relevance level based on the common path length in the specified path by the relevance level calculation unit 34. When the paths are partially common, it can be said that the first category and the second category corresponding to these paths exist in common under a certain category. For this reason, it can be said that the first category and the second category are similar as compared to the case where the paths are not common. Therefore, the teacher data generation server 3 can accurately calculate the degree of association between the first query and the second query.

  Further, the teacher data generation server 3 uses the relevance calculation unit 34 to determine the relevance based on the common path length from the highest category among the identified paths and the common path length in the identified paths. Is calculated. In this way, the teacher data generation server 3 can calculate the degree of association based on both the commonality of the path from the highest category and the partial commonality of the path.

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

  For example, in the present embodiment, the case where the teacher data generation server 3 is realized as a single server has been described as an example. However, the functions of the teacher data generation server 3 are distributed and implemented in a plurality of servers. It is also possible to realize the function of the teacher data generation server 3 by cooperation of the entire servers.

  Further, in the present embodiment, the category DB 33 is stored in the teacher data generation server 3, but the category DB 33 may be stored in a server different from the teacher data generation server 3. In this case, the teacher data generation server 3 accesses the server in which the category DB 33 is stored, and calculates the degree of association based on the category DB 33 stored in the server.

  In the present embodiment, the similarity determination system 1 is composed of three servers, that is, the query extraction server 2, the teacher data generation server 3, and the similarity determination server 4. However, the present invention is not limited to this. Absent. For example, the similarity determination system 1 may realize the functions provided by these servers with a single server, or may distribute the functions with four or more servers.

  In the present embodiment, the teacher data generation server 3 stores the generated teacher data in the teacher data DB 31. However, the teacher data is stored in a server different from the teacher data generation server 3. Also good. In this case, the similarity determination server 4 accesses a server in which teacher data is stored, and generates an identification model based on the teacher data stored in the server.

  In the present embodiment, when the relevance calculation fails, the relevance calculation unit 34 omits the processing of the feature information acquisition unit 35 and the storage control unit 36 that are performed following the relevance calculation. Although it was assumed, it is not limited to this. For example, when the relevance calculation unit 34 fails to calculate the relevance level, error information indicating the failure may be output to the feature information acquisition unit 35 and the storage control unit 36. And in each of the feature information acquisition part 35 and the storage control part 36, when error information is received, you may make it not perform the process which concerns on each function.

1 similarity determination system 2 query extraction server 21 search log DB
22 Query extraction unit 3 Teacher data generation server 31 Teacher data DB
32 Query receiver 33 Category DB
34 relevance calculation unit 35 feature information acquisition unit 36 storage control unit 4 similarity determination server 41 model generation unit 42 query DB
43 Similarity determination unit

Claims (7)

A teacher data generation device that generates teacher data used for machine learning,
Query acquisition means for acquiring a first query and a second query different from the first query;
A first category corresponding to the first query acquired by the query acquisition unit based on a category information storage unit that stores a category and information indicating a parent-child relationship between the category and another category; Relevance calculation means for calculating relevance with the second category corresponding to the second query acquired by the query acquisition means;
Feature information acquisition means for acquiring, as feature information, at least one of information representing the properties of the first query and the second query and information representing the relevance of the first query and the second query;
Storage control means for associating the first query, the second query, the relevance and the feature information and storing them in a predetermined storage means as teacher data;
A teacher data generation device comprising: Based on the category information storage unit, the relevance calculation unit is configured to pass the first category, a path of a category that is the highest level with respect to the first category, the second category, and the second category. A path with a category that is the highest level for the two categories is calculated, and the relevance is calculated based on the specified path.
The teacher data generation device according to claim 1. The relevance calculating means calculates the relevance based on a common path length from the highest category among the identified paths.
The teacher data generation device according to claim 2. The relevance calculating means calculates the relevance based on a common path length in the identified path;
The teacher data generation device according to claim 2. The relevance calculation means calculates the relevance based on a common path length from the highest category among the identified paths and a common path length in the identified paths.
The teacher data generation device according to claim 2. A teacher data generation method in which a computer executes generation of teacher data used for machine learning,
A query acquisition step of acquiring a first query and a second query different from the first query;
A first category corresponding to the first query acquired in the query acquisition step based on category information storage means for storing the category and information indicating a parent-child relationship between the category and another category in association with each other; A relevance calculation step of calculating relevance with the second category corresponding to the second query acquired in the query acquisition step;
A feature information acquisition step of acquiring, as feature information, at least one of feature information representing the properties of each of the first query and the second query and information representing a relationship between the first query and the second query;
A storage control step of associating the first query, the second query, the degree of association, and the feature information with each other and storing them in a predetermined storage unit as teacher data;
A teacher data generation method in which a computer executes. A teacher data generation program for causing a computer to generate teacher data used for machine learning,
A query acquisition step of acquiring a first query and a second query different from the first query;
A first category corresponding to the first query acquired in the query acquisition step based on category information storage means for storing the category and information indicating a parent-child relationship between the category and another category in association with each other; A relevance calculation step of calculating relevance with the second category corresponding to the second query acquired in the query acquisition step;
A feature information acquisition step of acquiring, as feature information, at least one of feature information representing the properties of each of the first query and the second query and information representing a relationship between the first query and the second query;
A storage control step of associating the first query, the second query, the degree of association, and the feature information with each other and storing them in a predetermined storage unit as teacher data;
A teacher data generation program for causing a computer to execute.

Images