Information processing apparatus, information processing method, and program

Abstract

To provide an information processing apparatus, an information processing method, and a program for classifying materials into the same material even if the materials are described in a document with different names.SOLUTION: An information processing apparatus includes: a storage section that stores a knowledge graph in which a material identifier and a synonym of a material name are accumulated for each material concept, a mapping model, and a material feature quantity dictionary; and an operation processing section. The operation processing section learns a word feature quantity model on the basis of the document and the material name in the knowledge graph, generates a word feature quantity dictionary using the word feature quantity model, generates a determination model for determining whether or not the material name belongs to the synonym of the material concept in the knowledge graph, estimates a word feature quantity of an input material name using the word feature quantity model, determines whether or not a material concept in which the input material name belongs to the synonym is in the knowledge graph using the determination model for the estimated word feature quantity, obtains a material feature quantity associated with a material identifier corresponding to the material concept when there is the material concept, and obtains a material feature quantity from the word feature quantity using the mapping model when there is no material concept.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present disclosure relates to a technology for assigning feature quantities to material names extracted from documents. [Background technology]

[0002] The latest research results are often published in literature. In addition to academic papers, the latest research results are published in patent documents such as published patent gazettes and patent publications. These documents contain knowledge based on the latest research results. Therefore, by using the knowledge contained in these documents, researchers can grasp research trends and conduct analyses based on the latest data. For example, for the purpose of developing new materials, researchers can create statistical models to predict the properties of new materials by utilizing knowledge such as experimental data contained in patent documents and academic papers.

[0003] Patent documents and academic papers are unstructured data consisting of natural language, figures, tables, and so on. However, unstructured data cannot be analyzed by computers. To utilize unstructured data for data analysis, knowledge must be extracted from the unstructured data and then converted into structured data, such as table data. If this extraction process were performed by a human, it would require the human to read and interpret each patent publication and academic paper to extract the necessary knowledge, which is a significant workload. Furthermore, considering that many new patent documents and academic papers are published every day, it is not easy for a human to extract the necessary knowledge from all of these documents.

[0004] To address this issue, a system has been proposed that uses natural language processing technology to automatically extract useful knowledge from documents (see, for example, Patent Document 1). The system disclosed in Patent Document 1 extracts necessary information from documents based on a domain-specific natural language processing engine and a domain-specific ontology.

[0005] To effectively utilize experimental data from patent documents and academic papers, it is necessary not only to extract information such as the material names and compounding amounts described in the documents, but also to identify the material referred to by the material names and assign features that reflect the material's structure and properties. For materials included in a materials database, information such as chemical structure can be obtained from the database and used as features. However, when a new material not included in a materials database is described in a document, assigning features to the material is difficult. Furthermore, when the material names described in the document are generic terms or cannot uniquely identify the material, assigning features to the material is also difficult.

[0006] To address this issue, Non-Patent Document 1 discloses a technology that learns feature quantities calculated from the sentences of entities described in a sentence and a mapping function of feature quantities in a knowledge graph, and uses this to calculate feature quantities in the knowledge graph of entities described in a sentence. By using the technology disclosed in Non-Patent Document 1, it is possible to calculate feature quantities of material names described in a sentence by utilizing a knowledge graph related to materials. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] International Publication No. 2021 / 156684 [Non-patent literature]

[0008] [Non-Patent Document 1] Haseeb Shah, et al. “An open―world extension to knowledge graph completion models.” Proceedings of the AAAI conference on artificial intelligence, Vol.33, No.01, 2019. Summary of the Invention [Problem to be solved by the invention]

[0009] However, the technology disclosed in Non-Patent Document 1 assumes a one-to-one correspondence between word embedding representations and knowledge graph embedding representations. Material names often have multiple synonyms, and there is no one-to-one correspondence between word embedding representations and knowledge graph embedding representations. With the technology disclosed in Non-Patent Document 1, if a single material has multiple material names, multiple word embedding representations are generated, which may result in the same material being analyzed as multiple different materials. Therefore, it is difficult to apply the technology disclosed in Non-Patent Document 1 to feature assignment for material names.

[0010] One object of the present disclosure is to provide an information processing device, an information processing method, and a program that classify materials into the same material even if the material is described in a document under a different name. [Means for solving the problem]

[0011] An information processing device according to one aspect included in the present disclosure includes: a storage unit that stores, for each ingredient concept, an ingredient knowledge graph that accumulates information on ingredient identifiers, which are identifiers corresponding to the ingredient concept, and information on synonyms of ingredient names; a mapping model that is a model that estimates ingredient features that reflect characteristics of the ingredient concept from word features; and an ingredient feature dictionary that associates the ingredient features with the ingredient identifiers; a word feature model learning unit that learns a word feature model that estimates the word features from the ingredient names in the literature, based on literature obtained from a literature database related to ingredients and the ingredient names accumulated in the ingredient knowledge graph; a word feature dictionary creation unit that uses the word feature model to determine the word features of the ingredient names accumulated in the ingredient knowledge graph and creates a word feature dictionary that associates the determined word features with the ingredient names; and a mapping model that stores, for each ingredient concept, information on synonyms of ingredient names of the same ingredient concept, which correspond to each of a plurality of different word features in the word feature dictionary. a word feature estimation unit that, when a user inputs an ingredient name, estimates the word feature of the input ingredient name using the word feature model; a synonym determination unit that uses the synonym determination model with respect to the estimated word feature to determine whether the same ingredient concept, which is a synonym of the input ingredient name, is stored in the ingredient knowledge graph, and, if it is determined that the same ingredient concept is stored in the ingredient knowledge graph, identifies the ingredient identifier corresponding to the same ingredient concept; and an ingredient feature estimation unit that, when the ingredient identifier is identified by the synonym determination unit, refers to the ingredient feature dictionary to determine the ingredient feature associated with the received ingredient identifier, and, if the ingredient identifier is not identified by the synonym determination unit, determines the ingredient feature from the estimated word feature. [Effects of the Invention]

[0012] According to one aspect of the present disclosure, for a material whose name has multiple synonyms, even if multiple synonyms are described in a document, different feature values ​​are not assigned for each synonym, so that appropriate feature values ​​can be assigned to the same material. As a result, it is possible to more clearly identify the features, including the properties and characteristics, of each material across a large number of documents. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a block diagram showing an example of the configuration of a feature assignment device according to a first embodiment; [Figure 2] 2 is a diagram illustrating an example of the hardware configuration of a processing unit illustrated in FIG. 1. FIG. [Figure 3] FIG. 2 is a diagram schematically illustrating a data flow in the feature assignment device shown in FIG. [Figure 4] 1 is a flowchart illustrating an example of a procedure of pre-learning performed by the feature assignment device according to the first embodiment. [Figure 5] 1 is a flowchart showing an example of an operation procedure in which the feature assigning device according to the first embodiment assigns a feature to an ingredient name input by a user. [Figure 6] FIG. 6 is an image diagram for explaining the processing of steps S402 to S405 shown in FIG. 5. [Figure 7] FIG. 10 is a block diagram showing an example of the configuration of a feature assignment device according to a second embodiment. [Figure 8] FIG. 8 is a diagram schematically illustrating a data flow in the feature assignment device shown in FIG. 7. [Figure 9] 10 is a flowchart showing an example of an operation procedure in which the feature assigning device according to the second embodiment assigns a feature to an ingredient name input by a user. [Figure 10] 10 is a diagram showing an example of an image displayed on the output unit shown in FIG. 7 in step S702 shown in FIG. 9. FIG. [Figure 11] FIG. 10 is a block diagram showing an example of the configuration of a feature assignment device according to a third embodiment. [Figure 12]FIG. 12 is a diagram schematically illustrating a data flow in the feature assignment device shown in FIG. [Figure 13] 11 is a flowchart illustrating an example of a procedure of pre-learning performed by the feature assignment device according to the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] The feature assignment device of this embodiment is an information processing device that extracts material names from documents and assigns the same feature to materials that are the same as the extracted material names. An example of the information processing device of this embodiment will be described below. [Example]

[0015] The configuration of the feature adding device according to the first embodiment will be described with reference to the drawings. FIG. 1 is a block diagram showing an example of the configuration of the feature adding device according to the first embodiment. As shown in FIG. 1, the feature adding device 1 is an information processing device such as a computer. The feature adding device 1 includes an input unit 2, an output unit 3, a calculation processing unit 4, and a storage unit 5.

[0016] The input unit 2 is an input device such as a keyboard, a mouse, or a touch panel. The input unit 2 is used when a user of the feature assignment device 1 inputs instructions or data to the feature assignment device 1. The output unit 3 is an output device such as a display device that displays information. When a user interactively exchanges information with the arithmetic processing unit 4, the output unit 3 displays information to be executed by the arithmetic processing unit 4 and displays the results of processing executed by the arithmetic processing unit 4. The memory unit 5 is a storage device such as an HDD (Hard Disk Drive) or SSD (Solid State Drive). A literature database (DB) 41 is constructed in the memory unit 5. The memory unit 5 stores a materials knowledge graph 42.

[0017] FIG. 2 is a diagram showing an example of the hardware configuration of the arithmetic processing unit shown in FIG. 1. As shown in FIG. 2, the arithmetic processing unit 4 has a memory 7 that stores a program, and a CPU (Central Processing Unit) 6 that executes processing in accordance with the program. The memory 7 is a non-volatile memory such as a flash memory. When the CPU 6 executes the program, the functions of each of the units shown in FIG. 1, namely, the word feature model learning unit 11, the word feature estimation unit 12, the word feature dictionary creation unit 13, the synonym determination learning unit 14, the synonym determination unit 15, the mapping model learning unit 16, the material feature learning unit 17, and the material feature estimation unit 18, are executed. The program may be stored not only in the memory 7 but also in a portable recording medium.

[0018] Next, the configuration of the feature assigning device 1 and input / output data of each component will be described with reference to Fig. 1 and Fig. 3. Fig. 3 is a diagram schematically showing the flow of data in the feature assigning device shown in Fig. 1.

[0019] The literature DB 41 stores literature such as patent documents and academic papers related to materials. The material knowledge graph 42 stores information related to each of a plurality of material concepts. Specifically, the material knowledge graph 42 stores information corresponding to each material concept, including a material ID, which is an identifier that uniquely identifies the material concept, a material name and synonyms for the material name, a material classification system, and relationships between material concepts. The material ID may be assigned to each material concept in advance, or may be assigned by the material feature learning unit 17. In this Example 1, a case will be described in which the material ID is assigned by the material feature learning unit 17.

[0020] A material concept encompasses not only a specific material with a certain chemical composition, but also a set of materials or substructures that have the same properties or the same chemical structure. Multiple material names may be stored in the material knowledge graph 42 corresponding to one material concept. In this case, the multiple material names are synonyms of each other. In other words, if one of these material names is designated as the base name, the other material names are synonyms of the base name.

[0021] For example, let us explain the information stored in the materials knowledge graph for the case where the material concept is poly(amide)polymer. The information stored in the materials knowledge graph associates the material names poly(amide)polymer, polyamide, and polyamides with the material concept poly(amide)polymer. These material names are synonymous with each other. Furthermore, the information stored in the materials knowledge graph for the material concept poly(amide)polymer is that it is classified into nylons and cyanophycin polymers in the material classification system. Furthermore, regarding the relationship between material concepts, the information stored in the materials knowledge graph is that poly(amide)polymer is composed of poly(amide) macromolecules. The material concept of poly(amide)polymer includes not only specific materials with the same chemical structure, but also material sets or substructures with the same properties or the same chemical structure, such as polymers.

[0022] The materials knowledge graph 42 may be constructed independently or based on a public database, such as the Chemical Entities of Biological Interest.

[0023] The word feature model training unit 11 trains the word feature model Xwm using documents stored in the document DB 41 and material names stored as information in the material knowledge graph 42. The word feature model Xwm is a model that calculates the word feature Xw of the material names stored as information in the material knowledge graph 42 from the documents stored in the document DB 41. A specific example will be described. The word feature model training unit 11 calculates the word feature Xw for each of the material names poly(amide)polymer, polyamide, and polyamides. The word feature Xw is a fixed-length vector representing the characteristics of the material name calculated from the context of the document. The word feature Xw is calculated using word embedding techniques such as Word2Vec, FastText, or BERT. The word feature model training unit 11 transmits the word feature model Xwm to the word feature estimation unit 12 and the word feature dictionary creation unit 13.

[0024] The word feature dictionary creation unit 13 uses the word feature model Xwm to calculate the word feature Xw for each ingredient name stored as information in the ingredient knowledge graph 42. The word feature dictionary creation unit 13 creates a word feature dictionary Xwd that associates the calculated word feature Xw with the ingredient name. The word feature dictionary Xwd is, for example, a list of pairs of ingredient names and word feature Xw. The word feature dictionary creation unit 13 transmits the word feature dictionary Xwd to the synonym determination learning unit 14 and the mapping model learning unit 16.

[0025] The synonym determination learning unit 14 creates a synonym determination model Sdm using the word feature dictionary Xwd received from the word feature dictionary creation unit 13. The synonym determination model Sdm is a model that determines whether material names corresponding to multiple different word features Xw in the word feature dictionary Xwd are synonyms of material names of the same material concept. For example, when the synonym determination model Sdm is given a word feature Xw corresponding to poly(amide)polymer and a word feature Xw corresponding to polyamide, it returns a True determination. A True determination means that the target multiple material names are synonyms. On the other hand, when the synonym determination model Sdm is given a word feature Xw corresponding to poly(amide)polymer and a word feature Xw corresponding to nylon, it returns a False determination. A False determination means that the target multiple material names are not synonyms. The synonym determination learning unit 14 refers to the word feature dictionary Xwd and learns the word features Xw of multiple material names that are synonyms associated with the same material concept, thereby acquiring the above-mentioned determination function in the synonym determination model Sdm. The synonym determination learning unit 14 transmits the synonym determination model Sdm to the synonym determination unit 15.

[0026] The material feature quantity learning unit 17 refers to the material knowledge graph 42, assigns a material ID to each material concept stored as information in the material knowledge graph 42, and calculates the material feature quantity Xm. The material feature quantity learning unit 17 then creates a material feature quantity dictionary Xmd that associates the calculated material feature quantity Xm with the material ID. The material feature quantity dictionary Xmd is, for example, a list of pairs of material IDs and material feature quantities Xm. The material feature quantity learning unit 17 transmits the material feature quantity dictionary Xmd to the mapping model learning unit 16 and the material feature quantity estimation unit 18. The material feature quantity Xm is a fixed-length vector calculated based on the classification system of material concepts stored in the material knowledge graph 42 and the relationships between material concepts. The material feature quantity Xm reflects the properties of the material concepts. The properties are, for example, chemical properties. The material feature quantity Xm is calculated for each material concept using, for example, a technique such as knowledge graph embedding.

[0027] The mapping model training unit 16 trains a mapping model Mpm that estimates an ingredient feature Xm from a word feature Xw based on the word feature dictionary Xwd and the ingredient feature dictionary Xmd. The mapping model training unit 16 sends the trained mapping model to the ingredient feature estimation unit 18. Ideally, there is one ingredient feature Xm for one ingredient concept. There may be multiple word features Xw for one ingredient concept. In this case, the mapping model training unit 16 selects a representative word feature Xw from the multiple word features Xw for one ingredient concept and calculates a mapping between the selected word feature Xw and the ingredient feature Xm. Alternatively, the mapping model training unit 16 may calculate an average value of the multiple word features Xw for one ingredient concept and calculate a mapping between the calculated average value and the ingredient feature Xm.

[0028] When the user inputs an ingredient name via the input unit 2, the word feature estimation unit 12 estimates a word feature Xw of the input ingredient name using the word feature model Xwm. The word feature estimation unit 12 transmits the estimated word feature Xw to the synonym determination unit 15 and the ingredient feature estimation unit 18. The word feature estimation unit 12 may re-train the word feature model Xwm as necessary.

[0029] When the synonym determination unit 15 receives the word feature Xw from the word feature estimation unit 12, it refers to the ingredient knowledge graph 42. Then, using the synonym determination model Sdm for the received word feature Xw, the synonym determination unit 15 determines whether or not an ingredient concept identical to the ingredient name entered by the user, which belongs to a synonym, is stored in the ingredient knowledge graph 42. Specifically, the synonym determination unit 15 refers to the word feature dictionary Xwd and determines, according to the synonym determination model Sdm, whether or not the ingredient name corresponding to the received word feature Xw is in a synonym relationship with the identical ingredient concept stored in the ingredient knowledge graph 42. If the synonym determination unit 15 determines that the ingredient concept identical to the ingredient name of the received word feature Xw is stored in the ingredient knowledge graph 42, it transmits an ingredient ID corresponding to the identical ingredient concept to the ingredient feature estimation unit 18. On the other hand, if the synonym determination unit 15 determines that the same ingredient concept as the ingredient name of the received word feature Xw is not stored in the ingredient knowledge graph 42, it does not send the ingredient ID to the ingredient feature estimation unit 18.

[0030] When the ingredient feature estimation unit 18 receives an ingredient ID from the synonym determination unit 15, it refers to the ingredient feature dictionary Xmd, finds ingredient feature Xm associated with the received ingredient ID, and outputs the found ingredient feature Xm. On the other hand, when the ingredient feature estimation unit 18 does not receive an ingredient ID from the synonym determination unit 15, it uses the mapping model Mpm to find ingredient feature Xm from the word feature Xw received from the word feature estimation unit 12, and outputs the found ingredient feature Xm.

[0031] In the first embodiment, the word feature model Xwm, the word feature dictionary Xwd, the synonym determination model Sdm, the mapping model Mpm, and the material feature dictionary Xmd are transmitted from the generation side to the user side. However, these models and dictionaries may be stored in the storage unit 5. These models and dictionaries may be periodically updated in the storage unit 5. In this case, the word feature estimation unit 12 and the word feature dictionary creation unit 13 read the word feature model Xwm from the storage unit 5. The synonym determination learning unit 14 and the mapping model learning unit 16 read the word feature dictionary Xwd from the storage unit 5. The synonym determination unit 15 reads the synonym determination model Sdm from the storage unit 5. The mapping model learning unit 16 reads the material feature dictionary Xmd from the storage unit 5. The material feature estimation unit 18 reads the mapping model Mpm and the material feature dictionary Xmd from the storage unit 5.

[0032] 1 shows a case where the document DB 41 is stored in the storage unit 5, but the document DB 41 does not have to be stored in the storage unit 5. For example, a server on which the document DB 41 is stored may be connected to a network. In this case, the processing unit 4 connects to the server via the input unit 2, the output unit 3, and the network, and reads documents from the document DB 41 stored in the server. The network is, for example, the Internet.

[0033] Next, the operation of the feature assignment device 1 of the first embodiment will be described. First, the operation of the feature assignment device 1 performing pre-learning to create a plurality of models and a plurality of dictionaries will be described. Fig. 4 is a flowchart showing an example of the pre-learning procedure performed by the feature assignment device according to the first embodiment.

[0034] In step S301, the word feature model learning unit 11 learns a word feature model from the literature DB 41 and the material knowledge graph 42. Specifically, the word feature model learning unit 11 learns the word feature model from the contents described in literature such as patent documents and academic papers stored in the literature DB 41. The word feature model learning unit 11 stores the learned word feature model in the storage unit 5. In step S302, the word feature dictionary creation unit 13 uses the word feature model to calculate word features of material names associated with material concepts stored in the material knowledge graph 42. The word feature dictionary creation unit 13 then creates a word feature dictionary that associates material names with word features. The word feature dictionary creation unit 13 stores the created word feature dictionary in the storage unit 5.

[0035] In step S303, the material feature learning unit 17 calculates material features for each material concept stored in the material knowledge graph 42. Then, the material feature learning unit 17 creates a material feature dictionary that associates the calculated material feature with a material ID. The material feature learning unit 17 stores the created material feature dictionary in the storage unit 5. In step S304, the mapping model learning unit 16 learns a mapping model based on the word feature dictionary and the material feature dictionary. The mapping model learning unit 16 stores the learned mapping model in the storage unit 5. In step S305, the synonym determination learning unit 14 learns a synonym determination model from the word feature dictionary. The synonym determination learning unit 14 stores the learned synonym determination model in the storage unit 5.

[0036] In this way, the feature assignment device 1 prepares a word feature dictionary, a material feature dictionary, a mapping model, and a synonym determination model. The feature assignment device 1 stores the multiple models and multiple dictionaries prepared by pre-learning in the storage unit 5.

[0037] Next, an operational procedure will be described in which the feature assignment device 1 assigns a feature to an ingredient name input by a user. Fig. 5 is a flowchart showing an example of the operational procedure in which the feature assignment device according to the first embodiment assigns a feature to an ingredient name input by a user.

[0038] In step S401, when a user inputs an ingredient name into the feature assignment device 1 via the input unit 2, the word feature estimation unit 12 estimates word features of the input ingredient name using a word feature model. The word feature estimation unit 12 transmits the estimated word features to the synonym determination unit 15 and the ingredient feature estimation unit 18. In step S402, the synonym determination unit 15 uses the synonym determination model for the received word features to determine whether an ingredient concept identical to the input ingredient name, which belongs to a synonym, is stored in the ingredient knowledge graph 42. If, as a result of the determination in step S402, the synonym determination unit 15 determines that an ingredient concept identical to the ingredient name of the received word feature is stored in the ingredient knowledge graph 42, it identifies an ingredient ID corresponding to the identical ingredient concept (step S403). Then, the synonym determination unit 15 transmits the identified ingredient ID to the ingredient feature estimation unit 18. In step S404, when the material feature estimation unit 18 receives the material ID from the synonym determination unit 15, it refers to the material feature dictionary and outputs the material feature associated with the received material ID via the output unit 3.

[0039] On the other hand, if the synonym determination unit 15 determines in step S402 that the same ingredient concept as the ingredient name of the received word feature is not stored in the ingredient knowledge graph 42, it does not send the ingredient ID to the ingredient feature estimation unit 18. If the ingredient ID is not received from the synonym determination unit 15, the ingredient feature estimation unit 18 estimates ingredient features from the estimated word features using a mapping model, and outputs the estimated ingredient features via the output unit 3 (step S405).

[0040] FIG. 6 is a conceptual diagram illustrating the processing of steps S402 to S405 shown in FIG. 5. Material names are extracted from documents stored in the document DB 41, and word features are calculated for each material name. The word features are expressed as multidimensional vector quantities. The word features of multiple material names are distributed in a multidimensional coordinate space, as shown in graph gh1. Each triangle in graph gh1 represents a word feature. For each material concept stored in the materials knowledge graph 42, a material feature is calculated as a multidimensional vector quantity. The material features of multiple material concepts are distributed in a multidimensional space, as shown in graph gh2. Each circle in graph gh2 represents a material feature. However, the materials knowledge graph 42 has a predetermined target range for the user for the materials stored as information. For example, the material is a "polymer material." Therefore, the distribution range of graph gh2 is narrower than that of graph gh1.

[0041] In steps S403 and S404, if the ingredient name corresponding to the word feature shown in graph gh1 is a synonym of an ingredient concept stored in the ingredient knowledge graph 42, the triangle mark indicating the ingredient feature of the ingredient name is absorbed into the circle represented in graph gh2 and reflected in graph gh3. On the other hand, if the ingredient name corresponding to the word feature shown in graph gh1 is not a synonym of an ingredient concept stored in the ingredient knowledge graph 42, the ingredient feature estimation unit 18 assigns ingredient features to the ingredient as a new ingredient concept using a mapping model (step S405). The ingredient feature in this case is represented by a dotted circle in graph gh3.

[0042] As shown in graph gh1 in Figure 6, many material names are extracted from many documents, but the extracted material names include many synonyms of a certain base material name. On the other hand, the material knowledge graph 42 has high accuracy of accumulated information, but increasing the amount of material concept data manually requires a large workload. In contrast, the feature assignment device 1 extracts many material names from many documents and performs synonym determination on the extracted material names, thereby reducing the number of material names that become noise.

[0043] The information processing device of Example 1 includes a storage unit 5, a word feature model learning unit 11, a word feature dictionary creation unit 13, a synonym determination learning unit 14, a word feature estimation unit 12, a synonym determination unit 15, and an ingredient feature estimation unit 18. The storage unit 5 stores an ingredient knowledge graph 42 that accumulates, for each ingredient concept, ingredient IDs, which are identifiers corresponding to the ingredient concept, and information on synonyms of ingredient names, a mapping model that estimates ingredient features reflecting the characteristics of the ingredient concept from word features, and an ingredient feature dictionary that associates ingredient features with ingredient IDs. The word feature model learning unit 11 learns a word feature model that estimates word features from ingredient names in literature, based on literature obtained from a literature DB 41 related to ingredients and the ingredient names accumulated in the ingredient knowledge graph. The word feature dictionary creation unit 13 uses a word feature model to determine word features of ingredient names stored in the ingredient knowledge graph 42, and creates a word feature dictionary that associates the determined word features with ingredient names. The synonym determination learning unit 14 creates a synonym determination model that determines whether ingredient names corresponding to multiple different word features in the word feature dictionary are synonyms of ingredient names of the same ingredient concept. When an ingredient name is input by the user, the word feature estimation unit 12 estimates word features of the input ingredient name using the word feature model. The synonym determination unit 15 uses the synonym determination model on the estimated word features to determine whether an ingredient concept identical to the input ingredient name, which is a synonym, is stored in the ingredient knowledge graph 42. If the determination result shows that the ingredient concept is stored in the ingredient knowledge graph 42, the synonym determination unit 15 outputs an ingredient ID corresponding to the ingredient concept. If the ingredient concept is not stored in the ingredient knowledge graph 42, the synonym determination unit 15 does not output an ingredient ID. When the ingredient feature estimation unit 18 receives an ingredient ID from the synonym determination unit 15, it refers to the ingredient feature dictionary and finds ingredient features associated with the received ingredient ID, and when it does not receive an ingredient ID from the synonym determination unit 15, it finds ingredient features from estimated word features using a mapping model.

[0044] According to the first embodiment, for a material whose name has multiple synonyms, even if multiple synonyms are described in a document, different feature values ​​are not assigned to each synonym, so that appropriate feature values ​​can be assigned to the same material. Therefore, when many documents are targeted, the features including the properties and characteristics of each material can be made more prominent.

[0045] In the first embodiment, the information processing device also includes a material feature learning unit 17 and a mapping model learning unit 16. The material feature learning unit 17 refers to the material knowledge graph 42, finds material features corresponding to material IDs, and creates a material feature dictionary by associating the found material features with the material IDs. The mapping model learning unit 16 learns a mapping model based on the word feature dictionary and the material feature dictionary. As a result, a material feature dictionary and a mapping model are generated and stored in the storage unit 5. [Example]

[0046] In this second embodiment, the user can modify the relationship between material concepts and material names for the information stored in the material knowledge graph 42. In this second embodiment, the same components as those described in the first embodiment are denoted by the same reference numerals, and detailed descriptions thereof will be omitted.

[0047] The configuration of a feature assignment device according to Example 2 will be described. FIG. 7 is a block diagram showing an example of the configuration of the feature assignment device according to Example 2. FIG. 8 is a diagram schematically showing the flow of data in the feature assignment device shown in FIG. 7. As shown in FIG. 7, the feature assignment device 1 is an information processing device having an input unit 2, an output unit 3, a calculation processing unit 4a, and a storage unit 5. An example of the hardware configuration of the calculation processing unit 4a is similar to the configuration described with reference to FIG. 2, and therefore a detailed description thereof will be omitted.

[0048] Compared to the calculation processing unit 4 shown in Fig. 1, the calculation processing unit 4a has a configuration in which a confirmation necessity determination unit 21, a confirmation correction unit 22, and a material feature re-estimation unit 23 are added. The CPU 6 shown in Fig. 2 executes a program to execute the functions of the confirmation necessity determination unit 21, the confirmation correction unit 22, and the material feature re-estimation unit 23 shown in Fig. 7. The configuration added in this second embodiment will be described.

[0049] The confirmation necessity determination unit 21 determines whether the ingredient feature Xm estimated for the input ingredient name is appropriate in relation to the ingredient concepts stored in the ingredient knowledge graph 42, and whether user confirmation is required based on the word feature Xw corresponding to the ingredient name and the word feature dictionary Xwd. A specific example of confirmation necessity determination will be described. For the input ingredient name, the confirmation necessity determination unit 21 compares the word feature Xw calculated by the word feature estimation unit 12 with the word feature Xw described in the word feature dictionary Xwd, and determines whether the difference between these values ​​is greater than a predetermined threshold. If the difference is greater than the threshold, the confirmation necessity determination unit 21 transmits a determination result to the confirmation correction unit 22 indicating that user confirmation is required. If the difference is equal to or less than the threshold, the confirmation necessity determination unit 21 transmits a determination result to the confirmation correction unit 22 indicating that user confirmation is not required.

[0050] When the confirmation correction unit 22 receives a determination result indicating that user confirmation is required from the confirmation necessity determination unit 21, it estimates the relationship between the input material name and the material concept stored in the material knowledge graph 42 based on the material feature quantity Xm calculated by the material feature quantity estimation unit 18. Then, the confirmation correction unit 22 outputs the estimation result via the output unit 3 to present it to the user and prompt the user to confirm whether the estimation result is correct. The checking and correcting unit 22 causes the output unit to display one option from a plurality of options for the relationship between the material name and material concept indicated by the estimation result, so that the user can select one option. If the estimation result by the confirming and correcting unit 22 is incorrect, the user inputs details to correct the relationship between the material name and the material concept stored in the material knowledge graph 42. When the user inputs the details of correction, the confirming and correcting unit 22 accepts the details of correction as a confirmation result and sends the confirmation result to the material feature re-estimation unit 23. The confirming and correcting unit 22 can use, for example, knowledge graph link prediction technology to infer the relationship with other material concepts. If the estimation result by the confirming and correcting unit 22 is correct, the user does not input any details of correction and ends the confirmation process. If the user does not input any details of correction, the confirming and correcting unit 22 sends a confirmation result to the material feature re-estimation unit 23 indicating that no correction has been made.

[0051] The material feature quantity re-estimation unit 23 re-estimates the material feature quantity Xm based on the confirmation result received from the checking and correcting unit 22 and the material knowledge graph 42. Specifically, if the confirmation result received from the checking and correcting unit 22 includes correction content, the material feature quantity re-estimation unit 23 re-estimates the material feature quantity Xm of the input material name based on the material knowledge graph 42 that reflects the correction content. The material feature quantity re-estimation unit 23 uses a technique such as knowledge graph embedding to re-estimate the material feature quantity Xm.

[0052] Next, the operation of the feature assignment device 1 in Example 2 will be described with reference to FIG. 9. FIG. 9 is a flowchart showing an example of an operation procedure in which the feature assignment device according to Example 2 assigns features to ingredient names input by a user. In Example 2, the feature assignment device 1 also performs the pre-learning described with reference to FIG. 4, as in Example 1, but a detailed description thereof will be omitted. Furthermore, the processing from steps S401 to S405 shown in FIG. 9 is the same as the processing from steps S401 to S405 described with reference to FIG. 5, and therefore a detailed description thereof will be omitted.

[0053] In step S701, the confirmation necessity determination unit 21 determines, based on the word features of the input ingredient name and the word feature dictionary, from the estimated ingredient features, whether the relationship between the ingredient name and the ingredient concepts stored in the ingredient knowledge graph 42 is appropriate and whether user confirmation is required. If the result of the determination in step S701 is that user confirmation is not required (step S701: No), the confirmation necessity determination unit 21 transmits the determination result that user confirmation is not required to the confirmation correction unit 22. On the other hand, if the result of the determination in step S701 is that user confirmation is required (step S701: Yes), the confirmation necessity determination unit 21 transmits the determination result that user confirmation is required to the confirmation correction unit 22.

[0054] When the confirmation correction unit 22 receives a determination result from the confirmation necessity determination unit 21 indicating that user confirmation is not required, it outputs the ingredient feature received from the ingredient feature estimation unit 18. On the other hand, when the confirmation correction unit 22 receives a determination result from the confirmation necessity determination unit 21 indicating that user confirmation is required, it estimates the relationship between the input ingredient name and the ingredient concept stored in the ingredient knowledge graph 42. Then, in step S702, the confirmation correction unit 22 outputs the estimation result to present it to the user. Then, when the user inputs confirmation details, the confirmation correction unit 22 accepts the correction details. In step S703, the ingredient feature re-estimation unit 23 re-estimates the ingredient feature of the input ingredient name based on the user's correction details and the ingredient knowledge graph 42. The ingredient feature re-estimation unit 23 outputs the re-estimated ingredient feature.

[0055] Fig. 10 is a diagram showing an example of an image displayed on the output unit shown in Fig. 7 in step S702 shown in Fig. 9. Fig. 10 shows a case where the output unit 3 shown in Fig. 7 is a display device. In step S702, the confirmation correction unit 22 causes the output unit 3 to display an image showing information prompting the user to confirm the material name of the estimated material feature quantity, as shown in Fig. 10.

[0056] In step S702 shown in FIG. 9 , the checking and correcting unit 22 displays on the output unit 3 one of multiple options for the relationship between the material name and the material concept indicated by the estimation result, allowing the user to select one option. The image shown in FIG. 10 is an example of the estimation result by the checking and correcting unit 22 for the material name "Polymer A" input via the input unit 2. Specifically, FIG. 10 shows that the material name "Polymer A" is estimated to be in a "subclass of" relationship with the material concept "Polymer X" in the material knowledge graph 42 and in a "has part" relationship with the material concept "Monomer a" in the material knowledge graph 42. The "relationship" is configured so that the user can select from multiple options using a pull-down menu. The options for "relationship" are not limited to "subclass of" and "has part." In addition to "subclass of" and "has part," options for "relationship" may also include, for example, "not related."

[0057] If the estimation result by the checking and correcting unit 22 is incorrect, when inputting correction details regarding the relationship between the material name and the material concept, the user can simply select the correct option while viewing multiple candidate options displayed on the output unit 3. In this case, the user can easily input the correction details, thereby reducing the burden of input operations.

[0058] The feature assignment device 1 of the second embodiment further includes a confirmation necessity determination unit 21, a confirmation correction unit 22, and an ingredient feature re-estimation unit 23 in addition to the components described in the first embodiment. When ingredient features are calculated for an input ingredient name using a mapping model, the confirmation necessity determination unit 21 determines, based on the word feature dictionary and the word features, whether the relationship between the ingredient name and the ingredient concepts stored in the ingredient knowledge graph 42 is appropriate and whether user confirmation is required. When the confirmation necessity determination unit 21 determines that user confirmation is required, the confirmation correction unit 22 outputs the relationship between the input ingredient name and the ingredient concepts stored in the ingredient knowledge graph 42 based on the calculated ingredient features. If the relationship is corrected by the user, the confirmation correction unit 22 accepts the correction as a confirmation result. The ingredient feature re-estimation unit 23 re-estimates ingredient features for the input ingredient name based on the confirmation result and the ingredient knowledge graph 42.

[0059] According to the second embodiment, it is determined whether the material feature amount estimated using the mapping model for the input material name is appropriate based on the relationship with the material concepts stored in the material knowledge graph 42, and whether user confirmation is required. Therefore, if it is determined that user confirmation is required, even if the relationship between the input material name and the material concepts stored in the material knowledge graph 42 is incorrect, these relationships can be corrected by the user. The material feature amount of the input material name is corrected to an appropriate value, and the accuracy of the material information stored in the material knowledge graph 42 is improved. [Example]

[0060] In this Example 3, documents to be acquired from the document DB 41 are selected in order to reduce noise in the information about materials accumulated in the material knowledge graph 42. In this Example 3, the same components as those described in Example 1 are denoted by the same reference numerals, and detailed descriptions thereof will be omitted.

[0061] The configuration of a feature assignment device according to Example 3 will be described. FIG. 11 is a block diagram showing an example of the configuration of the feature assignment device according to Example 3. FIG. 12 is a diagram schematically showing the flow of data in the feature assignment device shown in FIG. 11. As shown in FIG. 11, the feature assignment device 1 is an information processing device having an input unit 2, an output unit 3, a calculation processing unit 4b, and a storage unit 5. An example of the hardware configuration of the calculation processing unit 4b is similar to the configuration described with reference to FIG. 2, and therefore a detailed description thereof will be omitted.

[0062] Compared to the calculation processing unit 4 shown in Fig. 1, the calculation processing unit 4b has a configuration in which a text type determination unit 31 is added. The CPU 6 shown in Fig. 2 executes a program to execute the function of the text type determination unit 31 shown in Fig. 11. The configuration added in this third embodiment will be described below.

[0063] The sentence type determination unit 31 determines the type of the target sentence from the contents of words and other information in the target sentence that describes the name of the material to which a feature is to be assigned. For example, it determines whether the sentence is about an inorganic material or an organic material from keywords and other information in the target sentence. The type of the target sentence is not limited to inorganic materials and organic materials. For example, if the material name is "polymer," the type of the target sentence is "organic material." The name of the material to which a feature is to be assigned is, for example, the material name input by the user to the word feature estimation unit 12 in FIG. 12. The sentence type determination unit 31 transmits a sentence determination result indicating the determined type to the word feature model learning unit 11 and the material feature learning unit 17.

[0064] Next, the operation of the feature assignment device 1 of the third embodiment will be described with reference to Fig. 13. Fig. 13 is a flowchart showing an example of a procedure of pre-learning performed by the feature assignment device according to the third embodiment. Among the processes of steps S301 to S305 shown in Fig. 13, the processes of steps S302, S304, and S305 are the same as the processes of steps S302, S304, and S305 described with reference to Fig. 4, and therefore detailed description thereof will be omitted.

[0065] In step S901, the sentence type determination unit 31 determines the type of the target sentence from the description content of the target sentence. The sentence type determination unit 31 transmits a sentence determination result indicating the determined type to the word feature model learning unit 11 and the ingredient feature learning unit 17. In step S301, the word feature model learning unit 11 extracts documents containing sentences of the same type as the target sentence from the literature DB 41 based on the sentence determination result, and learns a word feature model based on the extracted documents and ingredient names stored as information in the ingredient knowledge graph 42. In step S303, the ingredient feature learning unit 17 extracts information related to the type of the target sentence from the ingredient knowledge graph 42 based on the sentence determination result, and calculates ingredient features using the extracted information. The ingredient feature learning unit 17 then creates an ingredient feature dictionary that associates the calculated ingredient features with ingredient IDs.

[0066] Here, a specific example of information extracted from the material knowledge graph 42 by the material feature learning unit 17 in relation to the type of target sentence will be described. For example, assume that the same material Z exhibits different functions when used as a polymer material and when used as a low-molecular-weight material. Furthermore, consider a case where a user collects information about material Z used as a polymer material. In this case, the type indicated by the sentence determination result is "polymer material." Then, the material feature learning unit 17 calculates material features by extracting information related to "polymer material" from the material knowledge graph 42 for the material concept of material Z, including the material name, synonyms of the material name, material classification system, and relationships between material concepts. As a result, the material feature when material Z functions as a low-molecular-weight material becomes smaller, making it easier for the user to obtain the information they need.

[0067] The feature assignment device 1 of this Example 3 has a sentence type determination unit 31 in addition to the configuration described in Example 1. The sentence type determination unit 31 determines the type of a target sentence in which an input material name is described from the description content of the target sentence, and outputs a sentence determination result indicating the determined type to the word feature model learning unit 11 and the ingredient feature learning unit 17.

[0068] According to the third embodiment, documents related to the target material for which the user needs information are extracted from the document DB 41, which increases the tendency for material names that are closely related to the target material to be accumulated in the material knowledge graph 42. This reduces noise in the information related to material concepts accumulated in the material knowledge graph 42. As a result, the accuracy of the information accumulated in the material knowledge graph 42 improves.

[0069] The above-described embodiments are illustrative examples of the present invention and are not intended to limit the scope of the present invention to these embodiments. Those skilled in the art can implement the present invention in various other aspects without departing from the scope of the present invention. For example, although the application of Example 3 to Example 1 has been described, Example 3 may also be applied to Example 2.

[0070] The above-described embodiment also includes the following features, but the features included in the present embodiment are not limited to those listed below.

[0071] (Item 1) a storage unit that stores a material knowledge graph that accumulates, for each material concept, information on material identifiers, which are identifiers corresponding to the material concept, and synonyms of material names; a mapping model that is a model that estimates material features that reflect the characteristics of the material concept from word features; and a material feature dictionary that associates the material features with the material identifiers; a word feature model learning unit that learns a word feature model that estimates the word feature from the material name in the document, based on the document obtained from a document database related to materials and the material name stored in the material knowledge graph; a word feature dictionary creation unit that uses the word feature model to determine the word features of the ingredient names stored in the ingredient knowledge graph, and creates a word feature dictionary that associates the determined word features with the ingredient names; a synonym determination learning unit that creates a synonym determination model that determines whether or not ingredient names corresponding to each of a plurality of different word features in the word feature dictionary are synonyms of the ingredient names of the same ingredient concept; a word feature estimation unit that, when a user inputs an ingredient name, estimates the word feature of the input ingredient name using the word feature model; a synonym determination unit that uses the synonym determination model for the estimated word feature to determine whether the same ingredient concept that belongs to the synonym of the input ingredient name is stored in the ingredient knowledge graph, and that specifies the ingredient identifier corresponding to the same ingredient concept when it is determined that the same ingredient concept is stored in the ingredient knowledge graph; an ingredient feature estimation unit that, when the ingredient identifier is identified by the synonym determination unit, refers to the ingredient feature dictionary to determine the ingredient feature associated with the received ingredient identifier, and, when the ingredient identifier is not identified by the synonym determination unit, determines the ingredient feature from the estimated word feature using the mapping model; An information processing device having the above.

[0072] (Item 2) In the information processing device according to item 1, a material feature learning unit that refers to the material knowledge graph, obtains the material feature corresponding to the material identifier, and creates the material feature dictionary by associating the obtained material feature with the material identifier; a mapping model learning unit that learns the mapping model based on the word feature dictionary and the material feature dictionary; The information processing device further comprises:

[0073] (Item 3) In the information processing device according to item 2, a confirmation necessity determination unit that, when the material feature is determined for the input material name using the mapping model, determines whether or not a relationship between the material name and the material concept stored in the material knowledge graph is appropriate, and whether or not confirmation with the user is necessary, based on the word feature dictionary and the word feature; a confirmation correction unit that outputs a relationship between the input material name and the material concept stored in the material knowledge graph based on the material feature quantity obtained when the confirmation necessity determination unit determines that the user's confirmation is necessary, and that receives the correction content as a confirmation result when the user corrects the relationship; a material feature quantity re-estimation unit that re-estimates the material feature quantity for the input material name based on the confirmation result and the material knowledge graph; The information processing device further comprises:

[0074] (Item 4) Item 3. In the information processing device according to item 3, an output unit for displaying information; The checking and correcting unit Inferring a relationship between the input material name and the material concept stored in the material knowledge graph, displaying the inference result on the output unit, and displaying one option from a plurality of options on the output unit so that the user can select it with respect to the relationship indicated by the inference result. Information processing device.

[0075] (Item 5) In the information processing device according to any one of items 2 to 4, a text type determination unit that determines the type of a target text containing the material name to be input based on the description content of the target text and outputs a text type determination result indicating the determined type; The word feature model learning unit extracting documents containing the same type of sentence as the target sentence from the document database based on the sentence determination result, and learning the word feature model based on the extracted documents and the material names stored in the material knowledge graph; The material feature amount learning unit extracting information related to the target sentence from the ingredient knowledge graph based on the sentence determination result, determining the ingredient feature quantity using the extracted information, and creating the ingredient feature quantity dictionary in which the determined ingredient feature quantity is associated with the ingredient identifier. Information processing device. [Explanation of symbols]

[0076] 1: feature assignment device, 2: input unit, 3: output unit, 4, 4a, 4b: calculation processing unit, 5: storage unit, 6: CPU, 7: memory, 11: word feature model learning unit, 12: word feature estimation unit, 13: word feature dictionary creation unit, 14: synonym judgment learning unit, 15: synonym judgment unit, 16: mapping model learning unit, 17: material feature learning unit, 18: material feature estimation unit, 21: confirmation necessity judgment unit, 22: confirmation correction unit, 23: material feature re-estimation unit, 31: text type judgment unit, 41: literature database, 42: material knowledge graph.

Claims

1. a storage unit that stores a material knowledge graph that accumulates, for each material concept, information on material identifiers, which are identifiers corresponding to the material concept, and synonyms of material names; a mapping model that is a model that estimates material features that reflect the characteristics of the material concept from word features; and a material feature dictionary that associates the material features with the material identifiers; a word feature model learning unit that learns a word feature model that estimates the word feature from the material name in the document, based on the document obtained from a document database related to materials and the material name stored in the material knowledge graph; a word feature dictionary creation unit that uses the word feature model to determine the word features of the ingredient names stored in the ingredient knowledge graph, and creates a word feature dictionary that associates the determined word features with the ingredient names; a synonym determination learning unit that creates a synonym determination model that determines whether or not ingredient names corresponding to each of a plurality of different word features in the word feature dictionary are synonyms of the ingredient names of the same ingredient concept; a word feature estimation unit that, when a user inputs an ingredient name, estimates the word feature of the input ingredient name using the word feature model; a synonym determination unit that uses the synonym determination model for the estimated word feature to determine whether the same ingredient concept that belongs to the synonym of the input ingredient name is stored in the ingredient knowledge graph, and that specifies the ingredient identifier corresponding to the same ingredient concept when it is determined that the same ingredient concept is stored in the ingredient knowledge graph; an ingredient feature estimation unit that, when the ingredient identifier is identified by the synonym determination unit, refers to the ingredient feature dictionary to determine the ingredient feature associated with the received ingredient identifier, and, when the ingredient identifier is not identified by the synonym determination unit, determines the ingredient feature from the estimated word feature using the mapping model; An information processing device having the above.

2. 2. The information processing device according to claim 1, a material feature learning unit that refers to the material knowledge graph, obtains the material feature corresponding to the material identifier, and creates the material feature dictionary by associating the obtained material feature with the material identifier; a mapping model learning unit that learns the mapping model based on the word feature dictionary and the material feature dictionary; The information processing device further comprises:

3. 3. The information processing device according to claim 2, a confirmation necessity determination unit that, when the material feature is determined for the input material name using the mapping model, determines whether or not a relationship between the material name and the material concept stored in the material knowledge graph is appropriate, and whether or not confirmation with the user is necessary, based on the word feature dictionary and the word feature; a confirmation correction unit that outputs a relationship between the input material name and the material concept stored in the material knowledge graph based on the material feature quantity obtained when the confirmation necessity determination unit determines that the user's confirmation is necessary, and that receives the correction content as a confirmation result when the user corrects the relationship; a material feature quantity re-estimation unit that re-estimates the material feature quantity for the input material name based on the confirmation result and the material knowledge graph; The information processing device further comprises:

4. 4. The information processing device according to claim 3, an output unit for displaying information; The checking and correcting unit Inferring a relationship between the input material name and the material concept stored in the material knowledge graph, displaying the inference result on the output unit, and displaying one option from a plurality of options on the output unit so that the user can select it with respect to the relationship indicated by the inference result. Information processing device.

5. 3. The information processing device according to claim 2, a text type determination unit that determines the type of a target text containing the material name to be input based on the description content of the target text and outputs a text type determination result indicating the determined type; The word feature model learning unit extracting documents containing the same type of sentence as the target sentence from the document database based on the sentence determination result, and learning the word feature model based on the extracted documents and the material names stored in the material knowledge graph; The material feature amount learning unit extracting information related to the type of the target sentence from the ingredient knowledge graph based on the sentence determination result, determining the ingredient feature amount using the extracted information, and creating the ingredient feature amount dictionary in which the determined ingredient feature amount is associated with the ingredient identifier. Information processing device.

6. An information processing method executed by an information processing device, a material knowledge graph that stores, for each material concept, information on material identifiers, which are identifiers corresponding to the material concept, and synonyms of material names; a mapping model that is a model that estimates material features that reflect the characteristics of the material concept from word features; and a material feature dictionary that associates the material features with the material identifiers; learning a word feature model that estimates the word feature from the material name in the literature, based on the literature obtained from a literature database related to materials and the material name stored in the material knowledge graph; Using the word feature model, the word feature of the material name stored in the material knowledge graph is obtained, and a word feature dictionary is created that associates the obtained word feature with the material name; creating a synonym determination model for determining whether or not material names corresponding to the plurality of different word features in the word feature dictionary are synonyms of the material names of the same material concept; When a user inputs an ingredient name, the word feature model is used to estimate the word feature of the input ingredient name; using the synonym determination model for the estimated word feature quantity, to determine whether the same ingredient concept that belongs to the synonym of the input ingredient name is stored in the ingredient knowledge graph, and if it is determined that the same ingredient concept is stored in the ingredient knowledge graph, identifying the ingredient identifier corresponding to the same ingredient concept; When the material identifier is specified, the material feature dictionary is referenced to determine the material feature associated with the specified material identifier, and when the material identifier is not specified, the material feature is determined from the estimated word feature using the mapping model. Information processing methods.

7. On the computer, a material knowledge graph that stores, for each material concept, information on material identifiers, which are identifiers corresponding to the material concept, and synonyms of material names; a mapping model that is a model that estimates material features that reflect the characteristics of the material concept from word features; and a material feature dictionary that associates the material features with the material identifiers; learning a word feature model that estimates the word feature from the material name in the literature, based on the literature obtained from a literature database related to materials and the material name stored in the material knowledge graph; Using the word feature model, the word feature of the material name stored in the material knowledge graph is obtained, and a word feature dictionary is created that associates the obtained word feature with the material name; creating a synonym determination model for determining whether or not material names corresponding to the plurality of different word features in the word feature dictionary are synonyms of the material names of the same material concept; When a user inputs an ingredient name, the word feature model is used to estimate the word feature of the input ingredient name; using the synonym determination model for the estimated word feature quantity, to determine whether the same ingredient concept that belongs to the synonym of the input ingredient name is stored in the ingredient knowledge graph, and if it is determined that the same ingredient concept is stored in the ingredient knowledge graph, identifying the ingredient identifier corresponding to the same ingredient concept; When the material identifier is specified, the material feature dictionary is referenced to determine the material feature associated with the specified material identifier, and when the material identifier is not specified, the material feature is determined from the estimated word feature using the mapping model. A program to make it happen.

Images