JP7773010B2 - Method, device, system, program and computer-readable storage medium - Google Patents

Description

The present invention relates to a method, an apparatus, a system, a program, and a computer-readable storage medium.

Knowledge graphs are used as knowledge networks that systematically link various types of knowledge and represent them in a graph structure. To create knowledge graphs, it is effective to extract information from text data written in natural language in the form of a machine-readable data structure. In this case, for example, a technique called open information extraction (IE) is sometimes used.

Patent document 1 describes the use of transfer learning from a high-resource language to extract information in the form of machine-readable data structures from a low-resource language.

Non-Patent Document 1 describes a method for iteratively extracting triples (subject, predicate, object) from natural language sentences, by first extracting easy slots, then using the easy slots as conditions to extract difficult slots, and so on.

Non-Patent Document 2 describes a technology for automatically converting English text in supervised OpenIE to train OpenIE systems for languages other than English. Specifically, it describes the Alignment-Augmented Consistent Translation (AACTRANS) model, which translates English sentences and their corresponding extracted sentences while maintaining consistency with each other.

Japanese Patent Application Publication No. 2022-151838

Kotnis, Bhushan, et al. “MILIE: Modular & Iterative Multilingual Open Information Extraction,” Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1:Long Papers).2022. Kolluru, K., Mohammed, M., Mittal, S., & Chakrabarti, S. (2022,May). “Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction,” In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)(pp.2502-2517).

It is necessary to efficiently create a Japanese knowledge graph.

The present invention aims to provide a method, device, system, program, and computer-readable storage medium for efficiently creating a Japanese knowledge graph.

The method, device, system, program, and computer-readable storage medium of the present invention are intended to achieve the above-mentioned objectives and have the following configuration.

(1) A method for learning according to one aspect of the present invention includes the steps of: acquiring a first character string written in a first language, the first character string including a plurality of first nouns; acquiring a trained model indicating a correspondence between the first character string and at least a first subject and a first predicate included in the first character string; translating the first character string into Japanese to generate a second character string, the second character string including a plurality of second nouns; and translating the first character string into Japanese to generate a second character string, the second character string including a plurality of second nouns. obtaining a first converted first string by changing the structure of the first string to correspond to the order of the first nouns in the first converted first string; obtaining a second converted first string by adding one or more Japanese particles before or after each of the first nouns included in the first converted first string based on the correspondence between the first nouns and the second nouns; and generating an updated trained model indicating the correspondence between the first string and at least a second subject and a second predicate included in the second string by performing transfer learning based on the second converted first string and the trained model.

(2) In the aspect (1) above, the step of generating the updated trained model by performing transfer learning based on the second converted first character string and the trained model corresponds to discontinuous and iterative extraction and learning, and in the discontinuous and iterative extraction, one or more keywords other than nouns and verbs are included between the first subject and the first predicate, and includes a step of iteratively extracting particles corresponding to each of the plurality of second nouns.

(3) In the above aspect (1), the one or more Japanese particles include a case particle.

(4) In the above aspect (1), the method further includes the steps of acquiring a third character string written in Japanese, and extracting a plurality of third nouns and a plurality of third particles contained in the third character string based on the third character string and the updated trained model, wherein the plurality of third particles include particles corresponding to each of the plurality of third nouns.

(5) In the above aspect (4), the plurality of third particles include a case particle.

(6) An apparatus according to one aspect of the present invention includes an acquisition unit and a processing unit. The acquisition unit is configured to acquire a first character string written in a first language and acquire a trained model indicating a correspondence between the first character string and at least a first subject and a first predicate included in the first character string, the first character string including a plurality of first nouns, the processing unit is configured to translate the first character string into Japanese to generate a second character string, the second character string including a plurality of second nouns, and the acquisition unit generates the first character string by changing the structure of the first character string so that the order of the plurality of first nouns corresponds to the order of the plurality of second nouns. and acquiring a converted first character string, adding one or more Japanese particles before or after each of the plurality of first nouns included in the converted first character string based on the correspondence between the plurality of first nouns and the plurality of second nouns to acquire a second converted first character string, and the processing unit is configured to perform transfer learning based on the second converted first character string and the trained model, thereby generating an updated trained model indicating the correspondence between the first character string and at least a second subject and a second predicate included in the second character string.

(7) In the aspect of (6) above, the acquisition unit is configured to acquire a third character string written in Japanese, and the processing unit is configured to extract a plurality of third nouns and a plurality of third particles included in the third character string based on the third character string and the updated trained model, and the plurality of third particles include particles corresponding to each of the plurality of third nouns.

(8) In the above aspect (6), the one or more Japanese particles include a case particle.

(9) An apparatus according to one aspect of the present invention includes an acquisition unit and a processing unit. The acquisition unit is configured to acquire a third character string written in Japanese and a trained model. The processing unit is configured to extract a plurality of third nouns and a plurality of third particles included in the third character string based on the third character string and the trained model. The trained model indicates a correspondence between a first character string written in a first language and at least a first subject and a first predicate included in the first character string. The plurality of third particles include particles corresponding to each of the plurality of third nouns.

(10) A system for learning and estimation according to one aspect of the present invention includes the device described in (6) above and the device described in (9) above.

(11) A program according to one aspect of the present invention, when executed on a computer, executes the method described in any one of (1) to (5) above.

(12) A computer-readable storage medium according to one aspect of the present invention stores a program that, when executed on a computer, causes the computer to execute the method described in any one of (1) to (5) above.

(1) to (12) provide a method, device, system, program, and computer-readable storage medium that can efficiently create a Japanese knowledge graph.

1 is a diagram illustrating an example of the configuration of an estimation system S according to an embodiment of the present invention. FIG. 1 is a diagram illustrating an overview of learning and estimation according to an embodiment of the present invention. 1 is a diagram illustrating an example of a functional configuration of a learning device 10 according to an embodiment of the present invention. FIG. 2 is a diagram illustrating an example of a functional configuration of an estimation device 30 according to an embodiment of the present invention. 10 is a flowchart of a learning process according to an embodiment of the present invention. 10 is a flowchart of an estimation process according to an embodiment of the present invention. FIG. 2 is an explanatory diagram illustrating the hardware configuration of each device according to the embodiment.

(First embodiment)
Hereinafter, with reference to the drawings, embodiments of the method, device, system, program, and computer-readable storage medium of the present invention will be described. Note that the following embodiments do not limit the scope of the invention as claimed, and not all combinations of features described in the embodiments are necessarily essential to the invention. Two or more of the features described in the embodiments may be arbitrarily combined. Furthermore, the same reference numerals are used for the same or similar components, and redundant descriptions will be omitted.

Figure 1 is a diagram showing an example configuration of an estimation system S according to an embodiment of the present invention.

The estimation system S includes a learning device 10 and an estimation device 30. The learning device 10 is connected to the estimation device 30 via a network NW. The learning device 10 generates a trained model through machine learning. The trained model indicates the correspondence between the Japanese text to be estimated and a Japanese knowledge graph. The learning device 10 provides the generated trained model to the estimation device 30. The estimation device 30 acquires Japanese text. The estimation device 30 estimates and outputs a Japanese knowledge graph based on the acquired Japanese text and the trained model. Details of the operation of the learning device 10 and the estimation device 30 will be explained separately.

Figure 2 is a diagram illustrating an overview of learning and estimation according to an embodiment of the present invention. First, the learning device 10 acquires English text and a Japanese translation of the English text. Based on the acquired English text, the learning device 10 may generate a corresponding Japanese translation of the English text by machine translation.

The learning device 10 performs text alignment processing. This processing rearranges the word order of English text to match the word order of the corresponding Japanese text. For example, if the English sentence is "He loves a pizza," the corresponding Japanese translation would be "He loves pizza." As a result of the text alignment processing, the word order contained in the English text is changed, and the English text becomes "He a pizza loves." This "Japanese-style English text" may also be referred to as "Japonized English text."

The learning device 10 adds a corresponding particle before the noun contained in the aligned Japanese-style English text. As a result of adding the particle, the Japanese-style English text becomes "(ha) He (ga) pizza loves." The Japanese-style English text could also be "(ha) He (wo) pizza loves" or "(ga) He (wo) pizza loves." The Japanese-style English text with the particle added may be referred to as "marked English text." By adding particles before the nouns, the learning device 10 can accurately and efficiently identify and learn whether the corresponding noun following each particle is a subject, predicate, object, etc.

To take a more complex example, the original English text might be, "He was awarded the Queen's Police Medal in the 1957 New Year Honours." The corresponding machine translation of Japanese might be, for example, "He was awarded the Queen's Police Medal in the 1957 New Year Honours." The Japanese-style English text might be, "He was 1957 the New in Queen's police Medal awarded." The marked English text might be, "(が (ga)) He was (で (de)) 1957 the New in (を (wo)) Queen's Police Medal awarded."

To give yet another example, the original English text could be, "Blagoja 'Billy' Celeski is an Australian footballer who plays as a midfielder for the Newcastle Jets." The corresponding machine translation of Japanese could be, "Blagoja 'Billy' Celeski is an Australian footballer who plays as a midfielder for the Newcastle Jets." The Japanese-style English text could be, "Blagoja 'Billy' Celeski is Newcastle Jets a plays Australian an footballer." and the marked English text could be, "(が (ga)) Blagoja 'Billy' Celeski for (で (de)) the Newcastle Jets as a midfielder an Australian (の (no)) footballer who plays is."

The learning device 10 generates a trained model through machine learning based on discontinuous iterative Open IE extraction and learning. This machine learning uses the correspondence between English text and an English knowledge graph as training data and uses transfer learning to learn the correspondence between marked English text and an English knowledge graph. Transfer learning is a method of learning only the data of a newly added model while leaving the data of an existing trained model untouched. Because corresponding particles are added before the nouns contained in the marked English text, it is possible to estimate whether each noun in the marked English text functions as a subject, predicate, etc. Therefore, the estimation device 30 can infer the meaning of the marked English text into appropriate Japanese based on the trained model and estimate an appropriate knowledge graph for Japanese. Note that the training data may also include similar correspondences for other languages other than English, such as French, Italian, and Spanish.

Traditionally, the OpenIE system has considered subjects, predicates, and objects to be consecutive. However, word order in Japanese is not always like this, and can be discontinuous. There may be predicates where the first part of an English text is at the beginning of the sentence and other parts are at the end of the sentence. The same is true for objects and subjects. For example, Japanese may contain the expression, "The girl, to her surprise, heard a strange sound." In this case, "the girl" and "heard" are discontinuous. In discontinuous repetition OpenIE, subjects, objects, and predicates can be split within a sentence. Here, learning includes two modes: continuous extraction and repetitive extraction.

i) Discontinuous extraction separates the subject, object, and predicate and spreads them across the sentence, a key difference from existing OpenIE systems, which assume these components are close together.
ii) Iterative learning is performed by predicting aspects one after another, and each time, the predicted aspects are added to the annotation. Aspects include subjects, objects, and predicates. Iterative learning is achieved by conditioning the prediction of missing aspects (subjects, objects, or predicates) on those that have already been predicted.

The estimation device 30 acquires Japanese text to be estimated and also acquires a trained model based on discontinuously repeated OpenIE extraction. This trained model is generated and provided by the learning device 10. The estimation device 30 estimates and outputs a Japanese knowledge graph through machine learning processing based on the acquired Japanese text and the trained model. The estimation device 30 may then perform link prediction. Link prediction refers to a predetermined prediction process based on the knowledge graph.

Examples of link prediction include performing automated decision-making in public safety and law enforcement. In a public safety application, link prediction could optimize the deployment of drones to inaccessible areas due to floods, earthquakes, or nuclear disasters. These drones could perform life-saving tasks, such as distributing medicine or water or searching for survivors.

The estimation device 30 may use Japanese text, such as social media or reports created by public institutions, as input, to estimate a knowledge graph by performing an estimation process based on a trained model, and may further predict the amount and type of drone assistance. Specifically, the estimation device 30 may estimate locations where assistance will be provided and an ordered list of the assistance needed at each location.

As another example of an application to public safety, the estimation device 30 may estimate effective countermeasures for hate speech in Japanese text. In this case, the estimation device 30 acquires Japanese text published in newspapers, portal sites, social media, etc., and estimates a knowledge graph by performing an estimation process based on a trained model. The estimation device 30 may then analyze the acquired text and compare it with the knowledge graph to estimate misinformation and hate speech. The estimation device 30 may then estimate and output an ordered list of news, posts, and threads to be deleted and users to be sued.

As an example of an application to public safety, the estimation device 30 may predict an increase in criminal activity by analyzing reports from police, social workers, and the like. In this case, the estimation device 30 acquires Japanese text from police stations and social workers, and the like. The acquired Japanese text may include, for example, a list of people with bad behavior, such as those with criminal records, along with information about their addresses, schools, or places of employment. The estimation device 30 then infers a knowledge graph by performing an estimation process based on the trained model. The estimation device 30 then analyzes the Japanese text. The estimation device 30 may analyze the acquired Japanese text and compare it with the knowledge graph to infer information about crimes. The estimation device 30 may then estimate the location and type of predicted crime, along with the degree of risk of occurrence. This may lead to the automatic dispatch of crime prevention teams organized by police officers to locations where criminal activity is likely to occur.

These estimation processes performed by the estimation device 30 enable authorities to receive automated decision-making support regarding measures necessary for public safety and/or maintaining order.

FIG. 3 is a diagram illustrating an example of the functional configuration of the learning device 10 according to an embodiment of the present invention.
includes an acquisition unit 110, a processing unit 120, a communication unit 130, a learning data storage unit 140, and an output unit 150. The learning data storage unit 140 includes a marked text storage unit 141, a knowledge graph storage unit 142, a learning result storage unit 143, and a learning result storage unit 144.

The acquisition unit 110 externally acquires a trained model to be trained. The trained model is training data and includes English text and a knowledge graph corresponding to the English text. The training data indicates the correspondence between the English text and the knowledge graph. The English text may be referred to as a first string. The first string includes a first subject and a first predicate. Note that English is an example of a first language. In other words, the first language may be any language other than Japanese, such as French, Italian, Spanish, or Chinese. The English text includes multiple nouns. The multiple nouns included in the English text are an example of multiple first nouns. The knowledge graph also includes these multiple nouns. Therefore, the training data indicates the correspondence between the first string and at least the first subject and first predicate included in the first string.

The processing unit 120 processes and processes information and data acquired from outside. The communication unit 130 transmits and receives information and data to and from external devices. The marked text storage unit 141 stores marked English text. The knowledge graph storage unit 142 stores an English knowledge graph. The learning result storage unit 143 stores trained models. The learning result storage unit 144 stores training results. The training results include updated trained models. Note that the marked text storage unit 141 may store marked text in another first language instead of marked English text. The knowledge graph storage unit 142 may store a knowledge graph in another first language instead of an English knowledge graph. The output unit 150 outputs information and data acquired from outside, as well as trained models, to the outside.

Figure 4 is a diagram showing an example of the functional configuration of the estimation device 30 in an embodiment of the present invention. The estimation device 30 includes an acquisition unit 310, a processing unit 320, a communication unit 330, an execution data storage unit 340, and an output unit 350. The execution data storage unit 340 includes a Japanese text storage unit 341, an estimation result storage unit 342, and a trained model storage unit 343.

The acquisition unit 310 acquires the Japanese text to be estimated and a trained model. The acquired trained model is an "updated trained model" generated by the learning device 10 and is generated based on discontinuous repetitive OpenIE extraction learning.

The processing unit 320 performs an estimation process based on the Japanese text to be estimated and the trained model, and estimates a knowledge graph corresponding to the Japanese text to be estimated. The processing unit 320 may also perform the link prediction described above.

The communication unit 330 transmits and receives information and data to and from external devices. The Japanese text storage unit 341 stores the Japanese text to be estimated. The estimation result storage unit 342 stores the results of the estimation process based on the Japanese text to be estimated and the trained model. The trained model storage unit 343 stores the trained model. The output unit 350 outputs information and data acquired from outside, as well as estimation results, to the outside. The output unit 350 may further output link prediction results.

Figure 5 is a flowchart of the learning process in an embodiment of the present invention.

In step S100, the acquisition unit 110 acquires English text. English is an example of a first language. The English text may be referred to as a first character string. The English text includes multiple nouns. The multiple nouns included in the English text may be referred to as first nouns.

In step S101, the acquisition unit 110 acquires a trained model. The trained model is training data that includes correspondences between English text and knowledge graphs corresponding to the English text. Instead of the acquisition unit 110, the communication unit 130 may receive the trained model from another external device. The knowledge graph storage unit 142 may store a knowledge graph, and the trained model storage unit 143 may store a trained model.

In step S102, the processing unit 120 translates the English text into Japanese using a machine translation engine to generate Japanese text. The Japanese text may be referred to as a second character string. The Japanese text includes multiple nouns. The multiple nouns included in the Japanese text may be referred to as second nouns.

In step S103, the processing unit 120 changes the order of multiple nouns contained in the English text to correspond to the order of nouns contained in the Japanese text obtained by machine translation, thereby generating English text with a changed word order. The English text with a changed word order may be referred to as a first converted first character string, or may be referred to as Japanese-style English text.

In step S104, the processing unit 120 adds, before each of the multiple nouns included in the English text with the changed word order, a particle that follows the noun included in the Japanese text, based on the correspondence between the multiple nouns included in the English text and the multiple nouns included in the Japanese text. The character string obtained as a result of this may be referred to as the second converted first character string, or may be referred to as the marked English text. The marked text storage unit 141 stores the marked English text.

In step S105, the processing unit 120 performs transfer learning based on the marked English text and the trained model to update the model parameters of the trained model and generate an updated trained model. The updated trained model indicates the correspondence between the English text and the Japanese version of the knowledge graph corresponding to the English text. Because the English text has corresponding marked English text and a corresponding Japanese translation, the updated trained model consequently indicates the correspondence between the Japanese text and the Japanese knowledge graph. The Japanese version of the knowledge graph corresponding to the English text includes Japanese nouns that correspond to multiple nouns contained in the English text. The learning result memory unit 144 stores the updated trained model.

In step S106, the output unit 150 outputs the updated trained model. Instead of or in addition to the output unit 150, the communication unit 130 may transmit the updated trained model to another device. In this case, the communication unit 130 may transmit the updated trained model to the estimation device 30.

Figure 6 is a flowchart of the estimation process in an embodiment of the present invention.

In step S300, the acquisition unit 310 acquires the Japanese text to be estimated. The Japanese text to be estimated may be referred to as a third character string. The Japanese text storage unit 341 stores the Japanese text to be estimated. Instead of the acquisition unit 310, the communication unit 330 may receive the Japanese text to be estimated from an external device.

In step S301, the acquisition unit 310 acquires an updated trained model. The trained model storage unit 343 stores the updated trained model. Instead of the acquisition unit 310, the communication unit 330 may receive the updated trained model.

In step S302, processing unit 320 executes a calculation process. Specifically, processing unit 320 executes an estimation process based on the Japanese text to be estimated and the updated trained model, and extracts multiple nouns included in the Japanese text to be estimated and multiple particles included in the Japanese text to be estimated. The multiple particles included in the Japanese text to be estimated may also be referred to as multiple third particles. As a result, processing unit 320 can estimate and obtain a knowledge graph corresponding to the Japanese text.

In step S303, the output unit 350 outputs a knowledge graph corresponding to the Japanese text. Instead of the output unit 350, the communication unit 330 may transmit the knowledge graph corresponding to the Japanese text to an external device.

In step S304, the processing unit 320 may further perform link prediction.

As described above, a learning method according to an embodiment of the present invention includes the steps of: acquiring English text (first character string) written in English (first language), the English text including a plurality of English nouns (a plurality of first nouns); acquiring a trained model indicating a correspondence between the English text and at least a first subject and a first predicate included in the English text, where the first subject and the first predicate may be an English knowledge graph; translating the English text into Japanese to generate Japanese text (second character string), the Japanese text including a plurality of second nouns; and generating English text with a changed word order (Japanese-style English text, or a first converted first character string) by changing the structure of the English text so that the order of the plurality of nouns included in the English text corresponds to the order of the plurality of second nouns included in the Japanese text. the step of acquiring a plurality of first nouns included in the English text with the changed word order, adding one or more Japanese particles before each of the plurality of first nouns included in the English text based on the correspondence between the plurality of first nouns included in the English text and the plurality of second nouns included in the Japanese text, thereby acquiring marked English text (second converted first character string); and the step of performing transfer learning based on the marked English text and the trained model to generate an updated trained model indicating the correspondence between the English text and the plurality of second nouns included in the Japanese text, where the plurality of second nouns included in the Japanese text may be the Japanese version of a knowledge graph.

This makes it possible to efficiently estimate and generate a Japanese knowledge graph from Japanese text.

Furthermore, in this embodiment, the process of generating an updated trained model by performing transfer learning by the processing unit 120 based on the marked English text and the trained model corresponds to discontinuous and iterative extraction and learning, and in the discontinuous and iterative extraction, one or more other keywords other than nouns and verbs are included between the first subject and first predicate included in the English text, and the processing unit 120 may iteratively extract particles corresponding to each of multiple second nouns included in the Japanese text. The other keyword may be, for example, a phrase such as "surprisingly" in the sentence "The girl, to her surprise, heard a strange sound."

In addition, in this embodiment, the particles added before the multiple first nouns included in the highlighted English text may include case particles.

Furthermore, in this embodiment, the acquisition unit 310 included in the estimation device 30 acquires the Japanese text to be estimated (third character string) and acquires an updated trained model, and the processing unit 320 extracts multiple third nouns and multiple third particles included in the Japanese text to be estimated based on the Japanese text to be estimated and the updated trained model, thereby obtaining a knowledge graph corresponding to the Japanese text to be estimated. Here, the multiple third particles include particles corresponding to each of the multiple third nouns.

In addition, in this embodiment, the multiple third particles may include case particles.

Furthermore, at least one of the learning device 10 and the estimation device 30 may be configured as a user terminal. The functions of at least one of the learning device 10 and the estimation device 30 may be realized as a program that executes each process in the learning device 10 and the estimation device 30. Execution of this program may result in a billing process for the user.

If at least one of the programs that executes each process in the learning device 10 and the estimation device 30 is configured to be downloaded from another server device, the server device may be configured to automatically transmit the program to at least one of the learning device 10 and the estimation device 30. In this case, the program may be configured to function without any separate operation by a user of at least one of the learning device 10 and the estimation device 30.

This makes it possible to efficiently estimate and generate a Japanese knowledge graph from Japanese text.

The learning device 10 and the estimation device 30 may be configured as a single device.

<Hardware configuration>
7 is an explanatory diagram illustrating the hardware configuration of each device according to this embodiment. Each device is a learning device 10 and an estimation device 30. Each device includes an input/output module I, a storage module M, and a control module P. The input/output module I is realized by including some or all of the communication module H11, connection module H12, pointing device H21, keyboard H22, display H23, button H3, microphone H41, speaker H42, camera H51, and sensor H52.

The storage module M is realized by including a drive H7. The storage module M may further be configured to include all or part of the memory H8. The control module P is realized by including the memory H8 and a processor H9. These hardware components are connected to each other via a bus so that they can communicate with each other, and are supplied with power from the power supply H6.

The connection module H12 is a digital input/output port such as a USB (Universal Serial Bus). In the case of a portable device, the pointing device H21, keyboard H22, and display H23 are touch panels. The sensor H52 is an image sensor that acts as an acquisition unit. The power supply H6 is a power supply unit that supplies the electricity necessary to operate each device. In the case of a portable device, the power supply H6 is a battery.

Drive H7 is an auxiliary storage medium such as a hard disk drive or solid state drive. Drive H7 may also be non-volatile memory such as an EEPROM or flash memory, or a magneto-optical disk drive or flexible disk drive. Drive H7 is not limited to being built into each device, but may also be an external storage device connected to the connector of IF module H12.

Memory H8 is a primary storage medium such as random access memory. Memory H8 may also be cache memory. Memory H8 stores instructions as they are executed by one or more processors H9.

Processor H9 is a CPU (Central Processing Unit). Processor H9 may also be an MPU (Microprocessing Unit) or GPU (Graphics Processing Unit). Processor H9 reads programs and various data from drive H7 via memory H8 and performs calculations to execute instructions stored in one or more memories H8.

The input/output module I implements an acquisition unit 110, a communication unit 130, an output unit 150, an acquisition unit 310, a communication unit 330, and an output unit 350.

Control module P implements processing unit 120 and processing unit 320. Note that in this specification, the terms "learning device 10" and "estimation device 30" may be replaced with "control units P10" and "P30," respectively, and the terms "each of these devices" may be replaced with "control module P."

Drive H7 or memory H8 implements the learning data storage unit 140 and execution data storage unit 340.

The above has described in detail each embodiment and variant as one aspect of the present invention with reference to the drawings, but the specific configuration is not limited to each embodiment or variant, and includes design modifications within the scope of the present invention. Furthermore, various modifications of this aspect of the present invention are possible within the scope of the claims, and embodiments obtained by appropriately combining the technical means disclosed in different embodiments are also included in the technical scope of the present invention. Also included are configurations in which elements described in the above embodiments and variants are substituted for elements that achieve the same effect.

For example, one aspect of the present invention may be realized by combining some or all of the above embodiments.

The particular order or hierarchy of operations in the processes/flowcharts disclosed above is exemplary. Based on appropriate design, the particular order or hierarchy of operations in the flow charts may be rearranged. Furthermore, some operations may be combined or omitted. The disclosure of the claimed methods presents elements of various operations in an exemplary order, and is not limited to the particular order or hierarchy presented.

For example, combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B, and C," "one or more of A, B, and C," and "A, B, C, or any combination thereof" include any combination of A, B, and/or C, and may include multiple As, multiple Bs, or multiple Cs. Specifically, combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B, and C," "one or more of A, B, and C," and "A, B, C, or any combination thereof" may be A only, B only, C only, A and B, A and C, B and C, or A, B, and C, and any such combination may include one or more members of A, B, or C.

S...estimation system, 10...learning device, 30...estimation device, 110...acquisition unit, 120...processing unit, 130...communication unit, 140...learning data storage unit, 150...output unit, 310...acquisition unit, 320...processing unit, 330...communication unit, 340...execution data storage unit, 350...output unit

Claims (12)

1. A computer-implemented method for learning, comprising:
obtaining a first string written in a first language, the first string including a plurality of first nouns;
acquiring a trained model indicating a correspondence between the first character string and at least a first subject and a first predicate included in the first character string;
a step of translating the first character string into Japanese to generate a second character string, the second character string including a plurality of second nouns;
obtaining a first transformed first string by modifying a structure of the first string such that an order of the plurality of first nouns corresponds to an order of the plurality of second nouns;
adding one or more Japanese particles before each of the plurality of first nouns included in the first converted first character string based on the correspondence between the plurality of first nouns and the plurality of second nouns, thereby obtaining a second converted first character string;
and performing transfer learning based on the second converted first character string and the trained model to generate an updated trained model indicating a correspondence between the first character string and at least a second subject and a second predicate included in the second character string. generating the updated trained model by performing transfer learning based on the second transformed first string and the trained model corresponds to discontinuous and iterative extraction and training;
In the discontinuous and repeated extraction,
one or more keywords other than nouns and verbs are included between the first subject and the first predicate;
The method of claim 1 , further comprising the step of iteratively extracting a particle corresponding to each of the plurality of second nouns. The method of claim 1 , wherein the one or more Japanese particles include a case particle. obtaining a third character string written in Japanese;
extracting a plurality of third nouns and a plurality of third particles included in the third character string based on the third character string and the updated trained model;
The method of claim 1 , wherein the plurality of third particles includes a particle corresponding to each of the plurality of third nouns. The method of claim 4 , wherein the plurality of third particles includes a case particle. An apparatus comprising an acquisition unit and a processing unit,
The acquisition unit
Obtaining a first string written in a first language;
acquire a trained model indicating a correspondence between the first character string and at least a first subject and a first predicate included in the first character string;
the first string includes a plurality of first nouns;
the processing unit is configured to translate the first character string into Japanese to generate a second character string, the second character string including a plurality of second nouns;
The acquisition unit
is configured to obtain a first transformed first string by modifying a structure of the first string such that an order of the plurality of first nouns corresponds to an order of the plurality of second nouns;
is configured to add one or more Japanese particles before or after each of the plurality of first nouns included in the converted first character string based on a correspondence between the plurality of first nouns and the plurality of second nouns, thereby obtaining a second converted first character string;
the processing unit is configured to perform transfer learning based on the second converted first character string and the trained model, thereby generating an updated trained model indicating a correspondence between the first character string and at least a second subject and a second predicate included in the second character string. the acquisition unit is configured to acquire a third character string written in Japanese;
the processing unit is configured to extract, based on the third character string and the updated trained model, a plurality of third nouns and a plurality of third particles included in the third character string;
The device according to claim 6 , wherein the plurality of third particles includes a particle corresponding to each of the plurality of third nouns. The apparatus of claim 6 , wherein the one or more Japanese particles include a case particle. An apparatus comprising an acquisition unit and a processing unit,
the acquisition unit is configured to acquire a third character string written in Japanese and a trained model generated by the method of claim 1 ;
the processing unit is configured to extract, based on the third character string and the trained model, a plurality of third nouns and a plurality of third particles included in the third character string;
the trained model indicates a correspondence relationship between a first character string written in a first language and at least a first subject and a first predicate included in the first character string;
The plurality of third particles include a particle corresponding to each of the plurality of third nouns. An apparatus according to claim 6;
Another device comprising an acquisition unit and a processing unit,
the acquisition unit is configured to acquire a third character string written in Japanese and a trained model generated by the device according to claim 6;
the processing unit is configured to extract, based on the third character string and the trained model, a plurality of third nouns and a plurality of third particles included in the third character string;
the trained model indicates a correspondence relationship between a first character string written in a first language and at least a first subject and a first predicate included in the first character string;
and another device, wherein the plurality of third particles include particles corresponding to each of the plurality of third nouns . A program for executing the method described in any one of claims 1 to 5 when executed on a computer. A computer-readable storage medium storing a program that, when executed on a computer, causes the computer to execute the method described in any one of claims 1 to 5.