JP2018014003A - Item value extraction model learning device, item value extraction device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To learn an extraction model for extracting the value of an item to be extracted, with high comprehensiveness.SOLUTION: A designated item extraction unit 24 extracts the value of an item that matches a designated item name list from each of the structured data of text data groups, as the value of an item to be extracted. A pseudo-supervised data creation unit 26 creates text to which is added an annotation indicating that it is the extracted value of the item to be extracted, as pseudo-supervised data. An extraction model learning unit 30 learns an extraction model for extracting the value of the item to be extracted from the text on the basis of the identity extracted from the text of pseudo-supervised data and the annotation added to the text.SELECTED DRAWING: Figure 14

Description

  The present invention relates to an item value extraction model learning device, an item value extraction device, a method, and a program, and more particularly to an item value extraction model learning device and an item value extraction device for extracting an item value to be extracted from a text. , Method and program.

  Conventionally, a synonym relation extraction method for Wikipedia (R) entry-redirection is known (Non-Patent Document 1). In this method, words having synonymous relations are extracted using the redirect relation of Wikipedia (R).

  Also, a labeling method for extracting a person alias and identifying a person using Wikipedia (R) is known (Non-Patent Document 2). In this method, patterns of expressions that are easily described as aliases are set in the text, and aliases are extracted for objects that match the pattern. For example, aliases are extracted using the patterns “X” and “X”.

Junichi Ohno et al., “Extraction of synonym relations between entry and redirect in Wikipedia”, Proc. Of the 17th Annual Conference, Language Processing Society of Japan, p. 296-299, March 2011. Daiki Saito et al., “Extraction of Person Alias Using Wikipedia and Labeling for Person Discrimination”, Language Processing Society, 20th Annual Conference, Proceedings, p. 63-66, March 2014.

  However, other names related to the title are often written in the body of the article, and as described in Non-Patent Document 1, it cannot be covered only by the redirect relationship.

  There are various actual description patterns with different names, and the pattern set by hand is not very comprehensive. However, it is expensive to manually expand the pattern. Actually, in Non-Patent Document 2, only aliases of person names are targeted. However, in Wikipedia (R), there are many aliases other than person names.

  The present invention has been made in view of the above circumstances, and has an item value extraction model learning apparatus, method, and the like, capable of learning an extraction model for extracting the value of an item to be extracted with high completeness. And to provide a program.

  It is another object of the present invention to provide an item value extraction apparatus, method, and program capable of extracting the value of an item to be extracted with high completeness.

  In order to achieve the above object, the item value extraction model learning device according to the first invention provides an item name specified as an item name that defines an item to be extracted, and an item name and an item value related to the value and the value. An input unit that accepts a text data group consisting of a set of structured data including the set and a text including the value, and an item that matches the specified item name from each of the structured data of the text data group The specified item extracting unit that extracts the value of the item to be extracted as the value of the item to be extracted, and the text to which the annotation indicating that the value of the item to be extracted is extracted by the specified item extracting unit Are generated from the text of the pseudo-teacher data created by the pseudo-teacher data creation unit and the pseudo-teacher data creation unit If, based on the annotations assigned to the text, from the text, it is configured to include a, and extracting the model learning unit for learning the extraction model for extracting the value of the item to be the extraction target.

  In the item value extraction model learning method according to the second aspect of the invention, the input unit specifies an item name specified as an item name that defines an item to be extracted, and a set of an item name and an item value related to the value and the value. Accepting a text data group consisting of a set of structured data including and a text including the value, and the designated item extraction unit extracts items of the item that match the designated item name from each of the structured data of the text data group. The value is extracted as the value of the item to be extracted, and the pseudo-teacher data creation unit is provided with an annotation indicating that the value of the item to be extracted is extracted by the specified item extraction unit Text is created as pseudo-teacher data, and the extraction model learning unit is extracted from the text of the pseudo-teacher data created by the pseudo-teacher data creation unit, and Based on the serial granted to text annotations, from the text, it learns the extraction model for extracting the value of the item to be the extraction target.

  An item value extracting apparatus according to a third aspect of the present invention is directed to the structure included in a text data group including a set of structured data including a set of a value, an item name related to the value, and an item value, and a text including the value. In the text to which the annotation indicating that the value of the item to be extracted is attached to the value of the item name specified as the item name that defines the item to be extracted extracted from each of the digitized data An extraction model for extracting the value of the item to be extracted from the text, learned in advance based on a feature extracted from the text of a pseudo-teacher data and an annotation attached to the text. An extraction model storage unit for storing, a feature extraction unit for extracting the features from the input text, the features extracted by the feature extraction unit, and the extraction Based on the Dell, from the text, it is configured to include a, and item value extracting unit for extracting the value of the item to be the extraction target.

  An item value extraction method according to a fourth aspect of the present invention is directed to the structure included in a text data group including a set of structured data including a set of a value, an item name related to the value, and an item value, and a text including the value. In the text to which the annotation indicating that the value of the item to be extracted is attached to the value of the item name specified as the item name that defines the item to be extracted extracted from each of the digitized data An extraction model for extracting the value of the item to be extracted from the text, learned in advance based on a feature extracted from the text of a pseudo-teacher data and an annotation attached to the text. An item value extraction method in an item value extraction device including an extraction model storage unit for storing, wherein the feature extraction unit extracts the feature from input text, and an item Extraction unit, the said feature extracted by the feature extraction unit, based on said extracted model, from the text, and extracts the value of the item to be the extraction target.

  An item value extraction model learning device according to a fifth invention includes a title, structured data including a combination of an item name and an item value, an article group including text including the value, and an extraction target for the title. An input unit that accepts an item name designated as an item name that defines an item to be selected, and an item value that matches the designated item name from each of the structured data of the article group Pseudo-teacher data for creating, as pseudo-teacher data, a designated item extraction unit that extracts as values, and the text that is annotated to indicate the value of the item to be extracted extracted by the designated item extraction unit A creation unit, a feature extracted from the text of the pseudo-teacher data created by the pseudo-teacher data creation unit, and an annotation attached to the text Based on the Deployment, from the text, it is configured to include a, and extracting the model learning unit for learning the extraction model for extracting the value of the item to be the extraction target.

  An item value extraction apparatus according to a sixth invention extracts from each of structured data included in an article group including a title, structured data including a combination of an item name and an item value, and text including the value. The pseudo-teacher that is the text in which the annotation indicating that the value of the item to be extracted is given to the value of the item name specified as the item name that defines the item to be extracted with respect to the title Extraction that stores an extraction model for extracting the value of the item to be extracted from the text, learned in advance based on the features extracted from the text of the data and the annotations attached to the text A model storage unit, a feature extraction unit that extracts the features from the input text, the features extracted by the feature extraction unit, and the extraction model Based on the bets, from the text, it is configured to include a, and item value extracting unit for extracting the value of the item to be the extraction target.

  A program according to a seventh aspect is a program for causing each of the item value extraction model learning device or the item value extraction device to function.

  According to the item value extraction model learning device, method, and program of the present invention, the value of the item that matches the specified item name is extracted as the value of the item to be extracted from each of the structured data. The extracted annotated text indicating the value of the item to be extracted is created as pseudo-teacher data, and the feature extracted from the text of the pseudo-teacher data is added to the text. By learning the extraction model based on the annotation that has been made, it is possible to learn the extraction model for extracting the value of the item to be extracted with high completeness.

  According to the item value extraction apparatus, method, and program of the present invention, an annotation is given to the value of the item name specified as the item name that defines the item to be extracted extracted from each of the structured data. Further, based on the feature extracted from the text of the pseudo-teacher data which is the text, the extraction model learned in advance based on the annotation given to the text, and the feature extracted from the input text Thus, by extracting the value of the item to be extracted from the text, it is possible to extract the value of the item to be extracted with high completeness.

It is a figure which shows the display on the web of Infobox. It is a figure which shows the description in XML of Infobox. It is a figure which shows the Infobox of an article, and the example of a text. It is a block diagram which shows the structure of the item value extraction model learning apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the article group input. It is a figure which shows the example of structured data. It is a figure which shows the pair of the title and the alternative name extracted. It is a figure for demonstrating the method of producing pseudo teacher data. It is a block diagram which shows the structure of the item value extraction apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the article input. It is a figure which shows the alternative name extracted from the input article. It is a flowchart which shows the item value extraction model learning process routine in the item value extraction model learning apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the item value extraction process routine in the item value extraction apparatus which concerns on the 1st Embodiment of this invention. It is a block diagram which shows the structure of the item value extraction model learning apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows the text data group input. It is a block diagram which shows the structure of the item value extraction apparatus which concerns on the 2nd Embodiment of this invention. It is a figure which shows the input text. It is a flowchart which shows the item value extraction model learning process routine in the item value extraction model learning apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the item value extraction process routine in the item value extraction apparatus which concerns on the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

<Overview>

  In order to generate pseudo-alias correct answer data in the article text from the synonym information described in the structured data (for example, Infobox) included in the article, and to extract the alias of the title from the article text Build an extraction model. As a result, only the designation of the item name of the structured data is set in advance, and a large amount of alternative name description patterns can be obtained at a low cost, and an extraction model for extracting alternative names can be constructed.

  Here, as shown in the Web display in FIG. 1 and the XML notation in FIG. 2, the Infobox is a part of Wikipedia (R) where information is structured, and the item name and its value. Is a structure that can be easily extracted.

  In addition, among the articles of Wikipedia (R), as shown in FIGS. 3 (A) and 3 (B), there are articles in which titles are also described in Infobox and text. For articles with nicknames in both, Infobox can automatically obtain nicknames, so the description pattern in the text is specified from the information and used as training data for the extraction model.

[First Embodiment]
<Configuration of Item Value Extraction Model Learning Device According to First Embodiment of the Present Invention>
Next, the configuration of the item value extraction model learning device according to the first embodiment of the present invention will be described. As shown in FIG. 4, the item value extraction model learning device 100 according to the first exemplary embodiment of the present invention includes a CPU, a RAM, a program for executing an item value extraction model learning process routine described later, It can be constituted by a computer including a ROM storing data. Functionally, the item value extraction model learning device 100 includes an input unit 10, a calculation unit 20, and an output unit 50 as shown in FIG.

  The input unit 10 accepts an article group that is an article group including a title, text, and structured data as shown in FIG. The structured data includes a pair of an item name and an item value, and the text includes the value of the item.

  Further, the input unit 10 extracts a title, structured data, and a text portion from each article in the accepted article group. For example, the first <h1> value of <body> is extracted as a title.

  Further, the input unit 10 accepts an item name list specified as an item name that defines an alternative title. For example, the following item name list is accepted.

[Abbreviation | common name | name | nickname | English name | English name | English name | Also | nickname | alias | Alias | Nickname | abbreviation]

  The calculation unit 20 includes an article group storage unit 22, a specified item extraction unit 24, a pseudo teacher data creation unit 26, a feature extraction unit 28, and an extraction model learning unit 30.

  The article group storage unit 22 stores an article group received by the input unit 10 and an item name list.

  The designated item extraction unit 24 extracts the value of the item that matches the designated item name list from each of the structured data included in each article of the article group as another title. For example, as shown in FIG. 6, the item value “water monster” corresponding to the item name matching the “nickname” included in the specified item name list is extracted from the structured data (see FIG. 7). . However, cleaning such as deletion of link information and division of a plurality of candidates is also performed.

  The pseudo-teacher data creation unit 26 specifies a part where the alternative name is described in the text of the article based on the alternative name extracted by the designated item extraction unit 24, and the specified description part is an alternative name. And annotated text is created as pseudo-teacher data.

  At this time, in the article, the part where the value of the item corresponding to the item name matching the item name list first appears is specified, and an annotation is given. At this time, the reason for limiting to the first appearing part is that it is often explained that the nickname is the nickname as the characteristic of the Wikipedia (R) article when it appears for the first time.

  Thereby, as shown in FIG. 8, various alias description patterns in the text of an article can be automatically annotated.

  The feature extraction unit 28 generates a feature vector for the text of each pseudo teacher data created by the pseudo teacher data creation unit 26. For example, the feature vector is generated based on the character notation of each character of the text or the distributed representation of each word.

  Based on the feature vector extracted from the text of each pseudo-teacher data by the feature extraction unit 28 and the annotation added to the text of each pseudo-teacher data, the extraction model learning unit 30 uses a title nickname from the text of the article. An extraction model for extracting the value of is learned and output by the output unit 50.

  Specifically, the extraction model learning unit 30 assigns a character level tag (for example, a BIO tag) to the text of each pseudo teacher data according to the annotation. For example, an O tag (others) is attached except where annotation is added, and a B tag indicating the start of the character string to be extracted and an I tag indicating the character string to be extracted are added to the location where the annotation is added. To do. For example, when the part to which the annotation is added is a three-character notation, a B tag, an I tag, and an I tag are added to the three characters.

  Then, the extraction model learning unit 30 learns the extraction model using a character-level series labeling model based on the feature vector generated for each pseudo-teacher data and the assigned tag. The extraction model can be constructed using a general sequence labeling method such as CRF (for example, the method described in Non-Patent Document 3).

[Non-Patent Document 3]: J. Lafferty, A. McCallum, and F. Pereira. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In Proceedings of the 18th International Conference on Machine Learning, 2001.

<Configuration of Item Value Extraction Device According to First Embodiment of the Present Invention>
Next, the configuration of the item value extraction apparatus according to the first embodiment of the present invention will be described. As shown in FIG. 9, the item value extraction apparatus 150 according to the first embodiment of the present invention stores a CPU, a RAM, a program for executing an item value extraction processing routine described later, and various data. It can be composed of a computer including a ROM. The item value extraction apparatus 150 functionally includes an input unit 60, a calculation unit 70, and an output unit 90 as shown in FIG.

  The input unit 60 accepts an article composed of a title and text as shown in FIG. The accepted articles do not contain structured data, and the text contains alternative titles. In addition, the input unit 60 extracts a title and a text part from the received article in the same manner as the input unit 10.

  The calculation unit 70 includes a feature extraction unit 72, an extraction model storage unit 74, and a title item value extraction unit 76.

  The feature extraction unit 72 generates a feature vector similar to the feature extraction unit 28 for the text of the article received by the input unit 60.

  The extraction model storage unit 74 stores an extraction model learned by the item value extraction model learning device 100.

  The title item value extraction unit 76 performs series labeling based on the feature vector extracted by the feature extraction unit 72 and the extraction model stored in the extraction model storage unit 74, and adds a tag to each character of the article text. (For example, a BIO tag) is assigned, and an alternative title is extracted. For example, in the example of FIG. 11, if a B tag, an I tag, and an I tag are given to the three characters “HHH”, this portion is extracted as an alternative name for the title, and a pair of the title and the alternative name is output. 90 for output.

  In the present embodiment, an extraction target portion with another name is a title definition sentence portion in the text. As a result of totaling the results of creating the pseudo-correct answer data, about 85% of the parts where the alternative names are described are the definition sentence part. By limiting to this part, efficient extraction becomes possible.

<Operation of Item Value Extraction Model Learning Device According to First Embodiment of the Present Invention>
Next, the operation of the item value extraction model learning device 100 according to the first embodiment of the present invention will be described. When the input unit 10 receives an article group including a title, text, and structured data, the title, structured data, and text portion are extracted from each article of the received article group and stored in the article group storage unit 22.

  In addition, when the input unit 10 receives an item name list designated as an item name that defines an alternative title, the item name list is stored in the article group storage unit 22. Then, the item value extraction model learning device 100 executes an item value extraction model learning process routine shown in FIG.

  First, in step S100, the value of an item that matches the specified item name list is extracted from the structured data of each article in the article group stored in the article group storage unit 22 as another title.

  Next, in step S102, for each article in the article group stored in the article group storage unit 22, in the text of the article, an alias for the title extracted from the structured data of the article in step S100 is used. The part described is specified, and an annotation indicating that it is another title is given to the specified description part to create pseudo teacher data.

  In step S104, a feature vector is generated for the text of each pseudo-teacher data created in step S102 based on the character notation of each character of the text and the distributed representation of each word. In step S106, a character level tag is attached to the text of each pseudo teacher data according to the annotation.

  In step S108, based on the feature vector generated for each pseudo-teacher data in step S106 and the tag given to the text of each pseudo-teacher data in step S104, a character-level series labeling model is used. The extraction model is learned and output by the output unit 50, and the item value extraction model learning processing routine is terminated.

<Operation of Item Value Extraction Device According to First Embodiment of the Present Invention>
Next, the operation of the item value extraction apparatus 150 according to the first embodiment of the present invention will be described. First, the extraction model learned by the item value extraction model learning device 100 is stored in the extraction model storage unit 74 of the item value extraction device 150. In addition, when an article including a title and text is received by the input unit 60, the item value extraction device 150 executes an item value extraction processing routine shown in FIG.

  In step S110, a feature vector is generated for the text of the article received by the input unit 60 based on the character notation of each character of the text, the distributed representation of each word, and the like.

  In step S112, series labeling is performed based on the feature vector generated in step S110 and the extraction model stored in the extraction model storage unit 74, and a tag is assigned to each character of the article text. Based on the assigned tag, the nickname of the title is extracted, a pair of the title of the article and the nickname is output by the output unit 90, and the item value extraction processing routine is terminated.

  As described above, according to the item value extraction model learning device according to the first embodiment of the present invention, the value of the item that matches the specified item name list is obtained from each of the structured data of the article group. The text extracted as an alias for the title and annotated to indicate that it is an alias for the title is created as pseudo-teacher data, the features extracted from the text of the pseudo-teacher data, and the annotation attached to the text By learning an extraction model based on the above, it is possible to learn an extraction model for extracting a title nickname with high completeness.

  Moreover, according to the item value extraction device according to the first exemplary embodiment of the present invention, based on the extraction model previously learned by the item value extraction model learning device and the feature extracted from the input text. By extracting the title alias from the article text, the title alias can be extracted with high completeness.

  Also, by linking the Infobox of the Wikipedia (R) article and the description of the text, the part of the nickname description in the text is automatically specified, and the model is constructed. As a result, it is possible to acquire and expand the description pattern of the natural sentence without giving the pattern of the natural sentence in advance, and it is possible to construct a nickname extraction model with a high reproduction rate at a low cost. In addition, the alternative names have various values, and new ones appear one after another, and it is particularly effective that they can be automatically extracted.

  In the above-described embodiment, the case where an alternative name of a title is an extraction target has been described as an example. However, the present invention is not limited to this, and other items related to the title other than the alternative name may be the extraction target. .

[Second Embodiment]
Next, a second embodiment will be described. Parts having the same configuration as those of the first embodiment are denoted by the same reference numerals and description thereof is omitted.

The second embodiment is different from the first embodiment in that an extraction model for extracting the value of an item to be extracted is learned from text data that is a set of text and structured data. Yes.
<Configuration of Item Value Extraction Model Learning Device According to Second Embodiment of the Present Invention>

  Next, the configuration of the item value extraction model learning device according to the second embodiment of the present invention will be described. As illustrated in FIG. 14, the item value extraction model learning device 200 according to the second exemplary embodiment of the present invention includes an input unit 10, a calculation unit 220, and an output unit 50.

  The input unit 10 receives a text data group that is a group of text data composed of a set of text and structured data as shown in FIG. The structured data includes a triple of a value (for example, a target word), an item name related to the value, and an item value, and the text includes the target word and the value of the item.

  Further, the input unit 10 extracts structured data and a text portion from each text data of the received text data group.

  Further, the input unit 10 receives an item name list specified as an item name that defines an item (for example, another name) to be extracted regarding the target word.

  The calculation unit 220 includes a text data group storage unit 222, a designated item extraction unit 24, a pseudo teacher data creation unit 26, a feature extraction unit 28, and an extraction model learning unit 30.

  The text data group storage unit 222 stores a text data group received by the input unit 10 and an item name list.

  The designated item extraction unit 24 extracts, as the value of the item to be extracted, the value of the item that matches the designated item name list from each of the structured data included in each text data of the text data group.

  Based on the value of the item to be extracted, which is extracted by the designated item extraction unit 24, the pseudo teacher data creation unit 26 specifies a place where the value of the item to be extracted in the text of the text data is described. Annotation indicating that the value of the item to be extracted is added to the specified description location, and an annotation indicating that the target word is described is added to the location where the target word is described. Created text as pseudo-teacher data.

  The extraction model learning unit 30 performs the first implementation based on the features extracted from the text of the pseudo-teacher data created by the pseudo-teacher data creation unit 26 and the annotation added to the text of the pseudo-teacher data. In the same manner as the above, the extraction model for extracting the value of the item to be extracted is learned from the text of the text data and is output by the output unit 50. In the present embodiment, features related to the target word are also included as features.

<Configuration of Item Value Extraction Device According to Second Embodiment of the Present Invention>
Next, the configuration of the item value extraction apparatus according to the second embodiment of the present invention will be described. As illustrated in FIG. 16, the item value extraction device 250 according to the second exemplary embodiment of the present invention includes an input unit 60, a calculation unit 270, and an output unit 90.

  The input unit 60 accepts text and target words as shown in FIG. The text has no structured data, and the text includes the target word and the value of the item to be extracted.

  The calculation unit 270 includes a feature extraction unit 72, an extraction model storage unit 74, and an item value extraction unit 276.

  The feature extraction unit 72 generates a feature vector similar to the feature extraction unit 28 for the text received by the input unit 60. In the present embodiment, features related to the target word are also included as features.

  The extraction model storage unit 74 stores an extraction model learned by the item value extraction model learning device 200.

  The item value extraction unit 276 performs series labeling based on the feature vector extracted by the feature extraction unit 72 and the extraction model stored in the extraction model storage unit 74, and assigns a tag to each character of the text. The value of the item to be extracted is extracted and output by the output unit 90.

<Operation of Item Value Extraction Model Learning Device According to Second Embodiment of the Present Invention>
Next, the operation of the item value extraction model learning device 200 according to the second embodiment of the present invention will be described. When the input unit 10 receives a text data group consisting of a combination of text and structured data, the structured data and text portion are extracted from each text data of the received text data group and stored in the text data group storage unit 222. .

  When the input unit 10 accepts an item name list designated as an item name that defines an item to be extracted, the item name list is stored in the text data group storage unit 222. Then, the item value extraction model learning device 200 executes an item value extraction model learning process routine shown in FIG. Note that the same processes as those in the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  First, in step S200, the value of the item that matches the specified item name list is extracted from the structured data of each text data of the text data group stored in the text data group storage unit 222. Extract as

  Next, in step S202, for each text data in the text data group stored in the text data group storage unit 222, the text of the text data is extracted from the structured data of the text data in step S200. The location where the value of the item to be extracted is described is specified, and an annotation indicating that the value of the item to be extracted is added to the specified description location, and is included in the structured data of the text data Annotation is given to the portion where the target word is described, and pseudo-teacher data is created.

  In step S104, a feature vector is generated for the text of each pseudo-teacher data created in step S202. In step S106, a character level tag is assigned to the text of each pseudo-teacher data.

  In step S108, the extraction model is learned based on the feature vector generated for each pseudo-teacher data in step S106 and the tag given to the text of each pseudo-teacher data in step S104, and the output unit 50 To complete the item value extraction model learning process routine.

<Operation of Item Value Extraction Device According to Second Embodiment of the Present Invention>
Next, the operation of the item value extraction apparatus 150 according to the second embodiment of the present invention will be described. First, the extraction model learned by the item value extraction model learning device 200 is stored in the extraction model storage unit 74 of the item value extraction device 250. In addition, when the text relating to the target word and the target word are received in the input unit 60, the item value extraction device 250 executes an item value extraction processing routine shown in FIG.

  In step S210, a feature vector is generated for the text received by the input unit 60 based on the character notation of each character of the text, the distributed representation of each word, the target word, and the like.

  In step S212, series labeling is performed based on the feature vector generated in step S210 and the extraction model stored in the extraction model storage unit 74, and a tag is assigned to each character of the text. Based on the tag, the value of the item to be extracted is extracted and output by the output unit 90, and the item value extraction processing routine is terminated.

  As described above, according to the item value extraction model learning device according to the second embodiment of the present invention, the value of the item that matches the specified item name list from each of the structured data of the text data group. Is extracted as the value of the item to be extracted, and the annotated text indicating that it is the value of the extracted item to be extracted is created as pseudo teacher data and extracted from the text of the pseudo teacher data By learning the extraction model based on the feature to be performed and the annotation added to the text, it is possible to learn the extraction model for extracting the value of the item to be extracted with high completeness.

  Moreover, according to the item value extraction device according to the second exemplary embodiment of the present invention, based on the extraction model previously learned by the item value extraction model learning device and the feature extracted from the input text, By extracting the value of the item to be extracted from the text, the value of the item to be extracted can be extracted with high completeness.

  The present invention is not limited to the above-described embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

For example, in the first embodiment and the second embodiment, the case where the BIO tag is assigned at the character level has been described as an example. However, the present invention is not limited to this, and tagging is performed at the word level. You may go. In this case, word division may be performed on the text in advance.
In the second embodiment, the case where the target word is input and the value of the item to be extracted is extracted from the text has been described as an example. However, the present invention is not limited to this. You may make it extract the pair of a target word and the value of an item used as the relationship of the item made into extraction object.

10, 60 Input unit 20, 70, 220, 270 Calculation unit 22 Article group storage unit 24 Designated item extraction unit 26 Pseudo teacher data creation unit 28, 72 Feature extraction unit 30 Extraction model learning unit 50, 90 Output unit 74 Extraction model storage Unit 76 title item value extraction unit 100, 200 item value extraction model learning device 150, 250 item value extraction device 222 text data group storage unit 276 item value extraction unit

Claims (8)

Text data group consisting of a combination of an item name specified as an item name that defines an item to be extracted, and structured data including a combination of a value, an item name related to the value, and an item value, and a text including the value An input unit that accepts
A designated item extraction unit that extracts the value of an item that matches the designated item name from each of the structured data of the text data group, as the value of the item to be extracted;
A pseudo-teacher data creation unit that creates, as pseudo-teacher data, the text with the annotation indicating that it is the value of the item to be extracted extracted by the designated item extraction unit;
The value of the item to be extracted is extracted from the text based on the feature extracted from the text of the pseudo teacher data created by the pseudo teacher data creation unit and the annotation given to the text. An extraction model learning unit for learning an extraction model for
Item value extraction model learning device including An extraction target extracted from each of the structured data included in a text data group including a set of structured data including a set of a value, an item name related to the value, and a value of the item, and a text including the value The feature extracted from the text of the pseudo-teacher data, which is the text to which the annotation indicating that the value of the item to be extracted is added to the value of the item name specified as the item name that defines the item An extraction model storage unit that stores an extraction model for extracting the value of the item to be extracted from the text, learned in advance based on the annotation given to the text;
A feature extraction unit that extracts the features from the input text;
An item value extraction unit that extracts the value of the item to be extracted from the text based on the features extracted by the feature extraction unit and the extraction model;
Item value extraction device including An article group consisting of a title, structured data including a pair of item name and item value, and text including the value, and an item name specified as an item name for defining an item to be extracted with respect to the title are accepted. An input section;
A designated item extraction unit that extracts the value of the item that matches the designated item name from each of the structured data of the article group, as the value of the item to be extracted;
A pseudo-teacher data creation unit that creates, as pseudo-teacher data, the text with the annotation indicating that it is the value of the item to be extracted extracted by the designated item extraction unit;
The value of the item to be extracted is extracted from the text based on the feature extracted from the text of the pseudo teacher data created by the pseudo teacher data creation unit and the annotation given to the text. An extraction model learning unit for learning an extraction model for
Item value extraction model learning device including The item to be extracted is determined for the title extracted from each of the structured data included in the article group including the title, the structured data including the item name and the value of the item, and the text including the value. The feature extracted from the text of the pseudo-teacher data, which is the text to which the annotation indicating that the value is the value of the item to be extracted, with respect to the value of the item name specified as the item name, An extraction model storage unit that stores an extraction model for extracting the value of the item to be extracted from the text, learned in advance based on annotations attached to the text;
A feature extraction unit that extracts the features from the input text;
An item value extraction unit that extracts the value of the item to be extracted from the text based on the features extracted by the feature extraction unit and the extraction model;
Item value extraction device including   The item value extraction model learning device according to claim 1, wherein the item to be extracted is referred to as another name. From the combination of the item name specified as the item name that defines the item to be extracted and the structured data including the value, the item name related to the value and the value of the item, and the text including the value Accept the text data group
The specified item extraction unit extracts the value of the item that matches the specified item name from each of the structured data of the text data group as the value of the item to be extracted,
The pseudo-teacher data creating unit creates, as pseudo-teacher data, the text with the annotation indicating that it is the value of the item to be extracted extracted by the specified item extracting unit,
The extraction model learning unit sets the extraction target from the text based on the feature extracted from the text of the pseudo-teacher data created by the pseudo-teacher data creation unit and the annotation attached to the text. Learning an extraction model for extracting item values Item value extraction model learning method. An extraction target extracted from each of the structured data included in a text data group including a set of structured data including a set of a value, an item name related to the value, and a value of the item, and a text including the value The feature extracted from the text of the pseudo-teacher data, which is the text to which the annotation indicating that the value of the item to be extracted is added to the value of the item name specified as the item name that defines the item And an item value extraction device including an extraction model storage unit that stores an extraction model for extracting the value of the item to be extracted from the text, which has been learned in advance based on the annotation given to the text The item value extraction method in
The feature extraction unit extracts the feature from the input text,
An item value extraction method, wherein an item value extraction unit extracts a value of an item to be extracted from the text based on the feature extracted by the feature extraction unit and the extraction model.   A program for causing a computer to function as each part of the item value extraction model learning device according to claim 1, claim 3, or claim 5, or the item value extraction device according to claim 2 or claim 4.