JP2021135931A - Document classification device, learning method and learning program - Google Patents

Abstract

To provide a document classification device, a learning method and a learning program that can correctly associate a numerical value portion with a feature portion to understand, and that can classify a document.SOLUTION: A document classification device classifies a document which includes an item associated with specific numerical data by means of a learning model including a neural network. The learning model is learned using a label imparted to a document that includes specific numerical data of numerical data extracted by a data extraction unit. A data extraction unit 400 divides text data of a collected document into a plurality of morphemes, identifies a predetermined feature portion and a numerical value portion from the plurality of divided morphemes, and determines a relationship between the feature portion and the numerical value portion. The data extraction unit 400 previously sets a condition to be excluded from the determination of the relationship between the feature portion and the numerical value portion.SELECTED DRAWING: Figure 3

Description

The present invention relates to a document classification device, a learning method, and a learning program.

In recent years, various materials have been developed, and many patent applications and treatises have been published based on the development. The amount of information such as the numerical values of the properties and the numerical range of the composition described in these patent documents and treatises is enormous, and it is impossible for the developer to develop a new material after grasping all the information. be.

Therefore, the developer needs to search the database containing a large number of patent documents and treatises for the information necessary for development. However, when searching for information from a database, even if it is possible to search using a search sentence consisting of natural sentences as shown in Patent Document 1, the numerical values or numerical range described in these documents are simply written in characters. I could only search as.

Further, when classifying a document related to a numerical value or a numerical range of interest from an unclassified document, in Patent Document 2, a plurality of texts are classified in advance according to predetermined items to form a plurality of sub-databases, and the sub-database of interest is selected. Documents are quickly classified by performing text mining processing on multiple texts stored in.

Japanese Unexamined Patent Publication No. 2000-259675 Japanese Unexamined Patent Publication No. 2004-5025

The classification device described in Patent Document 2 simply performs a text mining process on a document, and simply grasps the numerical value or the numerical range described in the document as characters to classify the document.

In order to classify a document based on a numerical value or numerical range, the data must be understood exactly which property or material is related to the numerical value or numerical range described in the document stored in the database. Need to be extracted. However, if the numerical value or numerical range (numerical part) is incorrect if the numerical value or numerical range (numerical part) is incorrect by simply associating the numerical value or numerical range described in the document with the characteristic or material whose position is close to each other (characteristic). There was a problem that the classification device could not correctly grasp the numerical value or the numerical range and classify it because it was associated with the part).

Therefore, an object of the present invention is to solve such a problem, and a document classification device and learning capable of correctly associating and grasping a numerical part and a characteristic part and classifying a document. To provide methods and learning programs.

The document classification device according to one embodiment of the present invention is a document classification device that classifies documents including items related to specific numerical data, and is an input unit in which a plurality of documents are input and a plurality of documents input in the input unit. The document is classified into a document including an item related to specific numerical data and a document not including the item by a learning model including a neural network, and an output unit for outputting information of the document classified by the classification unit. In order to train the training model, a data extraction unit that extracts numerical data from a plurality of documents, a labeling unit that assigns a label to a document containing specific numerical data among the numerical data extracted by the data extraction unit, and a labeling unit. The learning model is trained using a plurality of documents and labels given by the labeling unit. The data extraction unit includes a morphological analysis unit that divides the text data of the collected document into a plurality of morphological elements, and a phrase analysis unit that identifies a predetermined feature part and a numerical part from the plurality of morphological elements divided by the morphological element analysis unit. , A predetermined number of morphological elements indicating a relationship is included between the characteristic part and the numerical value part specified by the phrase analysis unit, and the number of morphological elements contained between the characteristic part and the numerical value part is less than or equal to the predetermined number. In some cases, the related analysis unit that determines that the feature part and the numerical part are related, the exclusion setting unit that presets the conditions to be excluded from the judgment of the relationship between the characteristic part and the numerical part in the related analysis unit, and the related analysis unit. Includes an extraction unit that extracts the feature part and the numerical part determined to be related in (1) as numerical data.

The learning method according to one embodiment of the present invention is to learn a learning model when a computer classifies a plurality of documents into a document including an item related to specific numerical data and a document not containing an item by a learning model including a neural network. This is a learning method that includes a step of extracting numerical data from a plurality of documents and a step of assigning a label to a document containing a specific numerical data among the extracted numerical data. Using the label, let the computer execute the process of training the training model. The steps for extracting numerical data include a step of dividing the text data of the collected document into a plurality of morphological elements, a step of specifying a predetermined feature part and a numerical part from the divided morphology elements, and a specified feature part. When a predetermined number of morphological elements indicating a relationship is included between and the numerical part and the number of morphological elements contained between the characteristic part and the numerical part is less than or equal to the predetermined number, the characteristic part and the numerical part Includes a step of determining that is related, a step of presetting conditions to be excluded from the determination of the relationship between the feature part and the numerical part, and a step of extracting the characteristic part and the numerical part determined to be related as numerical data. ..

The learning program according to one embodiment of the present invention learns a learning model when a computer classifies a plurality of documents into a document containing items related to specific numerical data and a document not containing items by a learning model including a neural network. This is a learning program that includes a data extraction process for extracting numerical data from a plurality of documents and a labeling process for assigning a label to a document containing specific numerical data among the extracted numerical data. Using the document and the assigned label, let the computer execute the process of training the training model. The data extraction process includes a morphological analysis process that divides the text data of the collected document into a plurality of morphological elements, a phrase analysis process that specifies a predetermined feature part and a numerical part from the divided morphological elements, and a specified feature. When a predetermined morphological element indicating a relationship is included between the part and the numerical part and the number of morphological elements contained between the characteristic part and the numerical part is less than or equal to the predetermined number, the characteristic part and the numerical part are included. The related analysis process that determines that is related, the exclusion setting process that presets the conditions to be excluded from the determination of the relationship between the feature part and the numerical part, and the feature part and the numerical part that are determined to be related are extracted as numerical data. Includes extraction processing.

According to the present invention, since the conditions for excluding from the determination of the relationship between the characteristic portion and the numerical portion are set in advance, the numerical portion and the characteristic portion can be correctly associated and grasped, and the document can be classified.

It is the schematic of an example of the document classification apparatus which concerns on this embodiment. It is a block diagram for demonstrating the structure of the document classification apparatus which concerns on this Embodiment. It is a functional block diagram for demonstrating the numerical search process in the document classification apparatus which concerns on this embodiment. It is a functional block diagram for demonstrating the process in the collection part of the document classification apparatus which concerns on this embodiment. It is a functional block diagram for demonstrating the process in the data extraction part of the document classification apparatus which concerns on this embodiment. It is a functional block diagram for demonstrating the process in the morpheme / lexical analysis part of the document classification apparatus which concerns on this embodiment. It is a conceptual diagram for demonstrating an example of processing in the morpheme / lexical analysis part of the document classification apparatus which concerns on this embodiment. It is a conceptual diagram for demonstrating another example of the processing in the morpheme / lexical analysis part of the document classification apparatus which concerns on this embodiment. It is a functional block diagram for demonstrating the process in the relation analysis part of the document classification apparatus which concerns on this embodiment. It is a conceptual diagram for demonstrating an example of processing in the relation analysis part of the document classification apparatus which concerns on this Embodiment. It is a conceptual diagram for demonstrating an example of erroneously determining a related word. It is a conceptual diagram for demonstrating an example of the process which considered the condition of the exclusion setting part in the related analysis part of the document classification apparatus which concerns on this embodiment. It is a conceptual diagram for demonstrating an example of the condition which excludes the association preset in the exclusion setting part. It is a histogram which shows the value of the particle diameter contained in the collected document. It is a graph explaining the importance of a word contained in a document. It is a list which shows the comparison result of 20 documents.

Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. The same or corresponding parts in the drawings are designated by the same reference numerals, and the description thereof will not be repeated.
[Embodiment]
FIG. 1 is a schematic view of an example of a document classification device 100 according to the present embodiment. The document classification device 100 classifies documents including items related to specific numerical data from unclassified documents. When the user inputs a search formula, the document classification device 100 classifies documents including contents related to the search formula from unclassified documents (for example, documents in a specific field collected from a patent document database). , Output the information of the classified document to the user.

Specifically, in the development of MLCC (Multilayer Ceramic Chip Capacitor) and LTCC (Low Temperature Co-fired Ceramics), when developing a material having a particle size of less than 80 nm, a search formula is used in the document classification device 100. Enter "particle size is 1.2 nm to 79 nm", and classify documents related to "particle size 1.2 nm to 79 nm" as "fine-grained documents" from unclassified documents. Output the information of the classified document to the user.

The document classification device 100 uses AI (Artificial Intelligence) to automatically classify documents containing content related to the search formula from unclassified documents, and classifies the documents related to the search formula. It is configured to perform certain actions. Specifically, when the user inputs a search formula into the document classification device 100, the document classification device 100 extracts feature words related to the search formula from the document based on a learning model including a neural network, and is related to the search formula. Performs a classification process that identifies whether it is a document or not.

Here, the "learning model" is labeled as a document containing a neural network and parameters used by the neural network, including feature words related to the search formula, and the document being a document related to the search formula. It is optimized (adjusted) by learning based on information. Specifically, when a document is input, the learning model extracts feature words related to the search formula from the document by a neural network, and classifies whether or not the document is related to the search formula based on the extracted feature words. ..

Then, the learning model does not update the parameters if both match, based on the document classified by itself and the information labeled on the input document (information on whether or not the document is related to the search formula). If they do not match, the parameters are optimized by updating the parameters so that they match. In this way, the learning model is learned by optimizing the parameters using the teacher data including the input document and the information labeled in the document which is the correct answer data.

The process of learning such a learning model is also referred to as a "learning phase". In addition, the learning model optimized by the learning phase is also referred to as a "learned model" in particular. That is, in the present embodiment, the pre-learning learning model and the learned learning model may be collectively referred to as a "learning model", while the learned learning model is distinguished from the pre-learning learning model. Therefore, it may also be called a "trained model".

Generally, when learning a learning model using teacher data, it is necessary for the user to classify documents related to the search formula and documents not related to the search formula in advance and label the documents. The more documents that serve as teacher data, the more accurate the unclassified documents will be classified by the trained model, so many documents need to be labeled.

However, if the teacher data is dozens of documents, the user can label the document, but if the teacher data is hundreds or thousands of documents, the user labels the document. The work of attaching is difficult.

Therefore, in the work of generating teacher data, it is conceivable to reduce the work of the user by automatically labeling the document searched by the search formula. However, it is not possible to generate "fine grain" teacher data by simply searching for documents collected in a document database (for example, a patent document database) using a simple search formula of "particle size of 1.2 nm to 79 nm". ..

That is, when inputting "particle size 1.2 nm to 79 nm" as a search formula and searching for a document using a material having a particle size of 1.2 nm to 79 nm, simply "particle size" is searched from the document collected in the document database. If only a document containing "1.2 nm" and "79 nm" is searched, a document using a material having a particle size of 50 nm will be omitted from the search result. Therefore, it is necessary to extract the characteristic "particle size" and the numerical values "1.2 nm" and "79 nm" from the documents collected in the document database in association with each other, create a database, and then perform a search.

In the document classification device 100 according to the present embodiment, as shown in FIG. 1, as a learning phase, a search formula (specific numerical data) is used for the teacher document (teacher data) collected from the document database of the target patent. A training model is trained by automatically assigning a label to a document containing. Further, in the document classification device 100, as shown in FIG. 1, after learning, the unclassified document is classified into "fine-grained document" and "non-fine-grained document" by using the learning model (learned model) learned in the learning phase. Classify as.

In the learning phase, the document classification device 100 identifies a document containing a search formula (specific numerical data) from a teacher document by a numerical search process and assigns a label (automatic label). Characteristic parts such as characteristics are associated with numerical values such as numerical values or numerical ranges, extracted as numerical data, stored in a database, and used for searching. Hereinafter, a numerical search process including a technique for associating a feature part and a numerical part from a document and extracting them as numerical data will be described.

First, the hardware configuration of the document classification device 100 will be described in detail. FIG. 2 is a block diagram for explaining the configuration of the document classification device 100 according to the present embodiment. The document classification device 100 is executed by the CPU 101 that executes various programs including an operating system (OS: Operating System), a memory unit 112 that temporarily stores data necessary for executing the program in the CPU 101, and the CPU 101. It includes a hard disk unit (HDD: Hard Disk Drive) 110 that stores programs in a non-volatile manner.

Further, the hard disk unit 110 stores in advance a program for realizing processing such as learning processing of a learning model using teacher data, classification processing of unclassified documents using a learned model, and numerical search processing for automatic labeling. Such a program is read from a storage medium 114a such as a CD-ROM (Compact Disk-Read Only Memory) by an optical drive 114 or the like.

The CPU 101 receives a search expression from a user or the like via an input unit 108 including a keyboard or a mouse, and outputs a document or the like classified by executing a program to the display unit 104. The parts are connected to each other via the bus 102. Further, the interface unit 106 is connected to a document database or the like (not shown). The document classification device 100 and the document database may be connected by wire or wirelessly.

Next, the numerical search process executed by the document classification device 100 will be specifically described. FIG. 3 is a functional block diagram for explaining the numerical search process in the document classification device 100 according to the present embodiment. The document classification device 100 includes functions of a collection unit 300, a data extraction unit 400, a database unit 500, and a labeling unit 600 in order to execute a numerical search process. The collection unit 300 collects a target document (a document that serves as teacher data) from the document database. The data extraction unit 400 extracts numerical data in which the characteristic portion and the numerical portion are associated with each other from the collected target document. The database unit 500 stores the numerical data extracted by the data extraction unit 400 in association with the document data of the extracted document. The labeling unit 600 automatically labels the searched document based on the input search formula.

Further, the collecting unit 300 will be described in detail. FIG. 4 is a functional block diagram for explaining the processing of the collection unit 300 in the document classification device 100 according to the present embodiment. The collection unit 300 downloads the full text of a document targeted by the user (for example, a patent document of MLCC containing BaCO3 as a main component) from a document database 301. For example, the user searches the document database for the MLCC document containing "BaCO3", and the collection unit 300 downloads the full text of the search result document. The downloaded document contains data in formats such as ".xml", ".html", and ".csv". In order to process the text data in the data extraction unit 400, the collection unit 300 converts the data of the document downloaded from the document database into the text data.

Specifically, the collecting unit 300 performs the process of converting the data in the ".xml" format into the text data in the ".xml conversion" 302. Further, the collecting unit 300 performs a process of converting the data in the ".htm" format into text data in the ".htm conversion" 303. Further, the collecting unit 300 performs a process of converting the data in the “.csv” format into text data in the “.csv conversion” 304.

The collecting unit 300 compares the text data of the documents converted in ".xml conversion" 302, ".html conversion" 303 and ".csv conversion" 304, and deletes the duplicate document data. (Delete duplicate data 305). The collection unit 300 outputs the text data of the document in which the duplicate data has been deleted 305 as the intermediate file 320.

On the other hand, instead of downloading the document data from the document database, a file containing the data of a plurality of documents is prepared in advance, and the collecting unit 300 can read the file 310 and collect the document data. .. The document data read from the file includes data in formats such as ".docx" and "text". The collecting unit 300 performs the process of converting the data in the ".docx" format into the text data in the ".docx conversion" 311. Since the document data included in the file does not include the duplicate document data, the collecting unit 300 uses the text data of the document converted in ".docx conversion" 311 and the document included in the file. The text data of is output as an intermediate file 320.

Next, a data extraction unit 400 that extracts numerical data in which a feature portion and a numerical value portion are associated with each other from the text data of the document collected by the collection unit 300 will be described. FIG. 5 is a functional block diagram for explaining the processing in the data extraction unit 400 of the document classification device 100 according to the present embodiment. The data extraction unit 400 includes a morpheme / lexical analysis unit 410 that analyzes a feature portion and a numerical value portion from text data of a document, and a related analysis unit 430 that associates the feature portion and the numerical value portion. Further, the data extraction unit 400 includes an exclusion setting unit 450 in which conditions for excluding associations are set in advance in processing by the relation analysis unit 430, and an extraction unit 470 that extracts numerical data associated with the relation analysis unit 430. I have.

The processing in the morpheme / lexical analysis unit 410 will be described in more detail. FIG. 6 is a functional block diagram for explaining the processing in the morpheme / lexical analysis unit 410 of the document classification device 100 according to the present embodiment. The morphological / lexical analysis unit 410 reads the intermediate file 320 output from the collection unit 300, takes in the text data of the document from the intermediate file 320, and processes the text data of the document taken in by the morphological analysis unit 412.

The morphological analysis unit 412 divides the text data of the imported document into columns of morphemes based on predetermined language grammar and dictionary information (for example, information such as part of speech of words), and for each morpheme. Analyze part of speech.

From the text data of the document divided into morpheme columns by the morphological analysis unit 412, a lexical analysis is performed to specify a feature portion and a numerical portion by a push down automaton (PDA). The lexical analysis is not limited to the pushdown automaton, but may be a finite automaton or a more complicated automaton.

Specifically, in the characteristic name attribution PDA413, words corresponding to predetermined characteristic names (for example, dispersant, thickness, resistance value, particle size, etc.) are specified by a pushdown automaton from a column of morphemes, and "feature" is used. Label it as one phrase. In the material name attribution PDA414, a word corresponding to a predetermined material name (for example, neodymium, titanium oxide, barium carbonate, etc.) is specified by a pushdown automaton from a column of morphemes, and labeled as "material". Treat as a lexical.

Further, in the unit attribution PDA415, a word corresponding to a predetermined unit (for example, g, mol, nm, etc.) is specified by a pushdown automaton from a column of morphemes, and labeled as "unit" to form one word. deal. Numerical attribution In PDA416, a phrase corresponding to a predetermined numerical value (for example, Arabic numeral, Chinese numeral, Roman numeral, etc.) is specified by a pushdown automaton from a column of morphological elements, and labeled as "amount" to make one phrase. Treat as. When the word "unit" specified by the unit-attribution PDA415 and the word "amount" specified by the numerically-attribution PDA416 exist in the vicinity (for example, within a predetermined number of morphological elements), Numerical attribution PDA416 treats the word "unit" and the word "amount" as one word "amount".

Further, in the numerical attribution PDA416, when one numerical value is included in a plurality of morphemes, a predetermined numerical range (for example, a range of ± 1% of the numerical value) is given to the numerical value to obtain one. It may be treated as the phrase "amount". As a result, when comparing numerical values, the numerical range having a range of numerical values is compared, and the comparison becomes easy. Further, in the numerical attribution PDA416, when a plurality of numerical values and a predetermined morpheme indicating a range (for example, "~") are included in the plurality of morphemes, a numerical range based on the plurality of numerical values is set. It may be treated as one "range" phrase. Of course, whether it is a numerical value or a numerical range, it may be treated as one "amount" word without distinction.

In the formula attribution PDA417, the words corresponding to the predetermined formulas (for example, x, 1-x, 0 <x ≦ 1, etc.) are specified by the pushdown automaton from the column of morphemes and labeled as “formula”. Treat as one phrase. Chemical formula attribution / composition restoration In PDA418, words corresponding to predetermined chemical formulas and compositions (for example, Ca, TiO2, BaCO3, etc.) are specified by pushdown automata from the morpheme column, and labeled as "material". Treat as one lexical.

If the word "amount" specified by the numerical attribution PDA416, the word "formula" specified by the formula attribution PDA417, and the "material" specified by the chemical formula attribution / composition restoration PDA418 exist in the vicinity, the chemical formula attribution -The composition restoration PDA418 treats the word "amount", the word "formula", and the word "material" as one word "material". Specifically, _{when analyzing the phrase "Ba 1-x} Ca _x TiO ₃ (0 <x ≤ 1)", the phrase "amount" is "3", and the phrase "formula" is "x", "1". The words "-x", "0 <x≤1", and "material" are analyzed as "Ba", "Ca", and "TIO", respectively. Since those words are present in the vicinity, the chemical formula attribution / composition restoration PDA418 finally changes the word "Ba _1-x Ca _x TiO ₃ (0 <x ≦ 1)" to one "material". Treat as a lexical.

The morpheme / lexical analysis unit 410 classifies the words labeled as "material" or "feature" into feature parts, and the words labeled as "amount" or "range" into numerical parts.

Next, a specific example will be shown to explain the processing in the morpheme / lexical analysis unit 410. FIG. 7 is a conceptual diagram for explaining an example of processing in the morpheme / lexical analysis unit 410 of the document classification device 100 according to the present embodiment. When the text data of the imported document includes a sentence 420 stating "1.20 g of TiO2 was used as a main component", the morphological analysis unit 412 first divides the sentence 420 into columns of morphemes.

The morphological analysis unit 412 divides the sentence 420 into columns of morphemes and outputs column data 421 including a plurality of morphemes. The column data 421 includes "TIO", "2", "o", "main component", "to", "", "te", "1", ".", "20", "g". , "Use", "", "ta", "." Contains morphemes.

In the lexical analysis, among the morphemes included in the column data 421, the morpheme 421a of "TIO" is specified as the word "material", and the morpheme 421b of "2" is specified as the word "amount". Further, in the lexical analysis, since the morpheme 421a of "TIO" and the morpheme 421b of "2" are present in the vicinity, "TIO2" is treated as one lexical element 422 of "material".

Further, in the lexical analysis, among the morphemes included in the column data 421, the morpheme 421c of "1", the morpheme 421d of ".", And the morpheme 421e of "20" are specified as the words of "amount", and the morpheme of "g" is specified. The morpheme 421f is specified as the word "unit". Further, in the lexical analysis, the morpheme 421c of "1", the morpheme 421d of ".", The morpheme 421e of "20", and the morpheme 421f of "g" exist in the vicinity, so "1.20 g" is used. Treat as one "amount" phrase 423.

FIG. 8 is a conceptual diagram for explaining another example of processing in the morpheme / lexical analysis unit 410 of the document classification device 100 according to the present embodiment. When the text data of the imported document contains the sentence 420a saying "1.20 g to 1.50 g was used as the main component of TiO2", the morphological analysis unit 412 first divides the sentence 420a into columns of morphemes. do.

The morphological analysis unit 412 divides the sentence 420a into columns of morphemes and outputs column data 421A including a plurality of morphemes. In the column data 421A, "TIO", "2", "o", "main component", "to", "", "te", "1", ".", "20", "g" , "~", "1", ".", "50", "g", "use", "", "ta", "."

In the lexical analysis, among the morphemes included in the column data 421A, the morpheme 421a of "TIO" is specified as the word "material", and the morpheme 421b of "2" is specified as the word "amount". Further, in the lexical analysis, since the morpheme 421a of "TIO" and the morpheme 421b of "2" are present in the vicinity, "TIO2" is treated as one lexical element 422 of "material".

Further, in the lexical analysis, among the morphemes included in the column data 421A, the morpheme 421c of "1", the morpheme 421d of ".", And the morpheme 421e of "20" are specified as the words of "amount", and the morpheme of "g" is specified. The morpheme 421f is specified as the word "unit". Similarly, in the lexical analysis, among the morphemes included in the column data 421A, the morpheme 421h of "1", the morpheme 421i of ".", And the morpheme 421j of "50" are specified as the morpheme of "amount", and "g" is specified. The morpheme 421k of is specified as the word "unit". Further, in the lexical analysis, a predetermined morpheme 421g indicating a range such as "~" is contained between the morpheme 421f of "g" and the morpheme 421h of "1". The morphemes 421k of the morphemes 421c to "g" are treated as one "range" phrase 424 having "1.20 g to 1.50 g". Of course, in the lexical analysis, "amount" and "range" may not be distinguished, and the lexical 424 of "range" may be treated as the lexical of "amount". As the predetermined morpheme indicating the range, it is possible to set a morpheme such as "from" or "to" in addition to "~".

Next, the related analysis unit 430 associates the feature portion selected by the morpheme / lexical analysis unit 410 with the numerical portion. FIG. 9 is a functional block diagram for explaining the processing in the related analysis unit 430 of the document classification device 100 according to the present embodiment.

In the related analysis unit 430 shown in FIG. 9, the wording of the material labeled as “material” and the wording of the numerical value labeled “amount” by the pushdown automaton are specified from the column of morphemes and associated with each other (material + numerical value). ) It has an extracted PDA431. In addition, the related analysis unit 430 identifies and associates the words of the material labeled "material" with the pushdown automaton and the words of the formula labeled "formula" from the morpheme column (material + formula). -Has a composition-restored PDA432. Furthermore, the related analysis unit 430 identifies and associates (characteristic + numerical value) the words of the material labeled as "feature" with the pushdown automaton and the words of the numerical value labeled as "amount" from the column of morphemes. It has a PDA433.

A specific example will be shown to explain the processing in the related analysis unit 430. FIG. 10 is a conceptual diagram for explaining an example of processing in the related analysis unit 430 of the document classification device 100 according to the present embodiment. The related analysis unit 430 includes the lexical 422 of "material" having one "TiO2" and the lexical 423 of "amount" having one "1.20 g" in the data processed by the morpheme / lexical analysis unit 410. If so, it is determined whether or not a predetermined morpheme 425 indicating a relationship such as "o" is included between the words 422 and the word 423. In addition, although "o" is shown as an example of a predetermined morpheme 425 indicating the relationship, "ga" of "material X is 1.0 g", "no" of "1.0 g material X", etc. It is considered that "as", ":", etc. are predetermined morphemes indicating the relationship. Further, the predetermined morpheme 425 indicating the relationship is not limited to a single morpheme such as "o", and may include a combination of a plurality of morphemes. As an example, "converted", which is a combination of two morphemes "converted" and "de", may be included.

When the morpheme 425 is included between the lexical 422 and the lexical 423, the related analysis unit 430 obtains the number of morphemes contained between the lexical 422 and the lexical 423. In the example of FIG. 10, the number of morphemes contained between the words 422 and the word 423 is five.

The related analysis unit 430 assumes that the distance between the obtained lexical 422 and the lexical 423 is close if the number of morphemes contained between the lexical 422 and the lexical 423 is less than or equal to a predetermined number (for example, 10). , Judge that they are related words. In the example of FIG. 10, since the number of morphemes is 5 which is equal to or less than a predetermined number, it is determined that the lexical 422 and the lexical 423 are related. In the example of FIG. 10, in order to extract the relationship between the word "material" 422 and the word "amount" 423, it is processed by the (material + numerical value) extraction PDA431.

Returning to FIG. 9, the extraction unit 470 extracts the words and phrases determined to be related by the association analysis unit 430 as associated numerical data, and outputs the data to the database unit 500. The database unit 500 stores the numerical data extracted by the extraction unit 470 in association with the document data of the document from which the numerical data is extracted (for example, the bibliographic information of the document). In the example of FIG. 10, the extraction unit 470 extracts "1.20 g containing TiO2 as a main component" determined to be related words as one related numerical data 431. The numerical data 431 includes information of material = TiO2 and amount = 1.20 g.

As described above, the related analysis unit 430 determines whether or not the token is related only by whether or not the number of morphemes contained between the tokens is equal to or less than a predetermined number. However, when the association analysis unit 430 determines whether or not a token is related only by the number of morphemes contained between the tokens, the tokens that are completely unrelated are erroneously related depending on the imported document. There is a risk of determining.

FIG. 11 is a conceptual diagram for explaining an example of erroneously determining a related word. The morpheme / lexical analysis unit 410 processes the sentence 420b of "1.20 g of TiO2 is used for 100 g of the main component" contained in the text data of the imported document. As a result of the processing by the morpheme / lexical analysis unit 410, the sentence 420b contains the lexical 422 of "material" having one "TiO2", the lexical 426 of "amount" having one "100g", and "1. It is analyzed that it contains the phrase 423 of "amount" with "20 g" as one.

The association analysis unit 430 includes a predetermined morpheme 425 indicating a relationship such as "o" between the lexical 422 and the lexical 426, and the number of morphemes contained between the lexical 422 and the lexical 426 is 2. Since the number is less than or equal to the predetermined number, it is determined that the words are related.

The extraction unit 470 extracts "TiO2 as a main component 100 g" determined by the related analysis unit 430 as a related phrase as one related numerical data 432. The numerical data 432 includes information of material = TiO2 and amount = 100 g.

However, in the sentence 420b of "Use 1.20 g of TiO2 for 100 g of the main component", it is necessary to relate that "TIO2" uses "1.20 g" from the original meaning. Therefore, the information associated with material = TiO2 and amount = 100g, such as the numerical data 432, is erroneous information (NG).

Therefore, in the document classification device 100 according to the present embodiment, as shown in FIG. 9, when the association analysis unit 430 determines the association of words, the exclusion setting unit 450 for which conditions for excluding the association are set in advance is provided. I have. That is, the relation analysis unit 430 does not determine whether or not the word is related only by the number of morphemes contained between the words, but the morpheme that meets the conditions preset in the exclusion setting unit 450. Is ignored to determine the lexical association.

FIG. 12 is a conceptual diagram for explaining an example of processing in which the condition of the exclusion setting unit 450 is taken into consideration in the related analysis unit 430 of the document classification device 100 according to the present embodiment. The morpheme / lexical analysis unit 410 processes the sentence 420b (see FIG. 11) of "1.20 g of TiO2 is used for 100 g of the main component" contained in the text data of the imported document. As a result of the processing by the morpheme / lexical analysis unit 410, the sentence 420b contains the lexical 422 of "material" having one "TiO2", the lexical 426 of "amount" having one "100g", and "1. It is analyzed that it contains the phrase 423 of "amount" with "20 g" as one.

The related analysis unit 430 determines the presence or absence of a morpheme that meets the conditions preset in the exclusion setting unit 450. As a preset condition, in FIG. 12, it is determined whether or not the sequence of morphemes "to" is included in the sentence 420b. When the column of the morpheme of "against" is included in the sentence 420b, the phrase 426 of "100 g" in front of the column of the morpheme is included in the exclusion target 451. That is, the related analysis unit 430 labels the sequence of morphemes "for 100 g" as "ignore" in consideration of the conditions set in advance in the exclusion setting unit 450, and treats them as one exclusion target 451.

By including the lexical 426 of "100 g" in the exclusion target 451, the related analysis unit 430 determines that a predetermined morpheme 425 showing a relationship such as "o" is included between the lexical 422 and the lexical 423. Will be done. Therefore, the related analysis unit 430 obtains the number of morphemes included between the lexical 422 and the lexical 423. Here, by treating the sequence of morphemes "for 100 g" as one exclusion target 451 the related analysis unit 430 predetermines the number of morphemes included between the words 422 and the word 423 as three. It can be determined that the number is less than or equal to the specified number, and it is determined that the lexical 422 and the lexical 423 are related.

As a result, the extraction unit 470 extracts "1.20 g of TiO2 with respect to 100 g of the main component" determined to be related words as one related numerical data 433. The numerical data 433 includes information of material = TiO2 and amount = 1.20 g. Therefore, the extraction unit 470 can avoid extracting erroneous information in which material = TiO2 and amount = 100g are associated with each other as in the numerical data 432.

The condition for excluding the association preset in the exclusion setting unit 450 will be described in more detail. FIG. 13 is a conceptual diagram for explaining an example of a condition for excluding the association preset in the exclusion setting unit 450. In FIG. 13 (a), when the word “amount” is followed by a row of morphemes “for” as described above, the row of morphemes “for 100 g” is treated as one exclusion target 451. Is shown as a condition.

FIG. 13B shows an example in which the morpheme / lexical analysis unit 410 processes the document “the amount of the dispersant 204 added is 0.3 g”. As a result of processing by the morpheme / lexical analysis unit 410, the sentence includes the phrase "feature" with one "dispersant", the phrase "amount" with one "204", and "0.3g". It is analyzed that it contains the word "amount" with "" as one. Therefore, if no condition is set in the exclusion setting unit 450, the related analysis unit 430 associates the word “204”, which is close to the word “dispersant”, with the word “dispersant”.

However, the phrase "204" is just the code for the phrase "dispersant." Therefore, the exclusion setting unit 450 excludes the word "amount" immediately after the word "feature" without specifying it as the word "amount", and includes the word "feature" and the word "amount" immediately after. Is set as a condition to treat as one "feature" word.

Therefore, the related analysis unit 430 treats the exclusion setting unit 450 as the phrase “feature” 452 with the “dispersant 204” as one in consideration of the condition, and associates it with the phrase “0.3 g”.

FIG. 13C shows an example in which the morpheme / lexical analysis unit 410 processes the document “It is not preferable when the Nd addition amount m exceeds 0.05”. As a result of processing by the morpheme / lexical analysis unit 410, the sentence includes the phrase "material" having one "Nd" and the phrase "amount" having one "0.05". Is analyzed. Therefore, if no condition is set in the exclusion setting unit 450, the related analysis unit 430 associates the word “Nd” with the word “0.05”.

However, if the document contains a word indicating negation such as "not" (a word labeled as "NOT"), the exclusion setting unit 450 exists in the vicinity of the word "NOT". A condition is set to exclude the word "amount" along with the word "NOT". Therefore, the related analysis unit 430 considers the condition of the exclusion setting unit 450, and as shown in FIG. 13 (c), one exclusion target 453 is a sequence of morphemes that "is not preferable when it exceeds 0.05". Treat as.

The condition shown in FIG. 13 is an example of a condition for excluding the association set in the exclusion setting unit 450, and other conditions may be set in the exclusion setting unit 450. In addition, the exclusion setting unit 450 can update the conditions set by the user as appropriate.

Returning to FIG. 3, the database unit 500 stores the numerical data extracted by the data extraction unit 400 from the documents collected by the collection unit 300. Then, the labeling unit 600 searches for the numerical data stored in the database unit 500 based on the search expression input by the user, and among the documents (documents to be teacher data) collected by the collection unit 300, the search expression (documents to be teacher data). Label documents that contain specific numerical data). In the search formula input by the user, it is conceivable to simply search for a document in which "particle size is 1.2 nm to 79 nm" and label the document containing the search formula as "fine particles". In this case, the labeling unit 600 searches the database unit 500 for a document containing information of feature = particle size and range = 1.2 nm to 79 nm, labels the searched document as "fine particles", and outputs the document.

The document classification device 100 classifies unclassified documents into "fine-grained documents" and "non-fine-grained documents" by learning a learning model based on the labeled teacher data generated by the labeling unit 600. Can be done. The learning model trained with the teacher data labeled "fine grain" extracts the feature word from the document labeled "fine grain", and determines whether or not the feature word is included in the unclassified document. It can be classified into "fine-grained documents" and "non-fine-grained documents". Therefore, the unclassified document does not include the term "particle size is 1.2 nm to 79 nm", but expressions related to the range (for example, "small particle size", "fine particle size", etc. Even if there is a document containing (representation of), the document classification device 100 can classify the document as a “fine-grained document”.

Hereinafter, a specific example of classifying unclassified documents using the document classification device 100 according to the present embodiment will be described. FIG. 14 is a histogram showing the value of the particle size contained in the collected documents. The document classification device 100 collects about 1700 MLCC-related patent documents from the document database 200, and uses a numerical search process to obtain characteristic parts such as chemical formulas, materials, and properties, and numerical parts such as numerical values or numerical ranges. It is associated and extracted as numerical data and stored in the database. In FIG. 14, the results of searching for the minimum value and the maximum value of the particle size using the database are visualized in a histogram with the particle size on the horizontal axis and the number of documents on the vertical axis.

The A region in the histogram shows documents in which numerical values having a particle size of 1.2 nm to 79 nm are described, and the number of documents is 59. The B region in the histogram shows a document in which a numerical value having a particle size of 200 nm to 270 μm is described, and the number of documents is 368. In the teacher data, the document contained in the A area is labeled as "fine grain", and the document contained in the B area is labeled as "not fine grained".

By using the histogram shown in FIG. 14, it was confirmed that the region A labeled as "fine particles" was visually sufficiently separated from the region B labeled as "not fine particles". Can be decided.

The document classification device 100 machine-learns the learning model using the teacher data determined above. For machine learning, for example, the Multinomial Naive Bayes method, which is a general text classification method, is used. Of course, for machine learning, a method other than the Multinomial Naive Bayes method may be used. Further, in the process of training the learning model, the teacher data is divided into the training data and the test data at a ratio of 7: 3. When splitting teacher data into training and testing, the proportion of documents labeled "fine" and "non-fine" should be the same as before splitting. Documents are randomly selected and divided.

After learning with the teacher data for learning, the learning model confirms the accuracy (correct answer rate) using the remaining teacher data for the test. In the learning model trained with the teacher data generated from the patent documents related to MLCC, the accuracy of correctly determining the teacher data for learning as "fine-grained document" or "non-fine-grained document" was 0.938. On the other hand, when the teacher data for the test is classified as "fine-grained document" or "non-fine-grained document" by this learning model, the accuracy is 0.912, and the learning with the teacher data for learning is normal. I was able to confirm that it was working.

In addition, we verified important words among the words included in the learned learning model. FIG. 15 is a graph illustrating the importance of the words contained in the document. In FIG. 15, the horizontal axis represents the number of documents and the vertical axis represents the tfidf value, and the importance of the three words “BRT”, “interface”, and “Ca” included in the learning model is shown. Here, the tfidf value is one of the methods for evaluating the importance of words contained in a document. Words appearing in many documents have low importance, and words appearing only in a specific document have high importance. .. Further, "Ca" indicates calcium, and "BRT" indicates a powder in which a rare earth element is solid-dissolved in barium titanate, for example, BaO-TiO2-ReO3 / 2 ceramic.

From the results shown in FIG. 15, the words included in the learning model include words that indicate the purpose and advantage of atomization such as "interface", words that indicate additives such as "Ca", and words that indicate materials such as "BRT". It can be seen that such as is a highly important word. Therefore, it can be seen that the learned learning model is appropriate as a model for classifying whether the learned learning model is a "fine-grained document" or a "non-fine-grained document".

Next, 20 documents are randomly selected from patent documents related to MLCC, LTCC, etc. (a total of about 6000 documents, unclassified documents that do not include patent documents used as teacher data), and a document classification device for a trained model. At 100, a "fine-grained document" or a "non-fine-grained document" is classified. In addition, since the correct answer label of "fine grain" or "not fine grain" is attached to these 20 documents in advance, the document classification device 100 of the trained model was able to correctly classify the documents. It is possible to compare whether or not.

FIG. 16 is a list showing the comparison results of 20 documents. In the comparison results shown in FIG. 16, documents that could not be searched by the search formula "particle size of 1.2 nm to 79 nm" could be classified as "fine-grained documents". Specifically, in the document with the number "CCCCC02", although there is a description of "size of 30-50 nm" in the document, there is no description of "particle size", so the search formula "particle size is 1.2 nm to 79 nm". It is a document that has not been searched completely. For humans, from the description that the size of nickel is 30-50 nm, it can be classified as "fine grain" as it falls within the range of "particle size of 1.2 nm to 79 nm". On the other hand, the document classification device 100 of the trained model was able to correctly classify the document of the number "CCCCC02" which the person classified as "fine grain" as "fine grain".

Further, in the document of the number "PPPPP01", although there is a description of "10 to 100 nm BaTiO3" in the document, there is a description including a range larger than "79 nm", so that the search formula "particle size is 1.2 nm to 79 nm" is used. This is a document that has not been completely searched. For humans, from the description that the size of BaTiO3 is 10 to 100 nm, it can be classified as "fine grain" as it is almost included in the range of "particle size of 1.2 nm to 79 nm". On the other hand, the document classification device 100 of the trained model was able to correctly classify the document of the number "PPPPP01" which the person classified as "fine grain" as "fine grain". On the contrary, the document classification device 100 of the trained model did not erroneously classify documents whose fields and particle sizes do not match.

As described above, the document classification device 100 according to the present embodiment is a document classification device that classifies documents including items related to specific numerical data. The document classification device 100 divides the input unit into which a plurality of documents are input and the plurality of documents input by the input unit into a document including an item related to specific numerical data and a document not including the item by a learning model including a neural network. It includes a classification unit for classifying and an output unit for outputting information on documents classified by the classification unit. The document classification device 100 labels a data extraction unit 400 that extracts numerical data from a plurality of documents and a document that includes specific numerical data among the numerical data extracted by the data extraction unit 400 in order to train a learning model. A learning model is trained by using a plurality of documents and labels given by the labeling unit. The data extraction unit 400 includes a morpheme / lexical analysis unit 410, a related analysis unit 430, an exclusion setting unit 450, and an extraction unit 470. The morpheme / lexical analysis unit 410 divides the text data of the collected document into a plurality of morphemes, and identifies a predetermined feature portion and a numerical portion from the divided plurality of morphemes. The relation analysis unit 430 includes a predetermined morpheme indicating a relationship between the specified feature part and the numerical part, and the number of morphemes contained between the feature part and the numerical part is equal to or less than the predetermined number. In the case of, it is determined that the feature part and the numerical part are related. The exclusion setting unit 450 sets in advance the condition to be excluded from the determination of the relationship between the feature portion and the numerical portion in the relation analysis unit 430. The extraction unit 470 extracts the feature portion and the numerical portion determined to be related by the related analysis unit 430 as numerical data.

As a result, in the document classification device 100 according to the present embodiment, the conditions for excluding from the determination of the relationship between the characteristic portion and the numerical portion are set in advance, so that the numerical portion and the characteristic portion are correctly associated and extracted. And can label documents that contain specific numerical data. Further, in the document classification device 100 according to the present embodiment, a document containing specific numerical data is labeled, and a learning model is trained with the labeled document to obtain a specific numerical value from an unclassified document. Documents related to the data can be categorized.

Further, the related analysis unit 430 may compress the conditions preset by the exclusion setting unit 450 as one morpheme and count the number of morphemes included between the feature portion and the numerical value portion. As a result, the document classification device 100 determines that the number of morphemes contained between the feature portion and the numerical portion is equal to or less than a predetermined number, and makes it easier to associate the feature portion with the numerical portion.

Further, when the morpheme / lexical analysis unit 410 includes one numerical value among the plurality of divided morphemes, the morpheme / lexical analysis unit 410 may assign a predetermined numerical range to the numerical value and specify it as a numerical part. As a result, when the document classification device 100 compares the numerical values with each other, the numerical values have a range of numerical values, and the comparison becomes easy.

Further, when the morpheme / lexical analysis unit 410 includes a plurality of numerical values and a predetermined morpheme indicating a range in the divided plurality of morphemes, the morpheme / lexical analysis unit 410 sets a numerical range based on the plurality of numerical values as a numerical part. May be specified. Thereby, the document classification device 100 can specify the numerical range as one numerical part.

Further, the morpheme / lexical analysis unit 410 may specify the unit and the numerical value as one numerical value portion when the unit is included in the plurality of divided morphemes. As a result, the document classification device 100 can specify the numerical portion including the unit, and it becomes easy to compare the numerical portion in consideration of the unit.

Further, the morpheme / lexical analysis unit 410 may specify any of a predetermined characteristic name, material name, chemical formula and composition as a characteristic portion from the plurality of divided morphemes. The morpheme / lexical analysis unit 410 may include a numerical range or a mathematical formula in either the chemical formula or the composition to specify the characteristic portion. As a result, the document classification device 100 can associate various types of feature portions with numerical portions.

This is a learning method for learning a learning model when a computer classifies a plurality of documents into a document containing items related to specific numerical data and a document not containing items by a learning model including a neural network. The learning method includes a step of extracting numerical data from a plurality of documents and a step of assigning a label to a document containing specific numerical data among the extracted numerical data. The steps of having a computer execute a process of learning a learning model using a plurality of documents and assigned labels and extracting numerical data include a step of dividing the text data of the collected document into a plurality of morphemes and a step of dividing the data. It includes a step of identifying a predetermined feature part and a numerical part from a plurality of morphemes. The step of extracting the numerical data includes a predetermined morpheme indicating the relationship between the specified feature part and the numerical part, and the number of morphemes contained between the feature part and the numerical part is predetermined. In the case of a number or less, a step of determining that the feature part and the numerical part are related is included. The step of extracting the numerical data includes a step of presetting a condition to be excluded from the determination of the relationship between the characteristic part and the numerical part, and a step of extracting the characteristic part and the numerical part determined to be related as numerical data. ..

This is a learning program in which a computer trains a learning model when a computer classifies a plurality of documents into a document containing an item related to specific numerical data and a document not containing the item by a learning model including a neural network. The learning program includes a data extraction process for extracting numerical data from a plurality of documents and a labeling process for assigning a label to a document containing specific numerical data among the extracted numerical data. Using a plurality of documents and assigned labels, a computer is made to execute a process of training a training model, and a data extraction process is a predetermined morpheme showing a relationship between a specified feature part and a numerical part. Includes, and includes a related analysis process for determining that the feature part and the numerical part are related when the number of morphemes contained between the feature part and the numerical part is less than or equal to a predetermined number. The data extraction process includes an exclusion setting process in which conditions for excluding from the determination of the relationship between the feature part and the numerical part are set in advance, and an extraction process for extracting the characteristic part and the numerical part determined to be related as numerical data. ..

[Modification example]
(1) In the above-described embodiment, an example in which a document is collected from a patent document database and numerical data in which a characteristic part and a numerical part are associated with each other is extracted from the collected document has been described. However, the present invention is not limited to this, and even in a configuration in which a document is collected from a document database other than the patent document database (for example, an academic paper database) and numerical data in which a characteristic part and a numerical part are associated with each other is extracted from the document. good. Further, a configuration may be configured in which documents are collected from a plurality of types of databases (for example, a patent document database and an academic paper database), and numerical data in which a characteristic portion and a numerical portion are associated with each other is extracted from the document.

(2) The various processes described above are supposed to be realized by one CPU, but are not limited to this. These various processes may be distributed and realized by a plurality of CPUs. The programs executed by these CPUs can execute the above-mentioned various processes by reading one or more instructions from at least one tangible readable medium.

Such media take the form of magnetic media (eg, hard disks), optical media (eg, compact discs (CDs, DVDs), volatile memories, memories of any type of non-volatile memory, and the like. It is not limited to.

Volatile memory may include DRAM (Dynamic Random Access Memory) and SRAM (Static Random Access Memory). The non-volatile memory may include ROM, NVRAM.

It should be considered that the embodiments disclosed this time are exemplary in all respects and not restrictive. The scope of the present invention is shown by the scope of claims, not the above description, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.

100 document classification device, 200 document database, 300 collection unit, 320 intermediate file, 400 data extraction unit, 410 morphological / lexical analysis unit, 412 morphological analysis unit, 430 related analysis unit, 450 exclusion setting unit, 470 extraction unit, 500 database Department, 600 Labeling part.

Claims (9)

A document classification device that classifies documents containing items related to specific numerical data.
Input section where multiple documents are input and
A classification unit that classifies a plurality of documents input by the input unit into a document that includes an item related to the specific numerical data and a document that does not include an item related to the specific numerical data by a learning model including a neural network.
It is provided with an output unit that outputs information on documents classified by the classification unit.
In order to train the learning model,
A data extractor that extracts numerical data from multiple documents,
A labeling unit for assigning a label to a document containing the specific numerical data among the numerical data extracted by the data extraction unit is provided.
The learning model is trained using a plurality of documents and the labels given by the labeling unit, and the learning model is trained.
The data extraction unit
A morphological analysis unit that divides the text data of the collected document into multiple morphemes,
A lexical analysis unit that identifies predetermined feature parts and numerical parts from a plurality of morphemes divided by the morphological analysis unit, and
A predetermined morpheme indicating a relationship is included between the characteristic portion and the numerical value portion specified by the lexical analysis unit, and the number of morphemes included between the characteristic portion and the numerical value portion is predetermined. A related analysis unit that determines that the characteristic part and the numerical value part are related when the number is less than or equal to
In the related analysis unit, an exclusion setting unit that presets conditions to be excluded from the determination of the relationship between the characteristic portion and the numerical value portion, and the exclusion setting unit.
A document classification device including an extraction unit that extracts the feature portion and the numerical value portion determined to be related by the related analysis unit as numerical data. The first aspect of the present invention, wherein the related analysis unit compresses the conditions preset by the exclusion setting unit as one morpheme and counts the number of morphemes contained between the feature portion and the numerical value portion. Document classification device. When one numerical value is included in a plurality of morphemes divided by the morphological analysis unit, the lexical analysis unit assigns a predetermined numerical range to the numerical value and identifies it as the numerical value portion. The document classification device according to claim 1 or 2. When a plurality of numerical values and predetermined morphemes indicating a range are included in the plurality of morphemes divided by the morphological analysis unit, the lexical analysis unit obtains a numerical range based on the plurality of numerical values. The document classification device according to any one of claims 1 to 3, which is specified as a numerical part. The lexical analysis unit identifies the unit and the numerical value as one of the numerical values when the unit is included in the plurality of morphemes divided by the morphological analysis unit. The document classification device according to any one item. The lexical analysis unit identifies any of a predetermined characteristic name, material name, chemical formula and composition as the characteristic portion from a plurality of morphemes divided by the morphological analysis unit, claim 1 to 5. The document classification device according to any one of the items. The document classification device according to claim 6, wherein the lexical analysis unit includes a numerical range or a mathematical expression in either the chemical formula or the composition to identify the characteristic portion. A learning method for learning a learning model when a computer classifies a plurality of documents into a document containing an item related to specific numerical data and a document not containing an item by a learning model including a neural network.
Steps to extract numerical data from multiple documents,
Including a step of assigning a label to a document containing the specific numerical data among the extracted numerical data.
Using a plurality of documents and assigned labels, a computer is made to execute a process of learning the learning model.
The step to extract numerical data is
Steps to divide the text data of the collected document into multiple morphemes,
A step to identify a predetermined feature part and a numerical part from a plurality of divided morphemes, and
When a predetermined morpheme indicating a relationship is included between the specified feature portion and the numerical value portion, and the number of morphemes contained between the feature portion and the numerical value portion is equal to or less than a predetermined number. In addition, a step of determining that the feature portion and the numerical value portion are related,
A step of presetting conditions to be excluded from the determination of the relationship between the feature portion and the numerical portion, and
A learning method including a step of extracting the characteristic portion determined to be related and the numerical portion as numerical data. A learning program for training a plurality of documents when a computer classifies a plurality of documents into a document containing an item related to specific numerical data and a document not containing an item by a learning model including a neural network.
Data extraction processing that extracts numerical data from multiple documents,
Includes a labeling process for assigning a label to a document containing the specific numerical data among the extracted numerical data.
Using a plurality of documents and assigned labels, a computer is made to execute a process of learning the learning model.
The data extraction process
Morphological analysis processing that divides the text data of the collected document into multiple morphemes,
Lexical analysis processing that identifies predetermined feature parts and numerical parts from a plurality of divided morphemes,
When a predetermined morpheme indicating a relationship is included between the specified feature portion and the numerical value portion, and the number of morphemes contained between the feature portion and the numerical value portion is equal to or less than a predetermined number. In addition, the related analysis process for determining that the feature portion and the numerical value portion are related,
Exclusion setting processing that presets conditions to be excluded from the determination of the relationship between the feature portion and the numerical portion, and
A learning program including an extraction process for extracting the characteristic portion determined to be related and the numerical portion as numerical data.

Images