JP2018156332A - Generation device, generation method, and generation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve accuracy of distributed representations generated from triples.SOLUTION: A generation device of the present disclosure comprises: an extraction unit configured to extract a first element and association information included in a predetermined triple; a determination unit configured to stochastically determine whether to limit a selection source for a second element or not; a selection unit configured to select at least a second element not included in the predetermined triple from among second elements included in arbitrary triples based on a result of the determination; and a generation unit configured to generate each distributed representation in such a way that a sum of distributed representation of the extracted first element and a distributed representation of the association information is not similar to a distributed representation of the selected second element.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a generation device, a generation method, and a generation program.

  2. Description of the Related Art Conventionally, a technique for constructing data called a triple including two related elements such as RDF (Resource Description Framework) and relation information indicating the relation between the elements is known. In order to facilitate such a search for triples, the distributed representation of each element and relationship information is learned so that the sum of the distributed representation of the first element and the distributed representation of the relationship information becomes the distributed representation of the second element. A technique for searching for a second element constituting a triple by calculating the sum of distributed expressions when a first element and relation information are received as a search query is known.

"Translating Embeddings for Modeling Multi-relational Data" Antoine Bordes Et al.

  However, the above-described conventional technology may not be able to guarantee the accuracy of distributed representation.

  For example, in the prior art, the sum of the distributed representation of the first element of a triple and the distributed information of the relationship information is similar to the distributed representation of the second element of the triple, Learns the distributed representation of each element and relation information so that they are not similar. For this reason, if there is a bias in the field to which the triple to be learned belongs, the learning in the field with a small number of triples does not proceed, and the accuracy of the distributed representation that is finally obtained decreases.

  The present application has been made in view of the above, and an object thereof is to improve the accuracy of distributed representation generated from triples.

  The generation apparatus according to the present application includes an extraction unit that extracts a first element and relationship information included in a predetermined triple, a determination unit that determines probabilistically whether to limit a selection source of the second element, A selection unit that selects at least a second element that is not included in the predetermined triple among second elements included in an arbitrary triple based on the determination result, a distributed expression of the extracted first element, and a distributed expression of relation information And a generation unit that generates each distributed representation so that the sum of and the distributed representation of the selected second element are not similar to each other.

  According to one aspect of the embodiment, it is possible to improve the accuracy of the distributed representation generated from the triple.

FIG. 1 is a diagram illustrating an example of processing executed by the information providing apparatus according to the embodiment. FIG. 2 is a diagram illustrating a configuration example of the information providing apparatus according to the embodiment. FIG. 3 is a diagram illustrating an example of information registered in the entity database according to the embodiment. FIG. 4 is a diagram illustrating an example of information registered in the relational information database according to the embodiment. FIG. 5 is a diagram illustrating an example of information registered in the distributed representation database according to the embodiment. FIG. 6 is a diagram illustrating an example of a second element selected as incorrect answer data by the information providing apparatus according to the embodiment. FIG. 7 is a flowchart illustrating an example of a flow of generation processing executed by the information providing apparatus according to the embodiment. FIG. 8 is a diagram illustrating an example of a hardware configuration.

  Hereinafter, a mode for carrying out a generation device, a generation method, and a generation program according to the present application (hereinafter referred to as “embodiment”) will be described in detail with reference to the drawings. Note that the generation device, the generation method, and the generation program according to the present application are not limited to the embodiment. In the following embodiments, the same parts are denoted by the same reference numerals, and redundant description is omitted.

[Embodiment]
[1. Regarding the processing provided by the information providing device]
First, an example of a generation process executed by an information providing apparatus as an example of a generation apparatus will be described with reference to FIG. FIG. 1 is a diagram illustrating an example of processing executed by the information providing apparatus according to the embodiment. In the following description, as a process executed by the information providing apparatus 10, a generation process for generating a distributed representation of information registered as a triple and a search for searching for information using the distributed representation generated by the generation process are performed. Processing will be described. In the following description, a process in which the information providing apparatus 10 “generates” a distributed representation of each information will be described. However, such processing appropriately “learns” the value of an appropriate distributed representation corresponding to each data. It is assumed that this is a process of “generating” a distributed representation.

[1-1. Overview of information providing device)
The information providing apparatus 10 is an information processing apparatus that can communicate with the user terminal 100 via a predetermined network N (for example, see FIG. 2) such as the Internet, and is realized by, for example, a server apparatus or a cloud system. Is done. The information providing apparatus 10 may be able to communicate with an arbitrary number of user terminals 100 via the network N.

  The user terminal 100 is an information processing device used by a user who requests information retrieval, and is realized by an information processing device such as a PC (Personal Computer), a server device, or a smart device. For example, the user terminal 100 transmits a search query for searching for information to the information providing apparatus 10. In such a case, the information providing apparatus 10 searches for information stored in the RDF format, that is, information corresponding to the search query from the triples, and provides the search result to the user terminal 100.

  Here, triples searched by the information providing apparatus 10 will be described. For example, the information providing apparatus 10 stores triple information composed of first and second elements called entities or the like and relation information indicating the relationship between the elements as triples. For example, the information providing apparatus 10 includes the entity of the subject (S: Subject) as the first element, the entity as the object (O: Object) as the second element, and the relationship information as the predicate (P: Predicate). Memorize pairs as triples.

  For example, the information providing apparatus 10 corresponds to various things in the world, such as various things that can be the subject of a person, an object, a building, etc. in the real world, attributes such as occupation and nationality, various states and events as entities. It is possible to systematically manage various knowledge by storing information to be stored and storing relationship information between entities. For example, when the entity # 1 corresponding to the predetermined person #A, the entity # 2 corresponding to the occupation “politician”, and the relationship information indicating “occupation” are defined as triples, the triple is “person #A”. "Occupation" can be shown to be "politician". As described above, the information providing apparatus 10 systematically manages knowledge using triples.

  The entity corresponding to the person and the entity corresponding to the person name may be different entities. In such a case, the name of a person can be held by storing the entity corresponding to the person, the relationship information indicating “name”, and the entity corresponding to the person name as triples. Such systematic management of knowledge using triples is used in, for example, a knowledge database.

  The entity may be not only text indicating things but also arbitrary information such as still images indicating things, moving images, sounds, web contents, URLs (Uniform Resource Locators) of web contents. Moreover, the entity does not need to be the information itself indicating things, and may be set as an entity indicating a concept.

[1-2. About generation processing)
Here, in order to facilitate the search for triples, a distributed representation of each element and relationship information is generated so that the sum of the distributed representation of the first element and the distributed representation of the relationship information becomes the distributed representation of the second element. A technique for searching for a second element constituting a triple by calculating the sum of distributed expressions when a first element and relation information are received as a search query is known. When such a distributed expression is generated, the information providing apparatus 10 calculates the sum of the distributed expression of the first element and the distributed expression of the relationship information when receiving the first element and the related information as a search query. By searching for a distributed expression similar to the calculated distributed expression, the first element accepted as a search query and the second element of the triple including the relationship information can be searched by a neighborhood search.

  However, there are cases in which the accuracy of distributed representation cannot be ensured with the prior art. For example, in the prior art, the sum of the distributed representation of the first element included in a triple and the distributed representation of the relationship information is similar to the second element included in the triple, that is, the distributed representation of correct data. The distributed expression of the second element not included, that is, the distributed expression is learned so as not to be similar to the incorrect answer data. However, in such a conventional technique, in the distributed expression space, the second element similar to the second element that is the correct answer data is located closest to the sum of the distributed expression of the first element and the distributed expression of the relationship information. A distributed representation may be placed. For this reason, in the prior art, the accuracy of the distributed representation can only be guaranteed up to a predetermined accuracy.

  In addition, when selecting incorrect answer data, the second element is selected at random from the stored triples. For this reason, when the field to which the second element belongs (arbitrary criteria that can classify the second element such as a person's name or occupation) is set, there is a bias in the number of second elements belonging to each field. In such a case, there is a risk that learning of distributed expressions in a field where the number of second elements to which the member belongs is small does not advance, and the accuracy deteriorates.

  Therefore, the information providing apparatus 10 generates a distributed representation of each element by executing the following generation process. First, the information providing apparatus 10 extracts a first element and relation information included in a predetermined triple. Subsequently, the information providing apparatus 10 determines probabilistically whether or not to limit the selection source of the second element.

  And the information provision apparatus 10 selects at least the 2nd element which is not contained in a predetermined triple among the 2nd elements contained in arbitrary triple based on a determination result. For example, when the information providing apparatus 10 determines that the selection source of the second element is limited, the information providing apparatus 10 is a second element that is not included in the predetermined triple among the second elements included in the arbitrary triple, and the predetermined field The second element belonging to is selected.

  As a more specific example, when the information providing apparatus 10 determines to limit the selection source of the second element, the second element included in the predetermined triple among the second elements included in the arbitrary triple. The second element that belongs to the same or similar field and is not included in the predetermined triple is selected. On the other hand, when determining that the selection source of the second element is not limited, the information providing apparatus 10 selects the second element that is not included in the predetermined triple among the second elements included in the arbitrary triple. Then, the information providing apparatus 10 distributes each element and the relationship information so that the sum of the extracted distributed representation of the first element and the distributed representation of the relationship information is not similar to the distributed representation of the selected second element. Generate a representation.

  More specifically, the information providing apparatus 10 performs positive sampling and negative sampling when generating the distributed representation. Here, positive sampling is a process of selecting a positive example used for learning distributed expressions. For example, the first element and the relationship information included in a predetermined triple and the second element ( That is, the correct data) is selected as a correct pair. Negative sampling is a process of selecting a negative example used for learning distributed expressions. For example, the first element and the relationship information included in the predetermined triple and the second element (not included in the predetermined triple) ( That is, it is a process of selecting the incorrect answer data) as an incorrect answer pair.

  When performing such negative sampling, the information providing apparatus 10 determines probabilistically whether or not to limit the selection destination of the second element. That is, the information providing apparatus 10 sets a stochastic bias in the selection destination of the second element to be a training target in the generation of the distributed expression (that is, the distributed expression training). When the information providing apparatus 10 determines that the selection destination of the second element is not limited, the information providing apparatus 10 does not include the second element that is not included in the predetermined triple among the second elements included in all the stored triples. Select randomly as correct data. On the other hand, when the information providing apparatus 10 determines to limit the selection destination of the second element, the information providing apparatus 10 is a second element belonging to the same or similar field as the second element included in the predetermined triple, A second element that is not included is randomly selected as incorrect answer data.

  As a result of such processing, the information providing apparatus 10 selects different data in the same or similar field as the correct answer data as the incorrect answer data in the negative sampling. Then, the information providing apparatus 10 can select data having an error as incorrect data although it is semantically close to the correct data. When the distributed representation is learned using such incorrect answer data, the relationship between the second elements belonging to the same or similar field can be dropped into the distributed representation space. In other words, for a set of a certain first element and relationship information, the relationship between the second element that is correct data and the second element that is similar to the correct data but is incorrect data may be dropped into the distributed expression space. it can. As a result, the information providing apparatus 10 can improve the accuracy of the generated distributed expression.

[1-3. (Example of generation process)
Next, an example of a generation process executed by the information providing apparatus 10 will be described with reference to FIG. In the following description, a triple including the second element serving as correct data, that is, a predetermined triple to be processed is described as a correct triple.

  First, the information providing apparatus 10 selects a correct triple from the stored triples, and extracts the first element and the relationship information included in the correct triple (step S1). For example, when the triple T1 including the first element “S1”, the relation information “P1”, and the second element “O1” is the correct triple, the information providing apparatus 10 sets the first element “S1” and the relation information “P1”. Is extracted.

  Subsequently, the information providing apparatus 10 calculates the accuracy d indicating the similarity between the second elements (step S2). For example, the information providing apparatus 10 extracts the second elements “O2”, “O3”, and “O4” included in other triples based on the second element “O1” included in the correct triple. The information providing apparatus 10 calculates the accuracy d12 indicating the similarity between “O1” and “O2”, the accuracy d13 between “O1” and “O3”, and the accuracy d14 between “O1” and “O4”. To do.

  Note that the information providing apparatus 10 may calculate the accuracy based on an arbitrary index as long as it calculates a semantic similarity, that is, a semantic similarity, as the accuracy. For example, the information providing apparatus 10 may calculate the accuracy indicating the similarity of the type of the second element, such as whether the second element is a name, occupation, or position. Further, the information providing apparatus 10 may calculate the accuracy based on a field previously given to the second element. The information providing apparatus 10 may calculate the accuracy based on the identity of the relationship information included in the triple together with the second element. For example, when the relationship information included in the triple together with the second element is the same, the information providing apparatus 10 may calculate the accuracy indicating that it is similar or belongs to the same field. That is, the information providing apparatus 10 can arbitrarily indicate whether or not the second element that is incorrect data belongs to the same field as the second element that is correct data (whether or not they are similar). Any information may be calculated as the accuracy d based on the information.

  Subsequently, the information providing apparatus 10 probabilistically determines whether or not to limit the field that is the source of selection of incorrect answer data (step S3). For example, each time the information providing apparatus 10 learns the distributed expression, the information providing apparatus 10 determines whether or not to limit the field with the probability “β”. As a more specific example, when the value of the probability “β” is “0.4”, the information providing apparatus 10 determines that the field is limited with a probability of 40 percent (that is, 40%). For example, the information providing apparatus 10 may adopt a predetermined value in the range of “0” to “0.4” as the value of the probability “β”, or any value (for example, “0”). In the range of “1” to “1”, etc.), the value may be dynamically changed as the distributed expression is learned.

  For example, in the example illustrated in FIG. 1, all categories CA include second elements “O1” to “O5”. In such a case, the information providing apparatus 10 categorizes each second element based on the calculated accuracy d. For example, when the accuracy d12 between “O1” and “O2” and the accuracy d14 between “O1” and “O4” are higher than a predetermined threshold value, the information providing apparatus 10 has the second element “O1”. ”,“ O2 ”, and“ O4 ”are classified into the category C1. Note that a threshold value of the accuracy d for determining whether or not they belong to the same field may be a predetermined value, or may be dynamically changed according to learning.

  And the information provision apparatus 10 produces | generates an incorrect answer pair and a correct answer pair, and learns a distributed expression (step S4). For example, the information providing apparatus 10 selects the second element included in the correct triple as the correct answer data. For example, when the information providing apparatus 10 determines not to limit the field, the second element “O2” that is not included in the correct triple among the second elements “O1” to “O5” belonging to all categories CA. ] To “O5”, the second element to be incorrect answer data is randomly selected. On the other hand, when the information providing apparatus 10 determines to limit the field, the information providing apparatus 10 randomly selects a second element not included in the correct triple from the second elements belonging to the predetermined category. More specifically, the information providing apparatus 10 randomly selects incorrect answer data from the second elements “O2” and “O4” belonging to the same category C1 as the correct answer data.

  Then, the information providing apparatus 10 learns distributed expressions. For example, the information providing apparatus 10 generates a pair of the first element “S1” and the relationship information “P1” and the second element “O1” as the correct answer pair. Further, the information providing apparatus 10 generates a pair of the first element “S1” and the relationship information “P1” and the second element “O4” as an incorrect answer pair. Then, the information providing apparatus 10 is similar to the distributed expression O1 of the second element “O1” in which the sum of the distributed expression S1 of the first element “S1” and the distributed expression P1 of the relationship information “P1” is similar to the first element “S1”. The value of each distributed expression is learned so that the sum of the distributed expression S1 of “S1” and the distributed expression P1 of the relationship information “P1” is not similar to the distributed expression O4 of the second element “O4”.

  Moreover, the information provision apparatus 10 repeatedly performs the process of step S1-S4. As a result, the information providing apparatus 10 can probabilistically generate a distributed expression reflecting the presence of incorrect answer data similar to the correct answer data, thereby improving the accuracy of the distributed expression of each element and each related information. Can be made.

  Next, an example of search processing executed by the information providing apparatus 10 will be described. First, the information providing apparatus 10 receives a search query from the user terminal 100 (step S5). For example, the information providing apparatus 10 receives the first element “S1” and the related information “P1” as a search query from the user terminal 100.

  In such a case, the information providing apparatus 10 searches for similar distributed expressions by adding the distributed expressions of the search query (step S6). For example, the information providing apparatus 10 identifies the distributed representation “S1” of the first element “S1” and the distributed representation “P1” of the relationship information “P1” from the generated distributed representations, and Calculate the sum. Then, the information providing apparatus 10 performs a neighborhood search for a distributed expression similar to the generated sum value.

  Here, with the generation process described above, the information providing apparatus 10 can drop the relationship between the correct answer data and the incorrect answer data similar to the correct answer data into the distributed expression. As a result, for example, the distributed expression most similar to the sum of the distributed expression “S1” and the distributed expression “P1” corresponds to the first element “S1” corresponding to the distributed expression “S1” and the distributed expression “P1”. There is a high probability of being a distributed expression corresponding to the second element of the correct triple including the relation information “P1”.

  Therefore, the information providing apparatus 10 searches for a distributed expression (for example, distributed expression “O1”) that is most similar to the sum of the distributed expression “S1” and the distributed expression “P1”, and the second corresponding to the searched distributed expression. Identify the element. And the information provision apparatus 10 outputs the specified 2nd element to the user terminal 100 as a search result (step S7). For example, the information providing apparatus 10 specifies a predetermined number of distributed expressions in the order close to the sum of the distributed expression “S1” and the distributed expression “P1”, and sets the second element corresponding to each specified distributed expression. The data may be output in a ranking format (that is, in the order in which the distributed representation is close to the sum).

[1-4. (Setting dynamic probability)
In the above description, the information providing apparatus 10 determines whether or not to limit the field from which incorrect answer data is selected with a predetermined probability “β”. Here, the information providing apparatus 10 may dynamically change the value of the probability “β” according to the progress of learning. For example, the information providing apparatus 10 sets the value of “β” to “0” until it learns the distributed expression a predetermined number of times, and after it learns the predetermined number of times, the value of “β” May be raised. That is, the information providing apparatus 10 may determine probabilistically whether or not to limit the field from which incorrect answer data is selected when the number of learnings exceeds a predetermined threshold. Further, the information providing apparatus 10 may increase the value of “β” as the number of learning increases.

  Further, the information providing apparatus 10 may probabilistically determine whether or not to limit the field from which incorrect answer data is selected according to the accuracy of the distributed representation. For example, the information providing apparatus 10 determines the ratio of the distributed representation of the second element included in the correct triple to the distributed representation closest to the sum of the distributed representation of the first element and the relationship information included in the correct triple. The accuracy value is calculated every time learning is performed. Then, when the calculated accuracy satisfies a predetermined condition, the information providing apparatus 10 determines probabilistically whether to limit the field from which the incorrect answer data is selected, and the accuracy satisfies the predetermined condition. If not, the selection source may not be limited.

  For example, if the information providing apparatus 10 continues to learn distributed expressions and the accuracy does not exceed a predetermined threshold, whether or not to limit the field from which incorrect answer data is selected. It may be determined probabilistically. In addition, when the accuracy exceeds a predetermined threshold, the information providing apparatus 10 may determine probabilistically whether or not to limit the field from which incorrect data is selected. Further, if the rate of increase in accuracy does not change for a predetermined period in spite of continuing learning, the information providing apparatus 10 probabilistically determines whether to limit the field from which incorrect answer data is selected. You may decide.

[1-5. (Setting accuracy)
In the above description, the information providing apparatus 10 selects, as the incorrect answer data, the second element whose accuracy with the correct answer data is higher than the predetermined threshold when limiting the field from which the incorrect answer data is selected. . However, the embodiment is not limited to this. For example, the information providing apparatus 10 may select incorrect answer data from a second element (for example, an entity such as a name or a face photograph) of the same type as the second element that becomes correct answer data.

  Further, the information providing apparatus 10 may adopt a classification based on an arbitrary criterion. For example, when the correct answer data indicates “the name of the politician's wife”, the information providing apparatus 10 may set the second element indicating “name” as the second element belonging to the same field as the correct answer data. The second element indicating “name” may be the second element belonging to the same field as the correct answer data, and the second element indicating “name of the politician's wife” may be the second element belonging to the same field as the correct answer data. In other words, the information providing apparatus 10 may act on a field set on an arbitrary basis as long as the second element belonging to the same or similar field as the second element which is correct answer data is used as incorrect answer data.

  In addition, for example, the information providing apparatus 10 may narrow the field for selecting incorrect answer data every time learning progresses or when the rate of increase in accuracy of classification information is delayed. That is, the information providing apparatus 10 may dynamically set the size of a field for selecting incorrect answer data.

  Moreover, the information provision apparatus 10 may set these fields based on the identity of the relationship information included in the triple together with each second element. For example, for the second element included in the plurality of triples including the relationship information indicating the “name” of the first element, the information providing apparatus 10 and the second element belonging to the same field (for example, the field “name”) May be handled. Moreover, the information provision apparatus 10 is good also considering the 2nd element contained in the other triple containing the relationship information similar to the relationship information contained in the correct answer triple as the 2nd element which belongs to the field | area similar to correct data. That is, when the information providing apparatus 10 determines to limit the selection source, the second element that is not included in the correct triple among the second elements included in any triple that includes the same relationship information as the correct triple information. May be selected as incorrect answer data.

  In addition, the information provision apparatus 10 may implement | achieve the various processes mentioned above by calculating the accuracy of 2nd elements. For example, the information providing apparatus 10 may set a binary accuracy indicating whether or not they belong to the same field. Further, the information providing apparatus 10 calculates the accuracy based on the identity and similarity of the type of the second element, the identity and similarity of the relationship information included in the triple together with the second element, and the calculated accuracy Whether or not the second elements belong to the same or similar field may be determined based on whether or not exceeds a predetermined threshold. For example, when it is determined that the selection source of the second element is limited, the information providing apparatus 10 is a second element that is not included in the correct triple among the second elements included in any triple, and is included in the correct triple. A second element whose accuracy is within a predetermined range may be selected.

  Further, the information providing apparatus 10 may calculate the accuracy by any method as long as it calculates the accuracy indicating the semantic similarity between the second elements. For example, if the information providing apparatus 10 calculates the accuracy based not only on the type of the second element or the relationship information included in the triple together with the second element, but also on the semantic similarity between the second elements, The accuracy may be calculated by any method.

  In addition, when the information providing apparatus 10 determines to limit the selection source, the second element belonging to a field more similar to the field to which the second element included in the correct answer triple belongs as the accuracy of the distributed representation increases. It may be selected as incorrect answer data. That is, as the accuracy of the distributed representation increases, the information providing apparatus 10 increases the accuracy threshold when determining whether the classification is the same as the second element that is the correct answer data, so that the incorrect answer data is obtained. The fields to be selected may be gradually narrowed.

  In addition, the information providing apparatus 10 may select a second element belonging to a field in which the number of second elements to which it belongs is smaller than other fields as incorrect answer data. For example, the information providing apparatus 10 specifies the second element of the triple including the same relationship information as the relationship information included in the correct triple as the second element belonging to the same classification as the correct data. In such a case, the information providing apparatus 10 identifies the field to which the identified second element belongs with a predetermined granularity, and counts the number of second elements belonging to the identified field. And the information provision apparatus 10 may preferentially learn the field | area with few selection opportunities by selecting incorrect answer data from the 2nd element which belongs to the field | area with the fewest number counted.

[2. Configuration of information providing device]
Hereinafter, an example of the functional configuration of the information providing apparatus 10 will be described. FIG. 2 is a diagram illustrating a configuration example of the information providing apparatus according to the embodiment. As illustrated in FIG. 2, the information providing apparatus 10 includes a communication unit 20, a storage unit 30, and a control unit 40.

  The communication unit 20 is realized by, for example, a NIC (Network Interface Card). The communication unit 20 is connected to the network N by wire or wireless, and transmits / receives information to / from the user terminal 100.

  The storage unit 30 is realized by, for example, a semiconductor memory device such as a RAM (Random Access Memory) or a flash memory, or a storage device such as a hard disk or an optical disk. In addition, the storage unit 30 stores an entity database 31, a relationship information database 32, and a distributed representation database 33.

  Hereinafter, an example of information registered in each of the databases 31 to 33 will be described with reference to FIGS. In the entity database 31, information on entities, that is, first elements and second elements included in a triple is registered. For example, FIG. 3 is a diagram illustrating an example of information registered in the entity database according to the embodiment. As shown in FIG. 3, information having items such as “entity ID”, “data type”, and “data” is registered in the entity database 31.

  Here, “entity ID” is an identifier of an entity. The “data type” is information indicating the type of information corresponding to the entity indicated by the associated “entity ID”. “Data” is information corresponding to the entity indicated by the associated “entity ID”.

  For example, in the example illustrated in FIG. 3, the entity ID “S1”, the data type “person name”, and the data “name # 1” are registered in association with each other. Such information indicates that the entity indicated by the entity ID “S1” is an entity corresponding to “person name”, and the person name is “name # 1”. In the example shown in FIG. 3, conceptual values such as “name # 1” are described, but actually, in the entity database 31, the names, photographs, and birth dates of persons corresponding to the associated entities. Various types of information indicating the date and the like are registered as data.

  In the relationship information database 32, relationship information between entities, that is, relationship information indicating the relationship between the first element and the second element included in the triple is registered. For example, FIG. 4 is a diagram illustrating an example of information registered in the relational information database according to the embodiment. In the example illustrated in FIG. 4, information including items such as “relation information ID”, “type”, “first element”, and “second element” is registered in the relationship information database 32.

  Here, the “relation information ID” is an identifier for identifying the relationship information. Further, the “type” is information indicating what kind of relationship between the elements the relationship information indicated by the “relation information ID” indicates. The “first element” and the “second element” are the first element and the second element in which the relationship information indicated by the associated “relation information ID” indicates the relationship.

  For example, in the example illustrated in FIG. 4, the relationship information ID “P1”, the type “spouse”, the first element “S1”, and the second element “O1” are registered in association with each other. In such information, the relationship information indicated by the relationship information ID “P1” is information indicating the relationship between the first element “S1” and the second element “O1”, and the second element “O1” is the first element. It shows that it is a “spouse” of “S1”. That is, such information indicates that the relationship information indicated by the relationship information ID “P1”, the first element “S1”, and the second element “O1” form a triple.

  In the distributed expression database 33, distributed expressions of each entity and relation information are registered. For example, FIG. 5 is a diagram illustrating an example of information registered in the distributed expression database according to the embodiment. As shown in FIG. 5, items such as “element ID / relation information ID” and “distributed expression” are registered in the distributed expression database 33. The “element ID / relation information ID” is an entity ID or relationship information ID corresponding to the distributed representation. The “distributed expression” is a distributed expression of the entity or the relationship information indicated by the associated “element ID / relationship information ID”.

  For example, in the example shown in FIG. 5, the element ID / relation information ID “S1” and the distributed expression “distributed expression # 1” are registered in the distributed expression database 33 in association with each other. Such information indicates that the distributed representation of the element indicated by the element ID / relationship information ID “S1” is the distributed representation “distributed representation # 1”. In the example shown in FIG. 5, a conceptual value such as “distributed expression # 1” is described. However, in practice, a multi-dimensional quantity that is a distributed expression is registered in the distributed expression database 33.

  Returning to FIG. 2, the description will be continued. The control unit 40 is a controller. For example, various programs stored in a storage device inside the information providing apparatus 10 are stored in a RAM or the like by a processor such as a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). Is implemented as a work area. The control unit 40 is a controller, and may be realized by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).

  As illustrated in FIG. 2, the control unit 40 includes a calculation unit 41, an extraction unit 42, a determination unit 43, a selection unit 44, a generation unit 45, and a search unit 46. The calculation unit 41 calculates the accuracy indicating the similarity between the second elements included in the triple. More specifically, the calculation unit 41 calculates the accuracy indicating the semantic similarity between the second elements. For example, the calculation unit 41 refers to the relationship information database 32 and identifies an entity registered as the second element. Further, the calculation unit 41 identifies the type and data of data corresponding to the identified entity from the entity database 31 and identifies the relationship information included in the triple together with the identified entity from the relationship information database 31.

  And the calculation part 41 is based on the identity and similarity of the data type corresponding to the entity, the identity and similarity of the type of the relationship information included in the triple together with the specified entity, etc. The accuracy indicating the semantic similarity is calculated. Note that the calculation unit 41 may calculate the accuracy between the second element and the other second element included in the triple selected as the correct triple by the extraction unit 42.

  The extraction unit 42 extracts the first element and the relationship information included in the predetermined triple. For example, the extraction unit 42 refers to the relationship information database 32 and selects one triple to be processed as a correct triple from the triples. Then, the extraction unit 42 specifies the entity of the first element and the relationship information included in the selected correct triple. That is, the extraction unit 42 extracts an entity included as information corresponding to the subject in the correct triple as the first element, and extracts an element included as information corresponding to the predicate in the correct triple as the relationship information.

  The determination unit 43 determines probabilistically whether or not to limit the selection source of the second element serving as incorrect answer data. For example, each time the extraction unit 42 selects the correct answer triple, the determination unit 43 probabilistically determines whether or not the selection source of the second element that is the incorrect answer data is limited. When the learning process is continuously performed, the determination unit 43 refers to the distributed representation database 33 and calculates the accuracy of the distributed representation. The determination unit 43 probabilistically determines whether or not to limit the selection source when the accuracy of the distributed representation satisfies a predetermined condition, and selects when the accuracy of the distributed representation does not satisfy the predetermined condition. You may decide not to limit the origin. For example, when the rate of increase in accuracy of the distributed representation does not change for a predetermined period, the determination unit 43 may determine probabilistically whether to limit the selection source. For example, the determination unit 43 may determine that the selection source of the second element is limited with a probability of 40% or less.

  When determining that the selection source of the second element is not limited, the selection unit 44 selects, as incorrect answer data, the second element that is not included in the correct triple among the second elements included in any triple. In addition, when the selection unit 44 determines to limit the selection source of the second element, the selection unit 44 is a second element that is not included in the correct triple among the second elements included in any triple, and belongs to a predetermined field. Select the second element. More specifically, the selection unit 44 selects an element included in the triple as information corresponding to the predicate as the second element serving as incorrect answer data.

  For example, if the selection unit 44 determines to limit the selection source of the second element, the selection unit 44 belongs to the same or similar field as the second element included in the correct triple among the second elements included in the arbitrary triple. A second element that is two elements and is not included in the correct triple is selected. In addition, when the selection unit 44 determines to limit the selection source, the second part not included in the correct triple among the second elements included in any triple including the same relationship information as the relationship information included in the correct triple is included. Select an element.

  To give a more specific example, when the determination unit 43 determines that the selection source of the second element that is incorrect data is not limited, the selection unit 44 refers to the relation information database 32 and converts it to a correct triple. From the second elements other than the second element included, a second element that is randomly incorrect data is selected. On the other hand, when the determination unit 43 determines that the selection source of the second element that is incorrect data is limited, the selection unit 44 uses the accuracy calculated by the calculation unit 41 to use the second element included in the correct answer triple. That is, the other 2nd element in which the accuracy between the 2nd element used as correct answer data is in a predetermined range is specified. And the selection part 44 selects the 2nd element used as incorrect answer data at random from the specified 2nd element.

  When the learning process is continuously performed, the selection unit 44 refers to the distributed representation database 33 and calculates the accuracy of the distributed representation. And the selection part 44 may select the 2nd element which belongs to the field | area similar to the field | area to which the 2nd element used as correct data belongs as incorrect data as the precision of distributed representation rises. For example, as the accuracy of the distributed representation increases, the selection unit 44 selects the second element that is incorrect data from the second elements that have a higher accuracy with the second element that is correct data. You may choose. In addition, the selection unit 44 may select a second element belonging to a field in which the number of second elements to which it belongs is smaller than other fields.

  Here, FIG. 6 is a diagram illustrating an example of a second element selected as incorrect answer data by the information providing apparatus according to the embodiment. FIG. 6 shows an example of the number (data amount) of the second element belonging to each field as well as the field to which each of the second elements “O1” to “O5” belongs.

  In the example illustrated in FIG. 6, the second element “O1” is included in the same triple together with the relationship information “P1”. The second element “O2” is an entity having a different type from the other second elements “O1”, “O3”, and “O4”, and is included in the triple that includes the same relationship information as the relationship information “P1”. Shall not. The second elements “O3” and “O5” are entities of the same type as the second element “O1”, and are included in a triple including the same relationship information as the relationship information “P1”. Further, it is assumed that the second element “O4” is included in the triple together with other related information “P2” similar to the related information “P1”.

  In such a case, the extraction unit 42 selects a triple including the first element “S”, the relationship information “P1”, and the second element “O1” as the correct triple CT. Then, the selection unit 44 selects the second element “O1” included in the correct triple CT as the correct data CS1.

  Subsequently, when the determination unit 43 determines that the selection source is not limited, the selection unit 44 includes the second elements “O1” to “O5” belonging to all categories CA that are not included in the correct triple CT. One second element is randomly selected as incorrect data from the two elements “O2” to “O5”. For example, the selection unit 44 selects the second element “O2” as the incorrect answer data NS1.

  On the other hand, when the determination unit 43 determines that the selection source is limited, the selection unit 44 determines that the second element “O1” selected as the correct answer data as the predetermined field is another second field that is the same as or similar to the second element “O1”. Identify the element. For example, the selection unit 44 specifies the second elements “O3” and “O5” included in the triple including the relationship information “P1”, similarly to the correct triple CT.

  Here, the selection unit 44 counts the data amounts of the fields to which the second elements “O3” and “O5” belong, respectively. In the example illustrated in FIG. 6, the data amount of the category CX to which the second element “O3” belongs is smaller than the data amount of the category CZ to which the second element “O5” belongs. Therefore, the selection unit 44 selects the second element “O3” belonging to the field with a smaller data amount as the incorrect answer data NS2.

  The selection unit 44 may expand the triple to be selected by selecting the second element from triples including the relationship information “P2” similar to the relationship information “P1” of the correct answer triple. For example, the selection unit 44 specifies a triple including the relationship information “P2” similar to the relationship information “P1” of the correct answer triple, and selects the second element “O4” included in the specified triple as the incorrect answer data NS3. May be.

  Returning to FIG. 2, the description will be continued. The generation unit 45 generates each distributed representation so that the sum of the extracted distributed representation of the first element and the distributed representation of the relationship information is not similar to the distributed representation of the selected second element. For example, the generation unit 45 adds the distributed representation of the first element included in the correct triple and the distributed representation of the relationship information included in the correct triple to the distributed representation of the second element that is the correct data, Distributed representation of each element and relationship information so that the sum of the distributed representation of the first element included and the distributed representation of the relationship information included in the correct triple is different from the distributed representation of the second element that is incorrect data Is generated. Then, the generation unit 45 registers the generated distributed expression in the distributed expression database 33.

  As a specific method for generating the distributed representation, any method can be applied as long as it is a distributed representation generation method using negative sampling.

  The search unit 46 executes a search process using a distributed expression. For example, the search unit 46 receives the first element and the related information from the user terminal 100 as a search query. In such a case, the search unit 46 refers to the distributed expression database 33 and identifies the distributed expression of the first element and the distributed expression of the relationship information received as the search query. Then, the search unit 46 calculates the sum of the specified distributed expressions, and specifies the distributed expression having the closest distance in the distributed expression space from the distributed expression database 33. Then, the search unit 46 reads the entity data corresponding to the specified distributed representation from the entity database 31, and transmits the read data to the user terminal 100.

  When the search unit 46 receives a first element or relationship information that is not registered in the entity database 31 or the distributed representation database 33 as a search query, the search unit 46 selects another first element that is most similar to the first element received as the search query. If one element or other relation information that is most similar to the relation information received as a search query is searched, and the corresponding second element is searched using the distributed expression of the searched first element and the distributed expression of the relation information. Good.

[3. Example of flow of processing executed by information providing apparatus]
Next, the flow of generation processing executed by the information providing apparatus 10 will be described with reference to FIG. FIG. 7 is a flowchart illustrating an example of a flow of generation processing executed by the information providing apparatus according to the embodiment. In addition, the information provision apparatus 10 can perform the process shown in FIG. 7 by arbitrary units at arbitrary timings.

  First, the information providing apparatus 10 extracts the first element and the relationship information from the correct triple, and sets the second element of the correct triple as correct data (step S101). Subsequently, the information providing apparatus 10 probabilistically determines whether or not the selection source of incorrect answer data is limited (step S102). If the information providing apparatus 10 determines not to limit (step S103: No), the information providing apparatus 10 randomly selects incorrect answer data from the second elements of triples other than the predetermined triple (step S104). On the other hand, when the information providing apparatus 10 determines to limit (step S103: Yes), it selects incorrect answer data from the second element belonging to the predetermined field according to the accuracy (step S105).

  Then, the information providing apparatus 10 generates the correct answer pair and the incorrect answer pair using the correct answer data, the incorrect answer data, the extracted first element and the relationship information (Step S106), and generates the correct answer pair and the incorrect answer pair. The distributed expression is learned by using (step S107), and the process is terminated.

[4. (Modification)
In the above, an example of generation processing and search processing by the information providing apparatus 10 has been described. However, the embodiment is not limited to this. Hereinafter, variations of generation processing and search processing executed by the information providing apparatus 10 will be described.

[4-1. (About entity types)
In the example described above, the information providing apparatus 10 has described the triple that associates a person with a person's name. However, the embodiment is not limited to this. That is, the information providing apparatus 10 may generate a distributed expression by executing the above-described generation process for a triple associated with an entity indicating an arbitrary event.

[4-2. Device configuration〕
Each of the databases 31 to 33 registered in the storage unit 30 may be held in an external storage server. Further, the information providing apparatus 10 may be realized by a front-end server that realizes a search process and a back-end server that realizes a generation process. In such a case, the search unit 46 shown in FIG. 2 is arranged in the front-end server, and the calculation unit 41, the extraction unit 42, the determination unit 43, the selection unit 44, and the generation unit 45 are arranged in the back-end server. Is done.

[4-3. Others]
In addition, among the processes described in the above embodiment, all or part of the processes described as being automatically performed can be performed manually, and conversely, the processes described as being performed manually. All or a part of the above can be automatically performed by a known method. In addition, the processing procedures, specific names, and information including various data and parameters shown in the document and drawings can be arbitrarily changed unless otherwise specified. For example, the various types of information illustrated in each drawing is not limited to the illustrated information.

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

  In addition, the above-described embodiments can be appropriately combined within a range in which processing contents do not contradict each other.

[4-4. program〕
Further, the information providing apparatus 10 according to the above-described embodiment is realized by a computer 1000 having a configuration as shown in FIG. FIG. 8 is a diagram illustrating an example of a hardware configuration. The computer 1000 is connected to an output device 1010 and an input device 1020, and an arithmetic device 1030, a primary storage device 1040, a secondary storage device 1050, an output IF (Interface) 1060, an input IF 1070, and a network IF 1080 are connected via a bus 1090. Have

  The arithmetic device 1030 operates based on a program stored in the primary storage device 1040 and the secondary storage device 1050, a program read from the input device 1020, and the like, and executes various processes. The primary storage device 1040 is a memory device such as a RAM that temporarily stores data used by the arithmetic device 1030 for various arithmetic operations. The secondary storage device 1050 is a storage device in which data used by the arithmetic device 1030 for various calculations and various databases are registered, such as ROM (Read Only Memory), HDD (Hard Disk Drive), flash memory, and the like. It is realized by.

  The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various types of information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), This is realized by a standard connector such as HDMI (registered trademark) (High Definition Multimedia Interface). The input IF 1070 is an interface for receiving information from various input devices 1020 such as a mouse, a keyboard, and a scanner, and is realized by, for example, a USB.

  The input device 1020 includes, for example, an optical recording medium such as a CD (Compact Disc), a DVD (Digital Versatile Disc), and a PD (Phase change rewritable disk), a magneto-optical recording medium such as an MO (Magneto-Optical disk), and a tape. It may be a device that reads information from a medium, a magnetic recording medium, a semiconductor memory, or the like. The input device 1020 may be an external storage medium such as a USB memory.

  The network IF 1080 receives data from other devices via the network N and sends the data to the arithmetic device 1030, and transmits data generated by the arithmetic device 1030 to other devices via the network N.

  The arithmetic device 1030 controls the output device 1010 and the input device 1020 via the output IF 1060 and the input IF 1070. For example, the arithmetic device 1030 loads a program from the input device 1020 or the secondary storage device 1050 onto the primary storage device 1040, and executes the loaded program.

  For example, when the computer 1000 functions as the information providing device 10, the arithmetic device 1030 of the computer 1000 implements the function of the control unit 40 by executing a program loaded on the primary storage device 1040.

[5. effect〕
As described above, the information providing apparatus 10 extracts the first element and the relationship information included in the correct answer triple. In addition, the information providing apparatus 10 determines probabilistically whether or not to limit the selection source of the second element that is incorrect answer data. Then, the information providing apparatus 10 selects at least a second element that is not included in the predetermined triple among the second elements included in any triple based on the determination result, and is related to the distributed representation of the extracted first element. Each distributed representation is generated so that the sum of the distributed representation of information and the distributed representation of the selected second element are not similar.

  For example, if the information providing apparatus 10 determines that the selection source of the second element that is the incorrect answer data is limited, the information providing apparatus 10 is the second element that is not included in the correct triple among the second elements included in any triple. The second element belonging to a predetermined field is selected. To give a more specific example, when the information providing apparatus 10 determines to limit the selection source of the second element serving as incorrect data, it is included in the correct triple among the second elements included in any triple. A second element belonging to the same or similar field as the second element and not included in the correct triple is selected as incorrect answer data. On the other hand, if the information providing apparatus 10 determines not to limit the selection source of the second element serving as the incorrect answer data, the second element that is not included in the correct triple among the second elements included in the arbitrary triple is rejected. Select as correct data.

  In this way, the information providing apparatus 10 learns the distributed expression using the incorrect answer data belonging to a predetermined field in a probabilistic manner. As a result, the information providing apparatus 10 can learn the distributed expression using the second element similar to the second element serving as the correct answer data as incorrect answer data. As a result, the information providing apparatus 10 can improve the accuracy of the generated distributed expression.

  The information providing apparatus 10 probabilistically determines whether or not to limit the selection source when the accuracy of the distributed representation satisfies a predetermined condition, and when the accuracy of the distributed representation does not satisfy the predetermined condition, It is determined that the selection source is not limited. For example, when the rate of increase in accuracy of the distributed representation does not change for a predetermined period, the information providing apparatus 10 determines probabilistically whether to limit the selection source. Further, for example, when the accuracy of the distributed representation exceeds a predetermined threshold, the information providing apparatus 10 determines probabilistically whether or not to limit the selection source. For this reason, in the distributed representation learning, the information providing apparatus 10 executes the same learning process at the initial stage and exceeds the accuracy estimated to be achievable by the conventional learning or the conventional learning. Then, when the accuracy does not increase, learning to improve the accuracy is performed by selecting incorrect answer data in a predetermined field in a probabilistic manner. As a result, the information providing apparatus 10 can realize more efficient learning.

  In addition, the information providing apparatus 10 calculates the accuracy indicating the similarity between the second elements included in the triple. When the information providing apparatus 10 determines that the selection source of the second element serving as the incorrect answer data is limited, the information providing apparatus 10 is a second element that is not included in the correct triple among the second elements included in any triple. Then, the second element whose accuracy with the second element included in the correct triple is included in a predetermined range is selected as incorrect answer data. For this reason, for example, the information providing apparatus 10 can set the second element similar to the second element serving as the correct answer data as incorrect answer data, so that the accuracy of the distributed expression to be generated can be improved.

  Moreover, the information provision apparatus 10 calculates the accuracy indicating the semantic similarity between the second elements. For this reason, since the information providing apparatus 10 can drop semantic similarity into the distributed expression space, for example, a semantic search process can be realized.

  In addition, the information providing apparatus 10 incorrectly determines, as the second element belonging to the predetermined field, the second element belonging to a field more similar to the field to which the second element included in the correct answer triple belongs as the accuracy of the distributed representation increases. Select as data. For this reason, as the learning progresses, the information providing apparatus 10 reflects fine similarities in the distributed representation, thereby improving the accuracy of the distributed representation and realizing efficient generation of the distributed representation.

  In addition, the information providing apparatus 10 selects, as incorrect answer data, a second element that belongs to a field in which the number of second elements to which it belongs is smaller than other fields. For this reason, the information providing apparatus 10 can prevent deterioration in accuracy of the distributed representation even when there is a bias in the field to which the second element belongs.

  Further, the information providing apparatus 10 extracts an element included as information corresponding to the subject in the triple as the first element, and extracts an element included as information corresponding to the predicate in the triple as the relationship information. Then, the information providing apparatus 10 selects an element included as information corresponding to the purpose in the triple as the second element. For this reason, the information providing apparatus 10 can generate a distributed representation from triples including information corresponding to S, P, and O, such as RDF.

  Further, the information providing apparatus 10 determines that the selection source of the second element is limited with a probability of 40% or less. For this reason, the information provision apparatus 10 can improve the precision of the distributed expression to generate.

  In addition, when the information providing apparatus 10 determines that the selection source of the second element is limited, the second element that is not included in the correct triple among the second elements included in the arbitrary triple including the relation information of the correct answer triple is selected. select. As a result, the information providing apparatus 10 can expand the selection range of the incorrect answer data, and therefore, it is possible to prevent the accuracy of the distributed representation from deteriorating even when the data amount is small.

  As described above, some of the embodiments of the present application have been described in detail with reference to the drawings. However, these are merely examples, and various modifications, including the aspects described in the disclosure section of the invention, based on the knowledge of those skilled in the art, It is possible to implement the present invention in other forms with improvements.

  Moreover, the above-mentioned “section (module, unit)” can be read as “means”, “circuit”, and the like. For example, the determination unit can be read as determination means or a determination circuit.

DESCRIPTION OF SYMBOLS 10 Information provision apparatus 20 Communication part 30 Storage part 31 Entity database 32 Relational information database 33 Distributed expression database 40 Control part 41 Calculation part 42 Extraction part 43 Determination part 44 Selection part 45 Generation part 46 Search part 100 User terminal

Claims (15)

An extraction unit for extracting the first element and the relationship information included in the predetermined triple;
A determination unit that probabilistically determines whether or not to limit the selection source of the second element;
A selection unit that selects at least a second element that is not included in the predetermined triple among second elements included in an arbitrary triple based on the determination result;
A generating unit that generates each distributed expression so that the sum of the extracted distributed expression of the first element and the distributed expression of the relationship information is not similar to the distributed expression of the selected second element. Generator to do. If the selection unit determines to limit the selection source of the second element, the selection unit is a second element that is not included in the predetermined triple among the second elements included in an arbitrary triple, and is in a predetermined field. The generation device according to claim 1, wherein the second element to which it belongs is selected. When it is determined that the selection source of the second element is limited, the selection unit belongs to the same or similar field as the second element included in the predetermined triple among the second elements included in the arbitrary triple. The generation device according to claim 2, wherein a second element that is not included in the predetermined triple is selected. When the selection unit determines that the selection source of the second element is not limited, the selection unit selects a second element that is not included in the predetermined triple among the second elements included in an arbitrary triple. The generation device according to any one of claims 1 to 3. When the accuracy of the distributed representation satisfies a predetermined condition, the determination unit probabilistically determines whether to limit the selection source, and when the accuracy of the distributed representation does not satisfy a predetermined condition, The generation apparatus according to claim 1, wherein the generation source is determined not to be limited. 6. The generation according to claim 5, wherein the determination unit probabilistically determines whether or not to limit the selection source when the rate of increase in accuracy of the distributed representation does not change for a predetermined period. apparatus. The generation apparatus according to claim 5 or 6, wherein the determination unit determines probabilistically whether or not to limit the selection source when the accuracy of the distributed representation exceeds a predetermined threshold. . A calculation unit that calculates the accuracy indicating the similarity between the second elements included in the triple;
When the selection unit determines to limit the selection source of the second element, the selection unit is a second element that is not included in the predetermined triple among the second elements included in an arbitrary triple, and the predetermined triple The generation device according to any one of claims 1 to 7, wherein a second element whose accuracy between the second element and the second element is included in a predetermined range is selected. The generation device according to claim 8, wherein the calculation unit calculates an accuracy indicating a semantic similarity between the second elements. The selection unit selects a second element belonging to a field more similar to a field to which the second element included in the predetermined triple belongs as the accuracy of the distributed representation increases. The production | generation apparatus as described in any one of these. The generation device according to any one of claims 1 to 10, wherein the selection unit selects a second element belonging to a field in which the number of second elements belonging to the field is smaller than other fields. The extraction unit extracts, as the first element, an element included as information corresponding to a subject in the triple, and as the relation information, extracts an element included as information corresponding to a predicate in the triple,
The generation device according to any one of claims 1 to 11, wherein the selection unit selects, as the second element, an element included in the triple as information corresponding to a purpose. If the selection unit determines to limit the selection source of the second element, the selection unit is not included in the predetermined triple among the second elements included in any triple including the relation information extracted by the extraction unit. Two elements are selected. The generation device according to any one of claims 1 to 12. A generation method executed by a generation device,
An extraction step of extracting the first element and the relationship information included in the predetermined triple;
A determination step of probabilistically determining whether or not to limit the selection source of the second element;
A selection step of selecting at least a second element not included in the predetermined triple among the second elements included in an arbitrary triple based on the determination result;
Generating the respective distributed expressions so that the sum of the extracted distributed expression of the first element and the distributed expression of the relationship information and the distributed expression of the selected second element are not similar to each other. How to generate. An extraction procedure for extracting the first element and the relationship information included in the predetermined triple;
A decision procedure for probabilistically determining whether to limit the selection source of the second element;
A selection procedure for selecting at least a second element not included in the predetermined triple among the second elements included in an arbitrary triple based on the determination result;
To cause the computer to execute a generation procedure for generating each distributed representation so that the sum of the extracted distributed representation of the first element and the distributed representation of the relationship information is not similar to the distributed representation of the selected second element Generator program.

Images