JP2018205978A - Information extracting device and information extracting method - Google Patents

Abstract

To improve a precision when information is extracted from a web page.SOLUTION: An information extracting device 1 includes: a contents obtaining unit 131 that obtains contents of multiple web sites; an image creating unit 132 that creates a screen shot image in a state in which the contents obtained by the contents obtaining unit 131 are displayed on a screen; and a specifying unit 133 that specifies information to be extracted contained in the screen shot image by utilizing the screen shot image as input data to a deep learning model created by deep learning on the basis of multiple learning image contents that contain learning information.SELECTED DRAWING: Figure 3

Description

  The present invention relates to an information extraction apparatus and an information extraction method for extracting information from a web page.

  Conventionally, a method of extracting information by crawling a website is known. Patent Document 1 discloses a search engine that extracts useful data included in a web page. In a conventional search engine, information is extracted based on text in a web page. For example, a conventional search engine extracts date information on the condition that there is a text such as “year”, “month”, “day” indicating the date.

Special table 2014-522030 gazette

  However, when extracting information based on text, the extracted information is often not desired information. For example, when extracting information about an event related to sports or amusement on a web page, the search engine extracts information indicating a date “April 1, 2017” as the date of the event. However, in some cases, the extracted date is not the date of the event, but the deadline for applying to participate in the event. Therefore, when the search engine extracts information based on the text, there has been a problem that incorrect information may be extracted.

  Therefore, the present invention has been made in view of these points, and an object thereof is to provide an information extraction apparatus and an information extraction method capable of improving accuracy when extracting information from a web page.

  An information extraction device according to a first aspect of the present invention creates a content acquisition unit that acquires content of a plurality of websites, and a screenshot image in a state where the content acquired by the content acquisition unit is displayed on a screen. Included in the screenshot image by using the screenshot image as input data to an image creation unit and a deep learning model created by deep learning based on a plurality of learning image contents including learning information And a specifying unit for specifying the extraction target information.

  The deep learning model is created by performing deep learning using first position information indicating a position where the learning information is included in the learning image content, and the specifying unit includes the first learning information. As the input data to the deep learning model associated with the position information, the extraction target information is specified by using the second position information indicating the position of the image region including the character string used in the extraction target information. Also good.

  The specifying unit creates a character image based on a predetermined character string included in the content acquired by the content acquisition unit, and an area having a correlation degree with the character image in the screen shot image is a threshold value or more. The position of the image area may be specified by specifying. The specifying unit may specify, as the extraction target information, event information including at least one of an event date and time, a place, and a content related to an event based on the position of the image area.

  Further, the specifying unit uses two or more of the text, the screenshot image, and the second position information included in the content acquired by the content acquisition unit as input data to the deep learning model. The extraction target information may be specified. In this case, when the accuracy of specifying the extraction target information using the text and the screenshot image as input data to the deep learning model is less than a threshold, the specifying unit further adds the second position information. It may be used as input data to the deep learning model.

  When the accuracy when the extraction target information is specified using the text as the first number of input data among the plurality of input data to the deep learning model is less than the threshold, The extraction target information may be specified using the screenshot image as the second number of input data.

  The specifying unit may receive specification of the type of the extraction target information to be specified, and specify the extraction target information using the deep learning model corresponding to the specified type.

  The information extraction device may further include an advertisement providing unit that provides an advertisement associated with the extraction target information specified by the specifying unit.

  The information extraction method according to the second aspect of the present invention includes a step of acquiring contents of a plurality of websites executed by a computer, a step of creating a screenshot image in a state where the acquired contents are displayed on a screen, By using the screenshot image as input data of a deep learning model created by deep learning based on a plurality of learning image contents including learning information, the extraction target information included in the screenshot image is obtained. Identifying.

  According to the present invention, it is possible to improve the accuracy when extracting information from a web page.

It is a figure for demonstrating the outline | summary of the information extraction apparatus of 1st Embodiment. It is a figure for demonstrating the outline | summary of the method by which an information extraction device extracts event information. It is a figure which shows the structure of an information extraction apparatus. It is a figure which shows an example of the content which a content acquisition part acquires. It is a flowchart which shows the procedure of the operation | movement which an identification part specifies event information based on position information. It is a figure which shows an example of the database with which event information was registered. It is an operation | movement flowchart of an information extraction apparatus. It is a figure which shows the structure of the information extraction apparatus of 2nd Embodiment. It is a figure which shows the advertisement information displayed with a content. It is an operation | movement flowchart of the information extraction apparatus of 2nd Embodiment.

<First Embodiment>
[Outline of Information Extraction Apparatus 1]
FIG. 1 is a diagram for explaining an overview of an information extraction apparatus 1 according to the first embodiment. The information extraction device 1 is a computer for extracting desired information from content included in web pages provided from a plurality of servers 2 accessible via the Internet N. The information extraction apparatus 1 extracts various types of extraction target information set in advance from the content of the web page, and registers the extracted extraction target information in the database 3. In the present embodiment, the information extraction apparatus 1 extracts event information related to various events as extraction target information and registers the extracted event information in the database 3, but the information extraction apparatus 1 extracts the information. The extraction target information is not limited to event information.

  An event is an event held on a specific day or period, for example, a sporting game, a festival, an exhibition, or a special sale. The event information is information including at least one of an event date or period, an event location, and an event content.

  The event information registered in the database 3 can be used by various applications. For example, a car navigation system mounted on a vehicle acquires event information registered in the database 3 and holds it within a predetermined range from a route to the current position of the vehicle or a destination based on the acquired event information. Extracted events. The car navigation system displays information on the extracted event, so that a person in the vehicle can recognize that the event is being held nearby.

[Outline of event information extraction method]
FIG. 2 is a diagram for explaining an outline of a method by which the information extraction apparatus 1 extracts event information. The information extraction apparatus 1 can extract event information with high accuracy by using various data obtained from web content included in a web page as input data of a deep learning model created in advance. The deep learning model is a model configured by a neural network in which coefficients are determined by learning a relationship between input variables and output variables using a large number of teacher data used as learning information.

  The information extraction device 1 can use a combination of text, an image, and position information indicating the coordinates of a predetermined image in a web page as input data to the deep learning model. When the text in a web page is used as input data for a deep learning model, the information extraction apparatus 1 extracts a plurality of pre-registered texts from the web page source code, and inputs the extracted plurality of texts into the deep learning model. Data. By doing in this way, the information extraction apparatus 1 can extract event information correctly with high probability based on the combination of a plurality of texts input to the deep learning model.

  The information extraction device 1 can also use a screenshot image obtained by rendering the source code in the web page instead of the text or together with the text as input data of the deep learning model. The screen shot image is an image in a state where a web page is displayed on a computer screen. An image deep learning model that can use a screen shot image as input data is created using a large number of screen shot images for learning. The information extraction apparatus 1 can extract information that is recognized as event information when the user visually recognizes the web page by using the screen shot image as input data of the deep learning model for images. The probability of extraction is further increased.

  For example, when it is considered that the accuracy when event information is extracted using only text is low, the information extraction apparatus 1 may use a screenshot image together with the text as input data of the deep learning model. The information extraction device 1 combines the text and the screenshot image, uses the text as part of the input data of the deep learning model, and uses it as the input data of the other part of the deep learning model. The probability that it can be extracted can be further increased.

  The information extraction apparatus 1 can further increase the probability that event information can be correctly extracted by specifying a position of a predetermined text in a screenshot image and using coordinates indicating the specified position as input data of a deep learning model. . Details of how the information extracting apparatus 1 identifies the position of a predetermined text in the screenshot image will be described later.

[How to create a deep learning model]
The deep learning model can be created using various known methods. When creating a deep learning model that can use text as input data, text included in a large number (for example, 1 million) of web pages is used as teacher data. The creator of the deep learning model identifies event information that can be grasped based on the web page by visually recognizing the web page used for learning. Then, it is possible to create a deep learning model for text using text as input data by associating the identified learning event information with text extracted from the source code of a web page used as teacher data it can. The event information specified by the creator of the deep learning model may be the same as the text included in the web page, or may be different content from the text included in the web page.

  Similarly, when creating a deep learning model that can use screen shot images as input data, screen shot images of many web pages are used as teacher data. The event information for learning specified by the creator of the deep learning model visually confirming the web page used for learning is obtained by rendering the source code of the web page used as teacher data. By associating with a screenshot image, it is possible to create a deep learning model for an image using the screenshot image as input data.

  Further, when creating a deep learning model that can use position information as input data, the positions of texts included in many web pages are used as teacher data. The event information for learning specified by the creator of the deep learning model visually confirming the web page used for learning is associated with the position information of the text included in the web page used as teacher data. Thus, it is possible to create a deep learning model for position using position information as input data.

  In addition, the creator of the deep learning model regularly updates the deep learning model by using a new learning web page to match the trend of the most recent web page configuration. A deep learning model can be created.

Further, the creator of the deep learning model can create a deep learning model for each type of information that is desired to be extracted from the web page. For example, the creator of a deep learning model creates a deep learning model using teacher data that includes event information related to a soccer game, thereby increasing the probability that event information related to a soccer game can be correctly extracted. A model can be created. The information extraction apparatus 1 can increase the probability that desired information can be correctly extracted by using the deep learning model selected based on the type of information to be extracted.
Hereinafter, the configuration and operation of the information extraction device 1 will be described in detail.

[Configuration of Information Extraction Apparatus 1]
FIG. 3 is a diagram illustrating a configuration of the information extraction device 1. The information extraction device 1 includes a communication unit 11, a storage unit 12, and a control unit 13.

  The communication unit 11 is a communication interface including a communication controller for the information extraction apparatus 1 to transmit and receive data between the server 2 and the database 3 via the Internet N. The communication unit 11 inputs the content of the web page received via the Internet N to the control unit 13. In addition, the event information output by the control unit 13 is transmitted to the database 3.

  The storage unit 12 includes a storage medium such as a ROM (Read Only Memory), a RAM (Random Access Memory), and a hard disk. The storage unit 12 stores a program executed by the control unit 13.

  The control unit 13 is, for example, a CPU (Central Processing Unit), and functions as a content acquisition unit 131, an image creation unit 132, a specifying unit 133, and a registration unit 134 by executing a program stored in the storage unit 12.

  The content acquisition unit 131 acquires content of a plurality of websites via the communication unit 11. The content acquisition unit 131 stores the acquired content in the storage unit 12.

  FIG. 4 is a diagram illustrating an example of content acquired by the content acquisition unit 131. The content shown in FIG. 4 includes information related to cherry-blossom viewing events in U Park. In the following description, a process for specifying event information based on the content shown in FIG. 4 will be described.

  When extracting event information based on text, the content acquisition unit 131 inputs the acquired content to the specifying unit 133. When the content acquisition unit 131 extracts event information based on a screenshot image, the content acquisition unit 131 also inputs the acquired content to the image creation unit 132.

  The image creation unit 132 creates a screenshot image in a state where the content acquired by the content acquisition unit 131 is displayed on the screen. The image creating unit 132 inputs the created screen shot image to the specifying unit 133.

  The specifying unit 133 uses the text, the screenshot image, and the position information included in the content acquired by the content acquisition unit 131, thereby generating event information as extraction target information included in the content acquired by the content acquisition unit 131. Identify.

  When the text included in the content shown in FIG. 4 is used as input data for the deep learning model, the specifying unit 133, for example, “sakura festival”, “opening time”, “venue”, “access”, “late March” "Early April", "March 25-April 5", "8: 00-21: 00", "U Park", "March 20, 2017" are used as input data for the deep learning model Extract as text. As a result, from the deep learning model, the specifying unit 133 has an event date and time of “March 25 to April 5” “8:00 to 21:00”, and the event is held at “U Park”. Yes, it is possible to obtain an output indicating that the content of the event is “Sakura Festival”.

  However, since the content shown in FIG. 4 also includes the text “March 20, 2017” indicating the date, the deep learning model incorrectly indicates that the event date is “March 20, 2017”. Output may be obtained. Therefore, the specifying unit 133 can increase the probability of specifying the correct extraction target information by using two or more of the text, the screenshot image, and the position information as inputs to the learning model. For example, the specifying unit 133 selects the second when the accuracy when the extraction target information is specified using text as the first number of input data among the plurality of input data to the deep learning model is less than the threshold value. The extraction target information may be specified using screen shot images as the number of input data.

  In the case of the example shown in FIG. 4, a net pattern is added around the text “March 25 to April 5” indicating the date of the cherry blossom festival. When the screen shot image is used as the input data to the deep learning model, the specifying unit 133 uses the network among the plurality of texts “March 25 to April 5” and “March 20, 2017” indicating the date. The text surrounded by the pattern “March 25th to April 5th” is output from the deep learning model. As described above, the specifying unit 133 can correctly specify event information by using the screen shot image as input data to the deep learning model.

  The specifying unit 133 also determines the position of the image area including the character string used in the event information when the accuracy of specifying the event information using the text and the screenshot image as the input to the deep learning model is less than the threshold. Is used as an input to the learning model. Specifically, the specifying unit 133 performs deep learning for position created by performing deep learning using position information indicating a position where a predetermined text is included as learning information in the learning image content. A predetermined character string included in the content acquired by the content acquisition unit 131 and coordinates indicating the position of the character string are input to the model. By doing so, event information corresponding to the position of the input character string is output from the deep learning model.

  The specifying unit 133 creates a character image based on a predetermined character string included in the content acquired by the content acquisition unit 131 in order to specify the position of the character string, and the degree of correlation with the character image in the screenshot image The region where is equal to or greater than the threshold is specified. The specifying unit 133 can specify the position of the character string used as the input data of the deep learning model by specifying the position of the image region including the text used for the event information. In this way, the specifying unit 133 can include an event including at least one of the date / time, location, and content of the event related to the event based on the position of the image area that includes the text used for the event information. Information can be specified as extraction target information.

  FIG. 5 is a flowchart showing a procedure of an operation in which the specifying unit 133 specifies event information based on position information. Hereinafter, an operation in which the specifying unit 133 specifies event information based on position information will be described with reference to FIGS. 4 and 5.

  First, the specifying unit 133 renders the content acquired by the content acquisition unit 131 to create a screenshot image (S1). Subsequently, the predetermined text included in the content acquired by the content acquisition unit 131 is converted into an image (S2). In the case of the example illustrated in FIG. 4, the specifying unit 133 performs “sakura festival”, “opening time”, “venue”, “access”, “late March to early April”, “March 25 to April 5”. The text “day”, “8:00 to 21:00”, “U park”, “March 20, 2017” is converted into an image.

  Subsequently, the specifying unit 133 searches for a position in the screen shot image created in step S1 of each of the converted images (S3). The specifying unit 133 specifies the position of each text on the web page by specifying an image region having the highest correlation with the image obtained by converting the text in the screen shot image (S4). The specifying unit 133 associates each text with the coordinates of the image area corresponding to the text, and stores them in the storage unit 12 (S5).

  Subsequently, the specifying unit 133 inputs the text and coordinates stored in the storage unit 12 in step S5 to the deep learning model (S6). The specifying unit 133 specifies event information output from the deep learning model based on the positional relationship between the texts (S7), and notifies the registration unit 134 of the event information.

  In the example shown in FIG. 4, text indicating the date of the event is arranged to the right of the text position indicating the content of the event “Sakura Festival”. In addition, a text indicating the place where the event is held is arranged to the right of the position of the text “venue”. On the other hand, the date and time arranged at the lower right of the web page is highly likely not the event date and time. As described above, it is considered that there is a certain relationship between the position where the text indicating the event information is arranged and the position where the predetermined text is arranged. Therefore, the specifying unit 133 inputs the text and the coordinates of the image region corresponding to the text into the deep learning model for the position created based on the teacher data including the position information of the text in many web pages. The accuracy of identifying event information can be improved.

  The specifying unit 133 specifies a large number of event information using at least one of text, screenshot images, and position information based on the contents of a large number of web pages, and the registration unit 134 sequentially stores the large number of event information. Register in database 3.

  FIG. 6 is a diagram illustrating an example of the database 3 in which event information is registered. In the event information database shown in FIG. 6, an event number, an event date, an event holding time, an event holding location, and an event content are associated with each other. The event information identified based on the web page shown in FIG. 4 is event information having an event number of 0002.

  When event information specified from different web pages is different, the registration unit 134 may register only event information that matches in a predetermined percentage or more of web pages in the database 3. For example, the registration unit 134 may not register the event information having different dates when the event dates are different among only one of the event information specified based on the plurality of web pages.

  The registration unit 134 registers the specified event information in the database 3 on condition that the event information specified by the specifying unit 133 is composed of a character string different from the event information already registered in the database 3. May be. By doing in this way, it can prevent that many information regarding the same event is registered into the database 3.

  Further, the registration unit 134 may register a numerical value corresponding to the number of web pages that is the basis for specifying the event information in the database 3 in association with the event information. An application that refers to event information registered in the database 3 can select event information with high reliability by using the registered numerical value.

  The registration unit 134 may register all event information specified by the specifying unit 133 in the database 3. In this case, the user who uses the application can acquire appropriate event information by selecting the event information to be used on the application side referring to the database 3 based on the accuracy required by the application.

[Operation Flowchart of Information Extraction Apparatus 1]
FIG. 7 is an operation flowchart of the information extraction apparatus 1. When the information extraction apparatus 1 starts processing for specifying event information, first, the content acquisition unit 131 acquires content of a large number of web pages (S11). When the content acquisition unit 131 acquires content, the specifying unit 133 specifies event information by inputting text included in the content to the deep learning model for text (S12).

  When the specifying unit 133 determines that the accuracy of the event information specified based on the text is equal to or higher than the threshold (Yes in S13), the process proceeds to step S17, and the registration unit 134 registers the event information in the database 3 (S17). ). On the other hand, when determining that the accuracy of the event information specified based on the text is less than the threshold (No in S13), the specifying unit 133 proceeds to step S14 and inputs the screenshot image to the deep learning model for images. Thus, event information is specified (S14). In step S14, the specifying unit 133 may specify event information using both the text and the screen shot image.

  When the specifying unit 133 determines that the accuracy of the event information specified based on the screenshot image is equal to or higher than the threshold (Yes in S15), the process proceeds to step S17, and the registration unit 134 registers the event information in the database 3. (S17). On the other hand, when determining that the accuracy of the event information specified based on the screenshot image is less than the threshold value (No in S15), the specifying unit 133 proceeds to step S16 to convert the position information into the deep learning model for position. Event information is specified by inputting (S16). The process in step S16 is the process from step S1 to S7 shown in FIG. In step S16, the specifying unit 133 may specify event information by combining all of the text, the screenshot image, and the position information.

  As described above, the information extraction device 1 combines the text included in the web page, the screen shot image created based on the source code, and the position information indicating the position of the text related to the event into the input data of the deep learning model. As a result, event information can be specified with high accuracy.

[Modification 1]
In the above description, the specifying unit 133 uses a predetermined deep learning model. However, depending on the type of event, the nationality of the person who created the web page, the language used in the web page, and the like, It is thought that the tendency of the method of posting extraction target information is different. Therefore, the specifying unit 133 may use different deep learning models based on the type of the extraction target information in order to increase the probability that the extraction target information can be correctly specified. Specifically, the specifying unit 133 receives a specification of the type of extraction target information to be specified from an external computer via the communication unit 11 and uses a deep learning model corresponding to the specified type. Extraction target information can be specified.

  The type of extraction target information is, for example, the language of the target web page, the country in which the web page was created, and the type of event. Specifically, the specifying unit 133 may select “a soccer related event posted on a Japanese site”, “a soccer related event posted on an English site”, “an art related event posted on an English site”, or “China Music-related events posted on a language site ”or the like can be used as the type of extraction target information. The specifying unit 133 can increase the probability that the extraction target information can be correctly specified by using a deep learning model suitable for specifying the extraction target information of the specified type.

[Effects of the information extraction apparatus 1 according to the first embodiment]
As described above, the information extraction apparatus 1 according to this embodiment includes the image creation unit 132 that creates a screenshot image in which the content acquired by the content acquisition unit 131 is displayed on the screen, and the input to the deep learning model. By using a screen shot image as data, it has a specifying unit 133 that specifies event information as extraction target information included in the screen shot image. In this manner, the event information included in the web page can be identified based on the tendency of the screen visually recognized by the person viewing the web page by identifying the event information using the screenshot image. The probability that event information can be correctly identified can be increased.

  In particular, the specifying unit 133 indicates the second position indicating the position of the image region including the character string used in the event information as input data to the deep learning model associated with the first position information of the predetermined text related to the event. By specifying the event information using the position information, the event information included in the web page can be specified based on the tendency of the positional relationship between the texts included in the web page. Therefore, even when a plurality of texts similar to the event information are included in the web page, the specifying unit 133 can correctly specify the event information with a high probability.

  The present invention is also effective when a search engine ranks web pages in descending order of relevance with a search keyword. When using a method that identifies the relationship between a search keyword and a web page based only on text, as in conventional search engines, take SEO (Search Engine Optimization) measures to embed the search keyword in a tag that the user does not see Thus, it becomes possible to display the ranking of the web page at the top. On the other hand, in this invention, a web page can be ranked based on the content which the user who browses a web page can visually recognize by using the pixel data of a snapshot image. Therefore, even when a code for SEO countermeasures is embedded in the web page, it is possible to improve the search accuracy.

Second Embodiment
FIG. 8 is a diagram illustrating a configuration of the information extraction device 4 according to the second embodiment. The information extraction device 4 is different from the information extraction device 1 in that an advertisement providing unit 135 is provided instead of the registration unit 134 in the information extraction device 1 of the first embodiment.

  The method for specifying information by the specifying unit 133 in the information extracting device 4 is the same as that in the first embodiment. Information other than event information is also specified as information. The extraction target information is, for example, text and images included in a dictionary created in advance.

  The storage unit 12 stores advertisement information in association with various types of information that can be specified by the specifying unit 133. When the information provided by the specifying unit 133 is acquired, the advertisement providing unit 135 provides the server 2 with the advertisement information stored in the storage unit 12 in association with the acquired information. The server 2 transmits the advertisement information provided from the advertisement providing unit 135 together with the content of the web page to the terminal that accesses the web page.

  FIG. 9 is a diagram showing advertisement information A1 and advertisement information A2 displayed together with content. The specifying unit 133 specifies various types of information in the web page shown in FIG. The specifying unit 133 specifies information such as “U Park”, “Ohanami”, “Cherry Blossom”, “Festival”, for example. The specifying unit 133 also specifies the position where the specified information is displayed, and determines the importance of the information based on the specified position.

  The specifying unit 133 notifies the advertisement providing unit 135 of information whose importance is greater than or equal to a threshold value. The specifying unit 133 notifies the advertisement providing unit 135 of information such as “U park”, for example. In this case, the advertisement providing unit 135 transmits the advertisement information A1 and the advertisement A2 stored in the storage unit 12 in association with “U Park” to the server 2, so that the advertisement information A1 and the advertisement A2 are displayed on the web page. The

  FIG. 10 is an operation flowchart of the information extraction device 4 of the second embodiment. S21 to S26 correspond to S11 to S16 in the operation flowchart shown in FIG. However, the information specified in S22, S24, and S26 is extraction target information that is not limited to event information. When the specifying unit 133 specifies the extraction target information in steps S22 to S26, the advertisement providing unit 135 selects the advertisement information based on the specified extraction target information. Subsequently, in S <b> 28, the advertisement providing unit 135 provides the selected advertisement information to the server 2.

[Effects of Information Extraction Device 4 of Second Embodiment]
As described above, in the information extraction device 4 of the present embodiment, the specifying unit 133 specifies the extraction target information in the web page, and the advertisement providing unit 135 has the advertisement information associated with the specified extraction target information. Is provided to the server 2. In this way, by using the information extraction device 4, advertisements regarding products and services related to the content of the web page can be displayed in the web page, so the probability that the user browsing the web page is interested in it. Can be increased.

  As mentioned above, although this invention was demonstrated using embodiment, the technical scope of this invention is not limited to the range as described in the said embodiment, A various deformation | transformation and change are possible within the range of the summary. is there. For example, the specific embodiments of device distribution / integration are not limited to the above-described embodiments, and all or a part of them may be configured to be functionally or physically distributed / integrated in arbitrary units. Can do. In addition, new embodiments generated by any combination of a plurality of embodiments are also included in the embodiments of the present invention. The effect of the new embodiment produced by the combination has the effect of the original embodiment.

DESCRIPTION OF SYMBOLS 1 Information extraction apparatus 2 Server 3 Database 4 Information extraction apparatus 11 Communication part 12 Storage part 13 Control part 131 Content acquisition part 132 Image creation part 133 Identification part 134 Registration part 135 Advertisement provision part

Claims (10)

A content acquisition unit for acquiring content of a plurality of websites;
An image creation unit that creates a screenshot image in a state where the content acquired by the content acquisition unit is displayed on a screen;
By using the screenshot image as input data to a deep learning model created by deep learning based on a plurality of learning image contents including learning information, the extraction target information included in the screenshot image is obtained. A specific part to identify;
An information extraction apparatus having The deep learning model is created by performing deep learning using first position information indicating a position where the learning information is included in the learning image content,
The specifying unit uses, as input data to the deep learning model associated with the first position information, second position information indicating a position of an image area including a character string used in the extraction target information. Specifying the extraction target information;
The information extraction device according to claim 1. The specifying unit creates a character image based on a predetermined character string included in the content acquired by the content acquisition unit, and an area having a correlation degree with the character image in the screen shot image is a threshold value or more By specifying the position of the image area,
The information extraction device according to claim 2. The specifying unit specifies, as the extraction target information, event information including at least one of an event date and time, a place, and contents related to an event based on the position of the image area.
The information extraction device according to claim 2 or 3. The specifying unit uses the two or more of the text included in the content acquired by the content acquisition unit, the screenshot image, and the second position information as input data to the deep learning model. Identify target information,
The information extraction device according to any one of claims 2 to 4. The specifying unit further uses the text and the screenshot image as input data to the deep learning model to specify the extraction target information when the accuracy of specifying the extraction target information is less than a threshold value. Used as input data to the model,
The information extraction device according to claim 5. When the accuracy when the extraction target information is specified using the text as the first number of input data among the plurality of input data to the deep learning model is less than the threshold, Identifying the extraction target information using the screenshot image as a second number of input data;
The information extraction device according to claim 5 or 6. The specifying unit receives the specification of the type of the extraction target information to be specified, and specifies the extraction target information using the deep learning model corresponding to the specified type.
The information extraction device according to any one of claims 1 to 7. An advertisement providing unit that provides an advertisement associated with the extraction target information identified by the identifying unit;
The information extraction device according to any one of claims 1 to 8. The computer runs,
Acquiring content from multiple websites;
Creating a screenshot image of the acquired content displayed on the screen;
By using the screenshot image as input data of a deep learning model created by deep learning based on a plurality of learning image contents including learning information, the extraction target information included in the screenshot image is identified. And steps to
An information extraction method comprising:

Images