JP6956133B2 - model - Google Patents

Description

The present invention relates to a model.

Conventionally, a technique called a knowledge base has been known in which the concept of a person or an event is used as an entity and the relationship between the entities is structured. Further, in order to efficiently create such a knowledge database, there is known a technique of clustering entities and updating relationships between entities based on the result of clustering.

Japanese Patent No. 6088091

“On Emerging Entity Detection” Michael Farber, Achim Rettinger, Boulos El Asmar

However, with the above-mentioned technique, it may not be possible to efficiently add a new entity.

For example, a method of extracting a new entity (hereinafter referred to as "new entity") from the content posted on the Internet such as a news article can be considered. However, it is difficult to estimate what kind of new entity is included in which content.

The present application has been made in view of the above, and an object of the present application is to streamline the addition of new entities.

The model according to the present application is a first layer that is any of an input layer, an output layer, and a layer from the input layer to the output layer into which post information regarding a predetermined element is input, and belongs to a layer other than the output layer. It includes an element and a second element whose value is calculated based on the first element and the weight of the first element, and belongs to each layer other than the output layer with respect to the information input to the input layer. By performing an operation based on the first element and the weight of the first element with each element as the first element, a value indicating whether or not the posted information is posted information related to a new element is output. It is characterized by making the computer function so that it outputs from the layer.

According to one aspect of the embodiment, the addition of new entities can be streamlined.

FIG. 1 is a diagram showing an example of processing executed by the information providing device according to the embodiment. FIG. 2 is a diagram showing a configuration example of the information providing device according to the embodiment. FIG. 3 is a diagram showing an example of information registered in the entity database according to the embodiment. FIG. 4 is a diagram showing an example of information registered in the triple database according to the embodiment. FIG. 5 is a diagram showing an example of information registered in the search log database according to the embodiment. FIG. 6 is a diagram showing an example of information registered in the posted information database according to the embodiment. FIG. 7 is a diagram showing an example of information registered in the learning data database according to the embodiment. FIG. 8 is a diagram showing an example of information registered in the model database according to the embodiment. FIG. 9 is a diagram showing an example of a process in which the information providing device according to the embodiment selects an element candidate for which the learning data is to be created. FIG. 10 is a flowchart showing an example of a flow of processing executed by the information providing device according to the embodiment. FIG. 11 is a diagram showing an example of a hardware configuration.

Hereinafter, embodiments of the model according to the present (hereinafter referred to as "embodiment".) Will be described in detail with reference to the drawings. It should be understood that model according to the present application is limited by this embodiment. Further, in each of the following embodiments, the same parts are designated by the same reference numerals, and duplicate description is omitted.

[Embodiment]
[1. About the processing provided by the information providing device]
First, an example of the selection process executed by the information providing device, which is an example of the selection device, will be described with reference to FIG. FIG. 1 is a diagram showing an example of processing executed by the information providing device according to the embodiment. In the following description, as the process executed by the information providing device 10, the learning data for creating a model for detecting a new entity to be registered in the knowledge database in which knowledge is systematized and organized information is registered. An example of the selection process for selecting the source entity will be described.

[1-1. Overview of information providing equipment]
The information providing device 10 is an information processing device capable of communicating with the log server 100 and the explanatory content server 200 via a predetermined network N (see, for example, FIG. 2) such as the Internet, and is, for example, a server device or a server device. It is realized by a cloud system or the like. The information providing device 10 may be able to communicate with an arbitrary number of log servers 100 and explanatory content servers 200 via the network N. Further, the information providing device 10 has a function of providing various information registered in the knowledge database based on a request from a user terminal (not shown) used by the user, for example.

For example, the user terminal transmits a search query for searching the knowledge database to the information providing device 10. In such a case, the information providing device 10 searches the knowledge database for information corresponding to the search query, and provides the search result to the user terminal.

Here, the knowledge database searched by the information providing device 10 will be described. Various kinds of knowledge are registered in the knowledge database in a systematic and organized state. For example, in the knowledge database, an entity that is an element to be registered (hereinafter, may be described as "element") and information indicating a relationship between the entities (hereinafter, described as "relationship information"). And are registered. An entity is information corresponding 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, and so on. Further, the relationship information is information indicating the relationship between two entities. The elements in the knowledge database (that is, the entities registered in the knowledge database) of the information providing device 10 may correspond to any thing or event.

The log server 100 is an information processing device that holds various histories, and is realized by a server device, a cloud system, or the like. For example, the log server 100 holds a log of search queries entered by the user when performing various searches via the Internet. For example, the log server 100 holds a log of search queries in any search such as a web search, a route search, a transaction target search in an electronic shopping street, a map search, and a content search.

In addition, the log server 100 holds posted information which is various contents posted on the Internet. For example, the log server 100 holds news, blogs, SNS (Social Networking Service), etc. posted on the website. The log server 100 may be realized by a server device that distributes news, a server device that manages blogs, a server device that manages SNS, and the like.

The explanatory content server 200 is an information processing device that manages and distributes explanatory content that explains various elements such as people, objects, buildings, contents, and events, and is realized by a server device, a cloud system, or the like. For example, the explanatory content server 200 manages and distributes explanatory content, which is content that explains various objects that can be elements in a knowledge database, such as Wikipedia (registered trademark) and a dictionary service via the Internet.

[1-2. About selection process]
Here, consider the process of registering an element indicating a new matter in the knowledge database. In order to register a new element (hereinafter referred to as "new element") in such a knowledge database, not only the character string indicating the new element but also what kind of new element is different from other elements. Relationship information indicating whether or not they have a relationship is required. In order to efficiently add such new elements and related information to the knowledge database, a method of extracting new elements from contents posted on the Internet such as news articles can be considered. However, it is difficult to determine what kind of new element is included in which content.

Therefore, the information providing device 10 executes the following selection process. First, the information providing device 10 specifies the search history regarding the element candidate which is a candidate for a new element and the posted information regarding the element candidate. Subsequently, the information providing device 10 generates learning data for making the model learn the characteristics of the information about the new element among the element candidates based on the specified search history and the specified posted information. Select element candidates for.

For example, when a word is used as an element candidate, the search status of the word can be estimated from the search history including the word. In addition, the status of posting of the word can be estimated from the posts including the word. The status of such a search or the status of posting can be an index of whether or not the word indicates a new element. For example, if the word indicates a new element, the search or posting of the word may be sudden from one day. Therefore, the information providing device 10 selects an element candidate that is likely to be a new element based on the search status indicated by the search history and the posting status indicated by the posted information.

For example, the information providing device 10 selects an element candidate for generating learning data based on the number of searches indicated by the search history of a certain element candidate and the number of posted information indicated by the posting history of the element candidate. conduct. For example, the information providing device 10 selects an element candidate for generating learning data based on the content of the change in the number of searches and the content of the change in the number of posts. To give a more specific example, the information providing device 10 is an element for generating learning data for element candidates that are searched on a certain day and posted on that day. Select as a candidate.

Further, the information providing device 10 executes a learning process for learning a model based on the element candidates selected by the selection process. That is, the information providing device 10 causes the model to learn the characteristics of the information regarding the new element by using the information regarding the selected element candidate. For example, the information providing device 10 generates a set of selected element candidates, posted information about the selected element candidates, and relationship information indicating the relationship between the selected element candidates and other elements as learning data. .. Then, the information providing device 10 learns the model by using the learning data.

More specifically, the information providing device 10 uses the posted information regarding the selected element candidate to learn a determination model for determining whether or not the posted information is a post related to a new element. Further, the information providing device 10 learns an element extraction model that extracts a new element included in the posted information by using the selected element candidate and the posted information about the selected element candidate. Further, the information providing device 10 uses the posted information regarding the selected element candidate and the relationship between the selected element candidate and the other element to form a new element and another element included in the posted information. The relationship estimation model that extracts the relationship information indicating the relationship is trained. For example, the information providing device 10 is a plurality of models for extracting other elements having a predetermined relationship with the new attribute from the posted information, and extracts other elements having different relationships with each other. Learn multiple models.

Further, the information providing device 10 extracts new elements and relational information from the posted information by using the model learned by the learning process, and updates the knowledge database based on the extracted new elements and relational information. Execute the update process.

[1-3. About an example of the processing executed by the information providing device]
Hereinafter, an example of a selection process, a learning process, and an update process (hereinafter, may be collectively referred to as “each process”) executed by the information providing device 10 will be described with reference to FIG. First, the information providing device 10 executes the selection process. In such a selection process, the information providing device 10 first acquires the explanatory content created within a predetermined period from the explanatory content server 200 (step S1).

For example, when explanatory content is newly registered for a certain matter, the matter explained by the explanatory content is likely to be a new element. Therefore, the information providing device 10 acquires the explanatory content from the explanatory content server 200 in which the date and time elapsed since the registration is within a predetermined period (for example, several months).

Subsequently, the information providing device 10 obtains, from the explanatory content, an element that is a candidate for a new element (hereinafter, referred to as "element candidate") and relational information indicating the relationship between the element candidate and another element. Extract (step S2). For example, in the example shown in FIG. 1, the information providing device 10 is the explanatory content "explanatory content # 1" in which the explanatory content ID (Identifier) for identifying the explanatory content is "content ID # 1", and the registration date and time is Acquire the explanation content that is "date and time # 1". In such a case, the information providing device 10 extracts a character string presumed to indicate the subject of the explanatory content as an element candidate, and also extracts the relational information indicating the relationship between the element candidate and another element as the explanatory content. Estimate from the contents of. That is, when the explanatory content is newly registered, the information providing device 10 acquires the element candidate from the explanatory content.

Here, the information providing device 10 may extract element candidates and relational information by using various sentence analysis techniques such as morphological analysis and meaning estimation. Also, for example, in the explanatory content, information indicating what is the subject matter of the explanatory content and what is the relationship between the subject matter and other matters (for example, called an infobox). Information) may be included. When such an infobox is included in the explanatory content, the information providing device 10 may extract element candidates and relational information from the infobox.

Further, for example, when the infobox is added within a predetermined period (for example, one month) after the explanatory content is registered, the information providing device 10 displays the related information registered in the infobox. It may be adopted as the relational information of the element candidate. In addition, the infobox may include classification information indicating classification such as a category or class of element candidates. When such classification information is included in the infobox, the information providing device 10 may extract the classification information of the element candidates from the infobox.

As a result of such processing, the information providing device 10 extracts a set of the element candidate and the relational information associated with the element candidate as candidate data. For example, the information providing device 10 includes element candidate # 1, relationship information # 1-1 indicating the relationship between element candidate # 1 and another first element, element candidate # 1 and others in the explanatory content # 1. When the relationship information # 1-2 indicating the relationship with the second element of the above is included, the element candidate # 1 and the relationship information group # 1 (relationship information # 1-1, and the relationship information # 1-1) are included from the explanation content # 1. Candidate data # 1 associated with the relationship information # 1-2) is generated. The information providing device 10 may include the classification information associated with the element candidate in the candidate data # 1.

Here, in addition to the explanation content related to the new element, the matters for which the explanation content is newly registered include explanations related to new matters such as explanation content obtained by translating the explanation content of another language that already exists into Japanese. Content other than content may be included. Therefore, the information providing device 10 selects candidate data of an element candidate having a high accuracy of a new element from the candidate data. More specifically, the information providing device 10 acquires the search history of the element candidate and the posted information about the element candidate (step S3). Then, the information providing device 10 selects an element candidate to be a new element based on the number of the specified search histories and the number of posted information.

For example, the information providing device 10 identifies an element candidate in which the amount of increase in the search history and the posting history at a predetermined date and time satisfies a predetermined condition, and generates learning data based on the specified element candidate (step S4). More specifically, in the information providing device 10, among the plurality of element candidates extracted from the explanatory content, the amount of increase in the number of search histories exceeds a predetermined threshold value at a predetermined date and time, and the number of posted information is increased. Element candidates whose increase amount exceeds a predetermined threshold are selected.

For example, consider a mode in which a new movie is a new element. When such a new movie is announced, the number of searches using the title of the movie as a search query is expected to increase sharply after the announcement compared to before the announcement. In addition, when such a new movie is announced, it is considered that the number of news and SNS including the title of the new movie will increase sharply after the announcement as compared with before the announcement. For this reason, it is thought that the number of searches and the number of posted information using new elements as search queries will increase sharply after a certain date and time. In addition, when such a new movie is announced, the explanatory content will be generated after a certain period of time has passed since the announcement. Furthermore, if the novel is made into a movie, it is considered that the element candidates are included in the search query and the posted information before the movie announcement date and time. Furthermore, if the increase in search queries that include the movie title and the increase in post information that includes the movie title are not linked, the string is the same as the string that is the title of the movie, but different. It is probable that a character string indicating the matter has been searched or posted.

Therefore, the information providing device 10 sets the date and time when the explanatory content from which the element candidates are extracted is registered as the reference date and time, and acquires the search history and the posted information for a predetermined period before the reference date and time. For example, the information providing device 10 is a search query input in the past one year of the reference date and time, and is a search query including element candidates and post information posted in the past one year of the reference date and time, and is an element candidate. Get post information including. Therefore, the information providing device 10 does not have the search history and the posted information before the predetermined date and time among the element candidates extracted from the explanatory content, and the number of the search histories sets a predetermined threshold value at the predetermined date and time. Select element candidates that exceed and the number of posted information exceeds a predetermined threshold.

For example, in the example shown in FIG. 1, the transition of the number of search queries including the element candidate # 1 is plotted with a dotted line for each date and time, and the transition of the number of posted information including the element candidate # 1 is plotted with a straight line for each date. Plotted. In the example shown in FIG. 1, as shown in FIG. 1 (A), the number of search queries and posted information becomes "0" until a certain date and time, and as shown in FIG. 1 (B), the search query is performed at a certain date and time. And the number of posted information is skyrocketing. In this way, if the number of search queries including element candidate # 1 and the number of posted information change from "0" to "1" or more on the same date and time, element candidate # 1 may be a new element. Highly sexual. In such a case, the information providing device 10 selects the element candidate # 1 as the learning target, and sets the candidate data # 1 of the element candidate # 1 as the learning data # 1.

Further, the information providing device 10 extracts the posted information posted at the date and time when the number of posted information suddenly increases from the posted information including the element candidates as learning data. For example, in the example shown in FIG. 1, the information providing device 10 posts information group posted at a date and time when the number of posted information changes from "0" to "1" or more among the posted information including the element candidate # 1. # 1 is extracted, and the extracted post information group # 1 is included in the learning data # 1. As a result of such a selection process, the information providing device 10 includes an element candidate that is likely to be a new element, relational information indicating the relationship between the element candidate and another candidate, and a post including the element candidate. Information and can be acquired as learning data.

Here, in the above-mentioned selection process, if the number of search queries and the number of posted information change from "0" to "1" or more on the same date and time, the posted information posted on that date and time is learned data. Extracted as. In this way, the posted information posted on the date and time when the number of search queries and the number of posted information changed from "0" to "1" or more is considered to be posted information that explains the element candidate as a new element. Be done.

It is considered that the posted information for explaining the element candidate as a new element includes a sentence having a feature indicating that the element candidate is a new element. Therefore, when learning the characteristics of such posted information, it is considered possible to create a model for estimating whether or not the posted information contains new elements.

Further, it is considered that the post information that explains the element candidate as a new element includes the character string of the element candidate that is the new element. Therefore, when learning the characteristics of such posted information, it is considered possible to create a model for extracting a character string presumed to be a new element from the posted information.

In addition, the post information that explains the element candidate as a new element includes a character string that explains the element candidate that is the new element, that is, a character string that indicates the relationship between the element candidate that is the new element and other elements. Is considered to be included. Therefore, when learning the characteristics of such posted information, it is considered possible to create a model for estimating the relationship information indicating the relationship between the new element and other elements from the posted information.

Therefore, the information providing device 10 uses the learning data generated by the selection process to determine whether or not the posted information includes new elements, an extraction model for extracting new elements from the posted information, and an extraction model. Learn the relationship estimation model that estimates the relationship information indicating the relationship between the new element and other elements from the posted information (step S5). For example, the information providing device 10 uses the posted information group # 1 included in the learning data # 1 as the correct answer data, and trains the model to learn the features of the correct answer data so that the posted information is the posted information indicating a new element. Generate a judgment model to determine whether or not. To give a more specific example, when the information providing device 10 inputs the posted information group # 1 included in the learning data # 1, the information providing device 10 provides information indicating that the input posted information includes a new element. When the post information that is output and includes the element candidate # 1 and is not included in the post information group # 1 is input, the information indicating that the input post information does not include the new element is output. The judgment model is trained so as to output.

Further, the information providing device 10 uses the element candidate # 1 and the posted information group # 1 as correct answer data, and trains the model to learn the features of the correct answer data to generate an extraction model that extracts new elements from the posted information. do. To give a more specific example, the information providing device 10 learns the extraction model so that when each posted information included in the posted information group # 1 is input, the information indicating the element candidate # 1 is output. conduct. That is, the information providing device 10 learns the extraction model so that the output data becomes the element candidate # 1 when the posted information group # 1 is used as the input data.

Further, the information providing device 10 estimates the relational information of the new element from the posted information by using the posted information group # 1 and the relational information group # 1 as correct answer data and letting the model learn the features of the correct answer data. Generate a relationship estimation model. To give a more specific example, the information providing device 10 estimates the relationship so as to estimate the relationship information included in the relationship information group # 1 when each posted information included in the posted information group # 1 is input. Train the model.

For example, the information providing device 10 is a plurality of models for extracting other elements having a predetermined relationship with a new element from posted information as a relationship estimation model, and each has a different relationship. Learn multiple models to extract the elements of. For example, the information providing device 10 uses an arbitrary sentence analysis technique to extract other elements having a predetermined relationship with the element candidates from each posted information included in the learning data # 1.

For example, when the posted information is a description of a new movie, the information providing device 10 extracts a character string of a performer, a character string of a director, a character string of a distribution company, and the like. Then, for example, the information providing device 10 learns the first relationship estimation model so as to extract the character string of the performer when the posted information is input, and when the posted information is input, the director's character string is input. The second relationship estimation model is learned so as to be extracted, and the third relationship estimation model is learned so as to extract the character string of the distribution company when the posted information is input.

Here, the information providing device 10 may learn the relationship estimation model using an arbitrary prediction model. For example, the information providing device 10 may learn the relationship estimation model based on the entity registered in the knowledge base and the relationship information. To give a more specific example, the information providing device 10 utilizes a characteristic that an element included in the posted information and another element having a predetermined relationship with the element candidate is an entity in the knowledge database. Identify from the knowledge database whether or not such "other factors" are humans, and if so, what their gender and occupation are. Then, the information providing device 10 may make the relationship estimation model learn the relationship between the specified relationship information, the element candidate, and the element. That is, the information providing device 10 may learn a relationship estimation model that predicts the relationship between each element included in the posted information and the element candidate.

Then, the information providing device 10 executes the update process using each model. That is, the information providing device 10 identifies the posted information including the new element, extracts the new element and the relational information from the specific information, and updates the knowledge database by using the extracted new element and the relational information. (Step S6). For example, when new post information # N1 and post information # N2 exist, the information providing device 10 uses a determination model to post information in which each post information # N1 and # N2 includes a new element (hereinafter, "" It may be described as "new post".) Judge whether or not it is. Then, when it is determined that the posted information # N1 is a new post, the information providing device 10 extracts new elements and related information from the posted information # N1 by using the extraction model and the relationship estimation model. do.

More specifically, the information providing device 10 extracts a new element from the posted information # N1 and extracts the new element and another element having a predetermined relationship with each other. Then, the information providing device 10 updates the knowledge database by using the extracted new elements and the relational information. For example, the information providing device 10 registers the new element in the knowledge database, and stores the new element, the extracted other element, and the relationship information indicating the relationship between the new element and the other element. Set the associated triple. For example, the information providing device 10 extracts the director's character string from the posted information # N1 when the relationship estimation model trained to extract the director's character string when the posted information is input. Information called a triple is registered in which the new element extracted by the model, the character string extracted by the relationship estimation model (that is, other elements), and the relationship information "supervisor" indicating the relationship between the elements are associated with each other.

That is, in the knowledge database updated by the information providing device 10, various kinds of knowledge are systematized by registering information called a triple consisting of two entities and relationship information indicating the relationship between the two entities. .. In other words, in the knowledge database, various knowledge is systematically represented by a triple consisting of an entity that is the subject (S), an entity that is the object (O), and relational information that is the predicate (P). ing.

In this way, the information providing device 10 specifies the search history regarding the element candidate which is a candidate for a new element and the posted information regarding the element candidate. Then, the information providing device 10 is learning data for making the model learn the characteristics of the information about the new element among the element candidates based on the number of the specified search histories and the number of the specified posted information. Select element candidates to generate. Therefore, the information providing device 10 can streamline the addition of new elements using the model.

[1-4. About the model]
Here, the information providing device 10 may learn an arbitrary type of determination model, extraction model, and relationship estimation model (hereinafter, may be referred to as “each model”). For example, the information providing device 10 may include a regression model, Bag of words, SVM (Support Vector Machine), DNN (Deep Neural Network), CRF (Conditional Random Fields), LSTM (Long short-term memory), or any other model. It can be adopted.

For example, the determination model in which the training is performed by the above-mentioned learning process is an input layer in which post information relating to a predetermined element is input, an output layer, or any layer from the input layer to the output layer, and is an output layer. The information input to the input layer includes the first element belonging to a layer other than the first element and the second element whose value is calculated based on the weights of the first element and the first element, and the information input to the input layer is other than the output layer. By performing an operation based on the first element and the weight of the first element with each element belonging to each layer as the first element, a value indicating whether or not the posted information is the posted information related to the new element is obtained from the output layer. It is a model for making the computer function to output.

Further, for example, the extraction model trained by the above-mentioned learning process is an input layer into which post information regarding a new element is input, an output layer, or any layer from the input layer to the output layer. An output layer for information input to the input layer, including a first element belonging to a layer other than the output layer and a second element whose value is calculated based on the weights of the first element and the first element. By performing an operation based on the first element and the weight of the first element with each element belonging to each layer other than the first element as the first element, a value indicating information indicating a new element among the information included in the posted information is output. It is a model for making the computer function so that it outputs from the layer.

Further, for example, the relationship estimation model trained by the above-mentioned learning process has an input layer in which post information about a new element is input, an output layer, and any layer from the input layer to the output layer. With respect to the information input to the input layer, including the first element belonging to a layer other than the output layer and the second element whose value is calculated based on the weights of the first element and the first element. By performing an operation based on the weights of the first element and the first element with each element belonging to each layer other than the output layer as the first element, a new element and other elements among the information included in the posted information can be obtained. It is a model for making a computer function so that a value indicating the relationship between them is output from the output layer.

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

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

The information providing device 10 acquires new elements and relational information by using each model having an arbitrary structure such as the regression model and the neural network described above. Specifically, the judgment model sets various parameters (for example, connection coefficient) so as to output a score indicating whether or not the posted information includes a new element when the posted information is input. Set. Further, in the extraction model, various parameters are set so as to output information indicating information such as a character string that is likely to be a new element from the posted information when the posted information is input. Further, in the relationship estimation model, various parameters are set so as to output information indicating other elements having a predetermined relationship with the new element indicated by the posted information when the posted information is input. ..

Each model according to the embodiment may be a model generated based on a result obtained by repeating input / output of data to a predetermined model. Further, when the information providing device 10 performs learning processing using GAN (Generative Adversarial Networks), each model may be a model forming a part of GAN.

For example, the information providing device 10 searches for element candidates selected based on the number of search histories related to element candidates and the number of posted information related to element candidates among the element candidates that are candidates for new elements. Each model is generated by performing the above-mentioned learning using the learning data including the history and the information indicating the relationship between the element candidate and the other element. The information providing device 10 may generate each using any learning algorithm. For example, the information providing device 10 generates a learning model by using a learning algorithm such as a neural network, a support vector machine, clustering, and reinforcement learning. As an example, when the information providing device 10 uses a neural network to generate each model, each model includes an input layer containing one or more neurons, an intermediate layer containing one or more neurons, and one or more neurons. It has an output layer.

The learning data is data for operating the information providing device 10 as the various models described above. That is, the learning data includes the element candidates selected based on the number of search histories related to the element candidates and the number of posted information related to the element candidates among the element candidates that are candidates for new elements, and the search history related to the element candidates. , Information indicating the relationship between the element candidate and other elements, and the data for making the computer function as the various models described above.

[1-5. About learning data]
In the above description, the information providing device 10 uses the posted information posted on the date and time when the number of search queries and posted information changes from "0" to "1" or more as learning data. However, the embodiment is not limited to this. For example, the information providing device 10 may use the posted information posted during the day when the number of search queries and posted information is changed from "0" to "1" or more as learning data, and the number of search queries and posted information is large. Posting information posted between the day when the number changes from "0" to "1" or more and the time when a predetermined date and time elapses may be used as learning data.

Further, for example, if the information providing device 10 selects an element candidate to be learning data based on the search status indicated by the search history and the posting status indicated by the posting history, the information providing device 10 may select an element candidate as learning data based on an arbitrary situation. Element candidates may be selected. For example, the information providing device 10 has a period in which the number of search histories and posting histories is equal to or less than a predetermined threshold value, and the number of search histories and posting histories is equal to or greater than a predetermined threshold value after a certain day. If this happens, the corresponding element candidate may be selected as training data. Further, the information providing device 10 may select an element candidate based on an arbitrary situation such as a search history or a statistical situation of a search or a post indicated by a post history.

Further, the information providing device 10 may generate learning data which is not only a positive example but also a negative example. For example, the information providing device 10 is the posted information about a certain element candidate, and the posted information posted on the day when the number of search queries and posted information changes from "0" to "1" or more is taken as a positive example. As a negative example, the posted information about the element candidate, which is posted after a predetermined date and time has elapsed from the day when the number of search queries and posted information changes from "0" to "1" or more. Then, the information providing device 10 may learn the determination model by using the positive example and the negative example.

The information providing device 10 may acquire the posted information registered in an arbitrary period when creating the information providing device 10. For example, the information providing device 10 may acquire post information and search queries for the past one year, and may determine whether or not the element candidate is a new element by using the acquired post information and search query. ..

In addition, it is predicted that the period from the occurrence of the new element to the generation of the explanatory content of the new element and the generation of the infobox will be within a predetermined range (for example, 40 days on average). .. Therefore, the information providing device 10 may acquire the posted information 40 days before the date and time when the explanatory content is generated or the infobox is generated, and the acquired posted information may be used as learning data.

[2. Configuration of information providing device]
Hereinafter, an example of the functional configuration of the information providing device 10 described above will be described. FIG. 2 is a diagram showing a configuration example of the information providing device according to the embodiment. As shown in FIG. 2, the information providing device 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) or the like. Then, the communication unit 20 is connected to the network N by wire or wirelessly, and transmits / receives information to / from the log server 100 and the explanation content server 200.

The storage unit 30 is realized by, for example, a semiconductor memory element such as a RAM (Random Access Memory) or a flash memory (Flash Memory), or a storage device such as a hard disk or an optical disk. Further, the storage unit 30 may be collectively referred to as an entity database 31, a triple database 32, a search log database 33, a post information database 34, a training data database 35, and a model database 36 (hereinafter, “each database 31 to 36”). .) Is memorized.

Hereinafter, an example of information registered in each database 31 to 36 will be described with reference to FIGS. 3 to 8. Information about an entity is registered in the entity database 31. For example, FIG. 3 is a diagram showing 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", "entity type", "node ID", "node type", and "data" is registered in the entity database 31.

Here, the "entity ID" is an identifier of an entity. Further, the "entity type" is information indicating the type of the entity indicated by the associated "entity ID". For example, whether the entity is an entity indicating a "person" or an entity indicating an "occupation". Information such as is shown. The "node ID" is an identifier of a node associated with the entity indicated by the associated "entity ID". The "node type" is information indicating the type of the node indicated by the associated "node ID", and is information such as whether the node indicates a name, a photograph, an occupation, or the like. Further, the "data" is the data of the node indicated by the associated "node ID".

For example, in the example shown in FIG. 3, the entity ID "E11", the entity type "person", the node ID "I111", the node type "name", and the data "name # 1" are registered in association with each other. In such information, the entity indicated by the entity ID "E11" (that is, the entity E11) is an entity indicating a "person", and the node ID "I111" is registered as a node indicating the "name" of the person indicated by the entity. Indicates that the name is "Name # 1". In the example shown in FIG. 1, conceptual values such as "name # 1" and "photo # 1" are described, but in reality, the entity database 31 contains a person corresponding to the associated entity. Various information indicating the name, photo, date of birth, etc. will be registered.

Information indicating the triple is registered in the triple database 32. For example, FIG. 4 is a diagram showing an example of information registered in the triple database according to the embodiment. In the example shown in FIG. 4, information having items such as "triple ID", "relationship information ID", "type", "first element", and "second element" is registered in the triple database 32.

Here, the "triple ID" is an identifier that identifies the triple. Further, the "relationship information ID" is an identifier that identifies the relational information included in the triple. Further, the "type" is information indicating the relationship between the entities included in the triple. Further, the "first element" and the "second element" are entity IDs of entities included in the triple indicated by the associated "triple ID".

For example, in the example shown in FIG. 4, the triple ID "triple # 1", the relationship information ID "C1", the type "occupation", the first element "E11", and the second element "E21" are registered in association with each other. There is. Such information is registered as a triple indicated by the triple ID "triple # 1" in which the entity E11, the entity E12, and the relationship information C1 are associated with each other, indicating that the entity E21 is the occupation of the entity E11.

The history of search queries, that is, the search history is registered in the search log database 33. For example, FIG. 5 is a diagram showing an example of information registered in the search log database according to the embodiment. As shown in FIG. 5, in the search log database 33, an identifier that identifies the search log, a "search log ID", a "search query" that indicates the entered search query, and a "search query" that indicates the date and time when the search was performed are displayed. Information having an item such as "search date and time" is registered.

For example, in the example shown in FIG. 5, the search log database 33 is associated with information such as the search log ID "search log # 1", the search query "search query # 1", and the search date and time "search date and time # 1". It is registered. Such information indicates that the search query of the search log indicated by the search log ID "search log # 1" is "search query # 1" and the search date and time is "search date and time # 1". In the example shown in FIG. 5, conceptual values such as "search log # 1", "search query # 1", and "search date and time # 1" are described, but in reality, the search log database 33 contains Numerical values and character strings that identify the search log, character strings entered as search queries, numerical values that indicate the search date and time, etc. are registered.

Post information is registered in the post information database 34. For example, FIG. 6 is a diagram showing an example of information registered in the posted information database according to the embodiment. As shown in FIG. 6, in the post information database 34, the "post log ID" which is an identifier for identifying the post information, the "post information" indicating the content of the post information, and the "post date and time when the post information was posted" are shown. Information with items such as "posting date and time" is registered.

For example, in the example shown in FIG. 6, the post information database 34 is associated with information such as the post log ID “post log # 1”, the post information “post information # 1”, and the post date and time “post date and time # 1”. It is registered. Such information indicates that the posting information indicated by the posting log ID "posting log # 1" is "posting information # 1" and the posting date and time is "posting date and time # 1". In the example shown in FIG. 6, conceptual values such as "post log # 1", "post information # 1", and "post date / time # 1" are described, but in reality, the post information database 34 contains Numerical values and character strings that identify the posted information, character strings entered as the posted information, numerical values that indicate the posting date and time, etc. are registered.

The learning data is registered in the learning data database 35. For example, FIG. 7 is a diagram showing an example of information registered in the learning data database according to the embodiment. As shown in FIG. 7, in the training data database 35, the “learning data ID” which is an identifier for identifying the training data, the “element candidate” which is the element candidate included in the training data, and the relational information included in the training data Information having items such as a certain "relationship information" and "posted information" which is posted information included in the learning data is registered.

For example, in the example shown in FIG. 7, the learning data database 35 has the learning data ID “learning data # 1”, the element candidate “element candidate # 1”, the relational information “relationship information group # 1”, and the posting information “posting information”. Group # 1 ”is registered in association with each other. Such information is registered in association with "element candidate # 1", "relationship information group # 1", and "posted information group # 1" as learning data indicated by the learning data ID "learning data # 1". Indicates that. Further, in such information, as "relationship information group # 1", another element having a predetermined relationship with "element candidate # 1" is "target element # 1-1", and the relationship is "relationship # 1". It indicates that the fact that it is "1-1" is registered. Further, such information indicates that "posted information # 1-1" and "posted information # 1-2" are registered as "posted information group # 1".

In the example shown in FIG. 7, concepts such as "learning data # 1", "element candidate # 1", "relationship information # 1-1", "relationship # 1-1", and "posted information # 1-1" However, in reality, a numerical value or a character string for identifying the training data, an element candidate, a relationship information, a relationship, a character string or the like as posting information or the like is registered in the training data database 35.

Data of each model is registered in the model database 36. For example, FIG. 8 is a diagram showing an example of information registered in the model database according to the embodiment. As shown in FIG. 8, the "model type" indicating the model type and the "model data" which is the model information are registered in the model database 36 in association with each other.

For example, in the example shown in FIG. 8, the model type “determination model” and the model data “model data # 1” are registered in association with each other in the model database 36. Such information indicates that the various parameters of the model that is the "determination model" are "model data # 1". In the example shown in FIG. 8, conceptual values such as "model data # 1" are described, but in reality, various information necessary for constructing a model such as connection relationships between nodes and connection coefficients. Is registered.

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

As shown in FIG. 2, the control unit 40 includes an acquisition unit 41, a specific unit 42, a selection unit 43, an extraction unit 44, a learning unit 45, and an update unit 46. When the explanatory content for explaining the element candidate is registered, the acquisition unit 41 acquires the element candidate from the explanatory content. For example, the acquisition unit 41 refers to the explanatory content server 200 at predetermined time intervals, and acquires the newly registered explanatory content within a predetermined period. In such a case, the acquisition unit 41 uses various character analysis techniques to extract text indicating the subject matter of the explanatory content, and uses the extracted text as an element candidate.

The identification unit 42 identifies the search history of the element candidate, which is a candidate for a new element, and the posted information about the element candidate. More specifically, the specific unit 42 specifies the search history of the element candidate acquired from the explanatory content by the acquisition unit 41 and the posted information regarding the element candidate. For example, the specific unit 42 identifies the posted information of the element candidate, which is posted within a predetermined period before the explanatory content is registered.

For example, the specific unit 42 acquires various search histories and posting information from the log server 100 at predetermined time intervals. Then, the specific unit 42 registers the acquired search history in the search log database 33, and registers the acquired post information in the post information database 34. Further, when the element candidate is acquired by the acquisition unit 41, the specific unit 42 refers to the search log database 33 and specifies a search log including the element candidate as a search query. In addition, the specifying unit 42 specifies the posted information including the acquired element candidates from the posted information database 34.

Here, the specific unit 42 may consider the search date and time and the posting date and time when specifying the search history and the posting information. For example, the specific unit 42 specifies the date and time when the acquired element candidate explanation content is registered, and searches the search history and posted information within a predetermined period prior to the specified date and time. History and posted information may be specified. For example, the specific unit 42 may specify the posted information for 40 days before the registration date of the explanatory content.

The selection unit 43 is an element candidate for generating learning data for making a model learn the characteristics of information about a new element among the element candidates based on the specified search history and the specified post information. Select. For example, in the selection unit 43, among a plurality of element candidates, an element in which the amount of increase in the number of search histories exceeds a predetermined threshold value and the amount of increase in the number of posted information exceeds a predetermined threshold value at a predetermined date and time. Select a candidate. For example, in the selection unit 43, among the plurality of element candidates, the search history and posting information before the predetermined date and time do not exist, the number of search histories exceeds the predetermined threshold value at the predetermined date and time, and Select element candidates for which the number of posted information exceeds a predetermined threshold.

For example, FIG. 9 is a diagram showing an example of a process in which the information providing device according to the embodiment selects an element candidate for which the learning data is to be created. In FIG. 9 (A), the transition of the number of search queries including a certain element candidate # 1 is plotted with a dotted line for each date and time, and the transition of the number of posted information including the element candidate # 1 is plotted for each date. Was plotted as a straight line. Further, in FIG. 9B, the transition of the number of search queries including a certain element candidate # 2 is plotted with a dotted line for each date and time, and the transition of the number of posted information including the element candidate # 2 is plotted for each date. Was plotted as a straight line.

For example, at the date and time shown in FIG. 9 (a), the search history and posted information including the element candidate # 1 suddenly occur, and the period shown in FIG. 9 (b) before FIG. 9 (a). In, there is no search history or posted information including element candidate # 1. When the number of search histories and the number of posted information follow such a transition, the element candidate # 1 is likely to be a new element. Therefore, in the selection unit 43, the element candidate # 1 in which the number of search histories exceeds a predetermined threshold value and the number of posted information exceeds a predetermined threshold value at a predetermined date and time from the state where the search history and the posted information do not exist. Is selected as the target for creating training data.

On the other hand, as shown in FIG. 9B, the number of search histories and posted information including the element candidate # 2 exists in a certain amount while repeatedly increasing and decreasing. Here, at the date and time shown in FIG. 9 (c), the search history and posted information including the element candidate # 1 are rapidly increasing, but as shown in FIG. 9 (d), the element candidate # 2 has been used in the past. There is a date and time when the number of search history and posted information including is rapidly increasing. If the number of search histories and the number of posted information follow such a transition, it is highly possible that element candidate # 2 is not a new element. To give a specific example, when element candidate # 2 is the title of a novel, element candidate # 2 becomes a topic at the date and time shown in FIG. 9 (d), and element candidate # 2 becomes a topic at the date and time shown in FIG. 9 (c). It is conceivable that the # 2 novel was made into a movie. When such element candidate # 2 is used as learning data, it is not possible to select post information that truly includes a new element, and as a result, the accuracy of the training data is lowered, and as a result, the extraction accuracy of the new element is lowered. There is a fear. Therefore, the information providing device 10 excludes the element candidate # 2 from the learning data creation target.

Returning to FIG. 2, the explanation will be continued. The extraction unit 44 extracts the posted information posted on the date and time when the number of posted information and search information suddenly increases from the posted information related to the element candidates selected by the selection unit 43 as learning data. In addition, the extraction unit 44 extracts the relationship between the element candidate and other elements as learning data from the explanatory content explaining the element candidate. For example, when the selection unit 43 selects the element candidate for which the learning data is to be created, the extraction unit 44 is the post information including the selected element candidate, and the number of the post information and the search information is from "0" to "0". Extract the posted information posted on the day when it increased to 1 "or more. Then, the extraction unit 44 associates the element candidates with the extracted post information and registers them in the learning data database 35.

Further, when the infobox is registered in the explanation content of the element candidate, the extraction unit 44 extracts the relationship information indicating the relationship between the element candidate and the other element from the infobox. The extraction unit 44 identifies the relationship information indicating the relationship between the element candidate and the other element from the character analysis technique, for example, the link relationship to the other explanatory content set in the explanatory content, and the like. You may. Then, the extraction unit 44 associates the extracted relationship information with the element candidates and registers them in the learning data database 35.

The learning unit 45 causes the model to learn the characteristics of the information about the new element by using the information about the selected element candidate. That is, the learning unit 45 learns each model using the learning data regarding the element candidates selected by the selection unit 43. For example, the learning unit 45 causes each model to learn various features of the learning data by using the selected element candidate, the posted information about the element candidate, and the relationship between the element candidate and other elements as learning data.

For example, the learning unit 45 uses the posted information of the learning data to learn a determination model for determining whether or not the posted information is a post related to a new element. For example, when the learning unit 45 inputs the posted information included in the learning data, the learning unit 45 outputs information indicating that the input posted information includes a new element, and when other posted information is input, the learning unit 45 outputs the information. The judgment model is trained so as to output information indicating that the input posted information does not include new elements.

Further, for example, the learning unit 45 learns an element extraction model that extracts a new element included in the posted information by using the element candidate of the learning data and the posted information about the element candidate. For example, the learning unit 45 learns the extraction model so as to output information indicating element candidates of the learning data when the posted information of a certain learning data is input.

Further, for example, the learning unit 45 uses the posted information of the learning data and the relationship between the element candidates of the learning data and other elements to relate the new element included in the posted information to the other elements. Learn the relationship estimation model that extracts sex. More specifically, the learning unit 45 is a plurality of models for extracting other elements having a predetermined relationship with a new element from the posted information, and the other elements having different relationships with each other. Learn multiple models to extract.

For example, the learning unit 45 prepares a plurality of models associated with different relationships as the relationship estimation model. Further, for example, when learning a model associated with the first relationship, the learning unit 45 refers to the learning data database 35 and has a first relationship between the element candidate and the element candidate. Identify the elements of. Then, when the posted information is input to the model associated with the first relationship, the learning unit 45 learns the model so as to output information indicating the specified other elements. By executing such processing for each relation estimation model, the learning unit 45 extracts another element having a predetermined relationship with the element candidate from the posted information, that is, the element candidate and the other element. You can learn a relationship estimation model for estimating the relationship between.

The update unit 46 updates the knowledge database using each model trained by the learning unit 45. For example, the update unit 46 acquires post information from the log server 100 for which a predetermined time has not passed since the post was posted. Then, the update unit 46 determines whether or not each posted information includes a new element by using the determination model.

Further, when the determination model determines that a certain posted information includes a new element, the update unit 46 extracts the new element from the posted information by using the extraction model. That is, the update unit 46 uses the extraction model to extract a character string that is likely to indicate a new element from the posted information determined to include the new element. In addition, the update unit 46 uses the relationship estimation model to extract other elements presumed to have a predetermined relationship with the new element from the posted information determined to include the new element.

Then, the update unit 46 updates the knowledge database by using the extracted new element and the relationship between the new element and other elements. For example, the update unit 46 registers an entity indicating a new element in the entity database 31. Further, the update unit 46 identifies an entity of another element having a predetermined relationship with the new element from the entity database 31. Then, as a triple, the update unit 46 triples a set of the entity of the new element, the entity of the specified other element, and the relationship (that is, "type") corresponding to the relationship estimation model extracted from the element. To be registered in the triple database 32.

[3. An example of the flow of processing executed by the information providing device]
Subsequently, the flow of processing executed by the information providing device 10 will be described with reference to FIG. FIG. 10 is a flowchart showing an example of a flow of processing executed by the information providing device according to the embodiment.

First, the information providing device 10 acquires the explanatory content created within a predetermined period (step S101). Subsequently, the information providing device 10 extracts element candidates from the explanatory content (step S102). Further, the information providing device 10 acquires the search history and the posting history of the extracted element candidates (step S103). Then, the information providing device 10 selects an element candidate whose increase amount between the search history and the posting history at a predetermined date and time exceeds a predetermined threshold value among the element candidates (step S104).

Further, the information providing device 10 uses the selected element candidate, the posting history posted on a predetermined date and time, and the related information as learning data (step S105). Then, the information providing device 10 learns each model using the learning data (step S106). Further, the information providing device 10 identifies the posted information including the new element by using the determination model (step S107), and uses the extraction model and the relationship estimation model from the identified posted information to obtain the new element. The relationship information is extracted (step S108). Then, the information providing device 10 registers the new element and the related information in the knowledge database (step S109), and ends the process.

[4. Modification example]
In the above, an example of the selection process, the learning process, and the update process by the information providing device 10 has been described. However, the embodiment is not limited to this. Hereinafter, variations of the provision process and the selection process executed by the information providing device 10 will be described.

[4-1. About the type of entity]
In the above example, as an example of the knowledge database, a knowledge database in which triples associated with a person and a person's occupation are registered is shown. However, the embodiment is not limited to this. That is, the information providing device 10 may update the knowledge database indicating arbitrary things. More specifically, the information providing device 10 may select arbitrary things as new elements and generate learning data, and may extract arbitrary things as new elements.

[4-2. Device configuration〕
The information providing device 10 does not have to manage the knowledge database by its own device. For example, each database 31 to 36 registered in the storage unit 30 may be stored in an external storage server. Further, the information providing device 10 may be realized by a front-end server that realizes a search process and a back-end server that realizes a selection process. In such a case, each of 41 to 46 shown in FIG. 2 is arranged on the back-end server, and a functional configuration for exerting a function of searching the knowledge database is arranged on the front-end server.

[4-3. others〕
Further, among the processes described in the above-described embodiment, all or a part of the processes described as being automatically performed can be manually performed, and conversely, the processes described as being manually performed. It is also possible to automatically perform all or part of the above by a known method. In addition, the processing procedure, specific name, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified. For example, the various information shown in each figure is not limited to the illustrated information.

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

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

[4-4. program〕
Further, the information providing device 10 according to the above-described embodiment is realized by, for example, a computer 1000 having a configuration as shown in FIG. FIG. 11 is a diagram showing an example of a hardware configuration. The computer 1000 is connected to the output device 1010 and the input device 1020, and the arithmetic unit 1030, the primary storage device 1040, the secondary storage device 1050, the output IF (Interface) 1060, the input IF 1070, and the network IF 1080 are connected by the bus 1090. Has.

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

The output IF 1060 is an interface for transmitting information to be output to an output device 1010 that outputs various information such as a monitor and a printer. For example, USB (Universal Serial Bus), DVI (Digital Visual Interface), and the like. It is realized by a connector of a standard such as HDMI (registered trademark) (High Definition Multimedia Interface). Further, 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, USB.

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

The network IF1080 receives data from another device via the network N and sends it to the arithmetic unit 1030, and also transmits the data generated by the arithmetic unit 1030 to the other device via the network N.

The arithmetic unit 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 unit 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 unit 1030 of the computer 1000 realizes the function of the control unit 40 by executing the program loaded on the primary storage device 1040.

[5. effect〕
As described above, the information providing device 10 specifies the search history regarding the element candidate which is a candidate for a new element and the posted information regarding the element candidate. Then, the information providing device 10 generates learning data for making the model learn the characteristics of the information about the new element among the element candidates based on the specified search history and the specified posted information. Select element candidates for. As a result, the information providing device 10 can make the model learn the characteristics of the information about the new element, and as a result, it is possible to realize the extraction of the new element using the model and to improve the efficiency of adding the new entity. can.

For example, the information providing device 10 selects an element candidate for generating learning data based on the search status indicated by the search history and the posting status indicated by the posting information. Further, for example, the information providing device 10 selects element candidates for generating learning data based on the number of searches indicated by the search history and the number of posts indicated by the posted information. Therefore, the information providing device 10 can select an element candidate that is likely to be a new element as an element candidate for generating learning data.

Further, when the explanatory content for explaining the element candidate is registered, the information providing device 10 acquires the element candidate from the explanatory content. Then, the information providing device 10 specifies the search history of the element candidate and the posted information about the element candidate. For example, the information providing device 10 identifies the posted information of the element candidate, which is posted within a predetermined period before the explanatory content is registered. Therefore, since the information providing device 10 selects the learning data generation target from the element candidates having a high possibility of new elements, the accuracy of the learning data can be improved.

Further, in the information providing device 10, the increase amount of the number of search histories exceeds a predetermined threshold value and the increase amount of the number of posted information exceeds a predetermined threshold value at a predetermined date and time among the plurality of element candidates. Select element candidates. For example, in the information providing device 10, among the plurality of element candidates, the search history and the posted information before the predetermined date and time do not exist, and the number of search histories exceeds the predetermined threshold value at the predetermined date and time. In addition, element candidates whose number of posted information exceeds a predetermined threshold are selected. Therefore, the information providing device 10 can improve the accuracy of the learning data.

Further, the information providing device 10 extracts the posted information posted on a predetermined date and time from the posted information related to the selected element candidates as learning data. Further, the information providing device 10 extracts the relationship between the element candidate and other elements as learning data from the explanatory content explaining the selected element candidate. Therefore, the information providing device 10 can generate learning data that realizes learning of the model for detecting the new element.

Further, the information providing device 10 causes the model to learn the characteristics of the information regarding the new element by using the information regarding the selected element candidate. For example, the information providing device 10 causes the model to learn the element candidate, the posted information about the element candidate, and the relationship between the element candidate and another element as learning data. For example, the information providing device 10 learns a determination model for determining whether or not the posted information is a post related to a new element, using the posted information related to the selected element candidate. Further, for example, the information providing device 10 learns an element extraction model that extracts a new element included in the posted information by using the selected element candidate and the posted information about the element candidate. Further, for example, the information providing device 10 uses the posted information regarding the element candidate and the relationship between the element candidate and another element to determine the relationship between the new element included in the posted information and the other element. Learn the relationship estimation model to be extracted. More specifically, the information providing device 10 is a plurality of models for extracting other elements having a predetermined relationship with a new element from the posted information as a relationship estimation model, and each has a different relationship. Learn multiple models to extract other elements of sex.

As a result of such processing, the information providing device 10 can accurately detect the new element from various posted information and estimate the relationship between the new element and the other element. As a result, the information providing device 10 can realize, for example, automatic updating of triples possessed by the knowledge database.

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

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

10 Information provider 20 Communication unit 30 Storage unit 31 Entity database 32 Triple database 33 Search log database 34 Post information database 35 Learning data database 36 Model database 40 Control unit 41 Acquisition unit 42 Specific unit 43 Selection unit 44 Extraction unit 45 Learning unit 46 Update 100 Log server 200 Description Content server

Claims (3)

An input layer where post information about a given element is entered,
Output layer and
A first element that is any layer from the input layer to the output layer and belongs to a layer other than the output layer.
Includes a first element and a second element whose value is calculated based on the weight of the first element.
With respect to the information input to the input layer, each element belonging to each layer other than the output layer is set as the first element, and the post is performed by performing an operation based on the weights of the first element and the first element. So that the output layer outputs a value indicating whether or not the information is posted information about a new element.
It ’s a model for making a computer work .
The weight of the first element is selected based on the search status indicated by the search history for the element candidate and the posting status indicated by the post information regarding the element candidate among the element candidates that are candidates for the new element. Based on learning using learning data based on element candidates
A model that features that. An input layer where post information about new elements is entered,
Output layer and
A first element that is any layer from the input layer to the output layer and belongs to a layer other than the output layer.
Includes a first element and a second element whose value is calculated based on the weight of the first element.
With respect to the information input to the input layer, each element belonging to each layer other than the output layer is set as the first element, and the post is performed by performing an operation based on the weights of the first element and the first element. A value indicating information indicating a new element among the information contained in the information is output from the output layer.
It ’s a model for making a computer work .
The weight of the first element is selected based on the search status indicated by the search history for the element candidate and the posting status indicated by the post information regarding the element candidate among the element candidates that are candidates for the new element. Based on learning using learning data based on element candidates
A model that features that. An input layer where post information about new elements is entered,
Output layer and
A first element that is any layer from the input layer to the output layer and belongs to a layer other than the output layer.
Includes a first element and a second element whose value is calculated based on the weight of the first element.
With respect to the information input to the input layer, each element belonging to each layer other than the output layer is set as the first element, and the post is performed by performing an operation based on the weights of the first element and the first element. Of the information contained in the information, a value indicating the relationship between the new element and another element is output from the output layer.
It ’s a model for making a computer work .
The weight of the first element is selected based on the search status indicated by the search history for the element candidate and the posting status indicated by the post information regarding the element candidate among the element candidates that are candidates for the new element. Based on learning using learning data based on element candidates
A model that features that.

Images