Processing method, processing device, and processing program

Abstract

The present invention improves analysis accuracy in analyzing a document that includes text information describing a substance. A processing method according to the present invention includes: a first output step for outputting a first distributed representation based on structure information pertaining to a substance having a substance name included in a document; a second output step for outputting a second distributed representation based on text information describing a substance having a substance name included in the document; and a training step for training a model by using the first distributed representation and the second distributed representation.

Description

Processing method, processing device, and processing program

[0001] The present disclosure relates to a processing method, a processing device, and a processing program.

[0002] Documents such as papers, magazines, and patent documents in the field of chemistry contain structural information that indicates the molecular structure of a substance and text information that describes the substance. When analyzing such documents using a machine learning model, the structural information and the text information have traditionally been analyzed separately. Furthermore, the text information has been analyzed in the same way for both substance names and other character strings.

[0003] Korean Patent Publication No. 2021-0044003

[0004] However, the above analysis method may not always provide the desired analysis results.

[0005] The present disclosure aims to improve the accuracy of analysis when analyzing documents that include text information describing substances.

[0006] A first aspect of the present disclosure is a processing method including: a first output step of outputting a first distributed representation based on structural information of a substance having a substance name included in a document; a second output step of outputting a second distributed representation based on text information describing a substance having a substance name included in the document; and a learning step of learning a model using the first distributed representation and the second distributed representation.

[0007] According to the first aspect of the present disclosure, it is possible to improve the accuracy of analysis when analyzing a document that includes text information describing a substance.

[0008] A second aspect of the present disclosure is the processing method according to the first aspect, wherein in the learning step, learning the model includes learning a model that has already been trained.

[0009] A third aspect of the present disclosure is a processing method according to the first or second aspect, wherein the text information is a sentence including a substance name, and in the learning process, learning the model includes learning the model using a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output process and the second output process.

[0010] A fourth aspect of the present disclosure is a processing method according to the third aspect, wherein, when the text information is a sentence including a substance name, a set of distributed representations is generated in which a first distributed representation is placed at a position corresponding to the position of the substance name and a second distributed representation is placed at a position corresponding to the position of the sentence other than the substance name.

[0011] A fifth aspect of the present disclosure is a processing method according to the third aspect, wherein the text information is a sentence including a substance name, and in the learning process, learning the model includes: masking a portion of the text information other than the substance name; and performing supervised learning on the text information after the masking process, using as input data a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output process and the second output process, and using as correct data the portion of the text information other than the substance name.

[0012] A sixth aspect of the present disclosure is a processing method according to the second aspect, wherein the text information is a sentence including a substance name, and in the learning process, learning the trained model includes: masking the substance name of the text information using a common token; and performing supervised learning on the text information after the masking process, using as input data a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output process and the second output process, and using the substance name of the text information as correct answer data.

[0013] A seventh aspect of the present disclosure is a processing method described in any of the first to sixth aspects, wherein the text information is a sentence including a substance name, and in the learning process, learning the model includes learning either a trained first model that outputs a first distributed representation, or a trained second model that outputs a second distributed representation, or the trained model to which the outputs of the trained first model and the trained second model are input.

[0014] An eighth aspect of the present disclosure is a processing method according to the seventh aspect, wherein in the learning process, the trained first model that outputs the first distributed representation is trained so that first distributed representations based on structural information of substances with different substance names that are included in sentences with similar contexts are arranged close to each other.

[0015] A ninth aspect of the present disclosure is a processing method according to the first aspect, comprising a third output step of inputting a first distributed representation based on structural information of a substance into a trained model trained in the training step, thereby outputting a token indicating the use or function of the substance.

[0016] A tenth aspect of the present disclosure is a processing method according to the first aspect, comprising a third output step of inputting a second distributed representation based on text information describing the use or function of a substance into a trained model trained in the training step, thereby outputting a token indicating the substance name of the substance.

[0017] An eleventh aspect of the present disclosure is a processing method according to the first aspect, comprising a third output step of inputting either or both of a first distributed representation based on structural information of a substance and a second distributed representation based on text information describing the use or function of the substance into a trained model trained in the training step, thereby outputting tokens indicating substance names of related substances related to the substance.

[0018] A twelfth aspect of the present disclosure is a processing method according to any one of the first to eleventh aspects, including an identification step of identifying structural information of a substance having a substance name or a substance name of a substance having structural information by referring to a database in which substance names are associated with structural information of substances having the substance names.

[0019] A thirteenth aspect of the present disclosure is a processing method according to the twelfth aspect, wherein if a substance having a substance name included in the document is a substance that becomes a single atom when structural information of the substance is converted into a string according to a predetermined rule, or if a substance having a substance name included in the document is a substance with an abnormal valence, the second output step outputs the second distributed representation based on the substance name included in the document.

[0020] A fourteenth aspect of the present disclosure is the processing method according to any one of the first to eleventh aspects, wherein the documents include papers, magazines, and patent documents.

[0021] A fifteenth aspect of the present disclosure is a processing device (100, 1500) having a control unit (200), wherein the control unit (200) outputs a first distributed representation based on structural information of a substance having a substance name included in a document, outputs a second distributed representation based on text information describing a substance having a substance name included in the document, and trains a model using the first distributed representation and the second distributed representation.

[0022] According to the fifteenth aspect of the present disclosure, it is possible to improve the accuracy of analysis when analyzing a document that includes text information describing a substance.

[0023] A sixteenth aspect of the present disclosure is a processing device (100, 1500) according to the fifteenth aspect, wherein the control unit (200) inputs a first distributed representation based on structural information of a substance into a trained model, thereby outputting a token indicating the use or function of the substance.

[0024] A seventeenth aspect of the present disclosure is a processing device (100, 1500) according to the fifteenth aspect, wherein the control unit (200) outputs a token indicating the substance name of the substance by inputting a second distributed representation based on text information describing the use or function of the substance into a trained model.

[0025] An eighteenth aspect of the present disclosure is a processing program that causes a control unit (200) to execute the following steps: a first output step of outputting a first distributed representation based on structural information of a substance having a substance name included in a document; a second output step of outputting a second distributed representation based on text information describing a substance having a substance name included in the document; and a learning step of learning a model using the first distributed representation and the second distributed representation.

[0026] According to the eighteenth aspect of the present disclosure, it is possible to improve the accuracy of analysis when analyzing a document that includes text information describing a substance.

[0027] A 19th aspect of the present disclosure is a processing program according to the 18th aspect, which causes a control unit (200) to execute a third output step of inputting a first distributed representation based on structural information of a substance into a trained model, thereby outputting a token indicating the use or function of the substance.

[0028] A twentieth aspect of the present disclosure is a processing program according to the eighteenth aspect, which causes a control unit (200) to execute a third output step of inputting a second distributed representation based on text information describing the use or function of a substance into a trained model, thereby outputting a token indicating the substance name of the substance.

[0029] FIG. 1 is a first diagram showing an example of the configuration of a system including a processing device. FIG. 2 is a diagram showing an example of the hardware configuration of the processing device. FIG. 3 is a first diagram showing an example of the functional configuration of the processing device. FIG. 4 is a first diagram showing an example of a database stored in a structure information storage unit. FIG. 5 is a flowchart showing the flow of processing by a first learning unit. FIG. 6 is a diagram showing a specific example of processing by the first learning unit. FIG. 7A is a first flowchart showing the flow of processing by a second learning unit. FIG. 7B is a second flowchart showing the flow of processing by the second learning unit. FIG. 8A is a first diagram showing a specific example of processing by the second learning unit. FIG. 8B is a second diagram showing a specific example of processing by the second learning unit. FIG. 9A is a first flowchart showing the flow of processing by an analysis unit. FIG. 9B is a second flowchart showing the flow of processing by the analysis unit. FIG. 10A is a first diagram showing a specific example of processing by the analysis unit. FIG. 10B is a second diagram showing a specific example of processing by the analysis unit. FIG. 11 is a second diagram showing an example of the functional configuration of a processing device. Fig. 12A is a third flowchart showing the flow of processing by the second learning unit. Fig. 12B is a fourth flowchart showing the flow of processing by the second learning unit. Fig. 13 is a second diagram showing an example of a database stored in the structure information storage unit. Fig. 14 is a third diagram showing a specific example of processing by the second learning unit. Fig. 15 is a second diagram showing an example of the configuration of a system including a processing device. Fig. 16 is a third diagram showing an example of the functional configuration of the processing device. Fig. 17 is a second flowchart showing the flow of processing by the analysis unit.

[0030] Hereinafter, each embodiment will be described with reference to the accompanying drawings. In this specification and drawings, components having substantially the same functional configuration are designated by the same reference numerals, and redundant description will be omitted.

[0031] [First embodiment] <System including processing device> An example of the configuration of a system including a processing device according to a first embodiment will be described. Fig. 1 is a first diagram showing an example of the configuration of a system including a processing device. As shown in Fig. 1, the system 10 includes a processing device 100 according to the first embodiment and a server device 110. In the system 10, the processing device 100 and the server device 110 are connected via a network 130.

[0032] The server device 110 functions as a data providing unit 111, and provides the document data stored in the document data storage unit 112 to the processing device 100 in response to a request from the processing device 100. The document data stored in the document data storage unit 112 is, for example, various document data such as papers, magazines, patent documents, etc. in the field of chemistry, and includes structural information indicating the molecular structure of a substance and text information describing the substance.

[0033] The processing device 100 is operated by a user 101, and functions as a learning device in the learning phase and as an analyzing device in the analysis phase. The processing device 100 has a machine learning model (hereinafter simply referred to as a model) for analyzing document data.

[0034] In the learning phase, the processing device 100 learns a model using, for example, document data provided from the server device 110. Note that learning a model here includes learning an untrained model and learning (relearning) an already trained model.

[0035] In the analysis phase, the processing device 100 analyzes, for example, document data using the trained model that has been retrained, and outputs the analysis results.

[0036] <Hardware Configuration of Processing Device> The hardware configuration of the processing device 100 will be described. Fig. 2 is a diagram showing an example of the hardware configuration of the processing device. As shown in Fig. 2, the processing device 100 has a processor 201, a memory 202, an auxiliary storage device 203, an I / F (Interface) device 204, a communication device 205, and a drive device 206. Note that the respective hardware components of the processing device 100 are connected to each other via a bus 207.

[0037] The processor 201 has various arithmetic devices such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), etc. The processor 201 reads various programs (for example, processing programs) into the memory 202 and executes them.

[0038] The memory 202 has a main storage device such as a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The processor 201 and the memory 202 form a so-called computer (also referred to as a "control unit 200"), and the processor 201 executes various programs read onto the memory 202, causing the computer to realize various functions.

[0039] The auxiliary storage device 203 stores various programs and various information used when the processor 201 executes the various programs.

[0040] The I / F device 204 connects the processing device 100 to an operation device 211 that inputs instructions from the user 101 and a display device 212 that displays the analysis results to the user 101 .

[0041] The communication device 205 is a device that is connected to the network 130 and performs communication processing with the server device 110 and the like.

[0042] The drive device 206 is a device for loading a recording medium 213. The recording medium 213 here includes media that record information optically, electrically, or magnetically, such as a CD-ROM, a flexible disk, a magneto-optical disk, etc. The recording medium 213 may also include semiconductor memory that records information electrically, such as a ROM, a flash memory, etc.

[0043] The various programs to be installed in the auxiliary storage device 203 are installed, for example, by setting the distributed recording medium 213 in the drive device 206 and reading the various programs recorded on the recording medium 213 by the drive device 206. Alternatively, the various programs to be installed in the auxiliary storage device 203 may be installed by being downloaded from the network 130 via the communication device 205.

[0044] <Functional Configuration of Processing Device> The functional configuration of the processing device 100 will be described. Fig. 3 is a first diagram showing an example of the functional configuration of the processing device. As described above, a processing program is installed in the processing device 100, and by executing the processing program, the processing device 100 functions as a data receiving unit 310, a first learning unit 320, a second learning unit 330, and an analyzing unit 340. The processing device 100 also functions as a first distributed representation generation unit 360, a second distributed representation generation unit 370, and a model execution unit 300.

[0045] (1) Data Receiving Unit The data receiving unit 310 receives document data provided by the server device 110. The data receiving unit 310 notifies the first learning unit 320 or the second learning unit 330 of the received document data.

[0046] (2) First Learning Unit In the learning phase, the first learning unit 320 generates first learning data based on the document data notified by the data receiving unit 310 and stores it in the first learning data storage unit 321.

[0047] Specifically, the first learning unit 320 identifies the substance name contained in the document data, and for the identified substance name, identifies the structural information of the substance having that substance name by referring to the database stored in the structural information storage unit 350.

[0048] The first learning unit 320 converts the identified structural information into a string of characters that conforms to predetermined rules, and then inputs it into the first distributed representation generation unit 360, thereby obtaining the first distributed representation output by the first distributed representation generation unit 360.

[0049] The first learning unit 320 inputs text information contained in the document data, which describes a substance having the above-mentioned substance name, into the second distributed representation generation unit 370, thereby obtaining the second distributed representation output from the second distributed representation generation unit 370.

[0050] The first learning unit 320 stores a set of distributed representations including the acquired first distributed representation and the acquired second distributed representation in the first training data storage unit 321 as first training data.

[0051] In addition, when the text information is a sentence containing the above-mentioned substance name, the first distributed representation is placed at a position corresponding to the position of the substance name in the sentence, and the second distributed representation is placed at a position corresponding to the position of the part of the sentence other than the substance name. In this way, the first learning unit 320 can acquire a set of distributed representations corresponding to the word order of the original sentence.

[0052] In the learning phase, the first learning unit 320 learns the model possessed by the model execution unit 300 by inputting a set of distributed representations, which are the first learning data stored in the first learning data storage unit 321, into the model execution unit 300. At this time, the first learning unit 320 learns the model through pre-learning using a large set of sentences, thereby generating a general-purpose trained model.

[0053] As a result, according to the trained model trained by the first training unit 320, when analyzing structural information indicating the molecular structure of a substance and text information describing the substance (e.g., its use, function, etc.), it is no longer necessary to analyze the structural information and the text information separately, as was done conventionally. In other words, conventionally, structural information indicating the molecular structure of a substance and text information describing the substance (e.g., its use, function, etc.) were treated as a single entity without any correlation. In contrast, according to the trained model, it is possible to perform analysis by taking into account, for example, information on the semantic similarity of use or function and information on the structural similarity of substances.

[0054] As a result, according to the trained model, when analyzing a document containing text information describing a substance, it is possible to input a first distributed representation based on structural information and a second distributed representation based on text information, thereby improving the accuracy of analysis.

[0055] (3) Second Learning Unit In the learning phase, the second learning unit 330 performs masking on part of the text information included in the document data notified by the data receiving unit 310, and generates second learning data in which the input data and the correct answer data are associated with each other. The second learning unit 330 stores the generated second learning data in the second learning data storage unit 331.

[0056] Specifically, in order to be able to analyze the use, function, etc. of a substance having the substance name from the substance name, the second learning unit 330 performs a masking process on part of the text information other than the substance name (the part describing the substance's use, function, etc.).

[0057] The second learning unit 330 identifies structural information of a substance having a substance name included in the text information by referring to a database stored in the structural information storage unit 350. The second learning unit 330 converts the identified structural information into a character string that follows a predetermined rule, and then inputs the character string to the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0058] The second learning unit 330 acquires second distributed representations by inputting the parts of the text information other than the substance names (including the masked parts) to the second distributed representation generation unit 370. Note that the second learning unit 330 uses a set of distributed representations including the acquired first distributed representation and the acquired second distributed representation as input data.

[0059] The second learning unit 330 identifies a part of the text information other than the substance name (the part describing the substance's use, function, etc.). The second learning unit 330 sets the identified part (the part describing the substance's use, function, etc.) as correct answer data.

[0060] The second learning unit 330 stores second learning data in which the input data and the correct answer data are associated with each other in the second learning data storage unit 331 .

[0061] Furthermore, in order to be able to analyze the substance name from the substance's use, function, etc., the second learning unit 330 performs a masking process on the substance name in the text information using tokens that are common to the substance name.

[0062] The second learning unit 330 inputs tokens common to the substance names to the first distributed representation generation unit 360, thereby acquiring the first distributed representations output from the first distributed representation generation unit 360. The second learning unit 330 inputs the parts of the text information other than the substance names to the second distributed representation generation unit 370, thereby acquiring the second distributed representations output from the second distributed representation generation unit 370. The second learning unit 330 receives as input data a set of distributed representations including the acquired first distributed representation and the acquired second distributed representation.

[0063] The second learning unit 330 identifies the substance name contained in the text information and sets the identified substance name as correct answer data.

[0064] The second learning unit 330 stores second learning data in which the input data and the correct answer data are associated with each other in the second learning data storage unit 331 .

[0065] In the learning phase, the second learning unit 330 inputs the second learning data stored in the second learning data storage unit 331 to the model execution unit 300, thereby re-learning the trained model (general-purpose trained model) held by the model execution unit 300. In this way, the second learning unit 330 re-learns the general-purpose trained model through supervised learning assuming various analysis purposes. In this case, the second learning unit 330 uses the first distributed representation and the second distributed representation.

[0066] As a result, according to the trained model retrained by the second training unit 330, when analyzing structural information indicating the molecular structure of a substance and text information describing the substance (such as its use and function), it is no longer necessary to analyze the structural information and the text information separately, as was done conventionally. In other words, conventionally, structural information indicating the molecular structure of a substance and text information describing the substance (such as its use and function) were treated as a single entity without any correlation. In contrast, according to the trained model, it is possible to perform analysis by taking into account, for example, information on the semantic similarity of use or function and information on the structural similarity of substances.

[0067] As a result, according to the trained model, when analyzing a document containing text information describing a substance, the analysis accuracy can be improved by inputting a first distributed representation based on structural information and a second distributed representation based on text information.

[0068] (4) Analysis Unit In the analysis phase, the analysis unit 340 generates analysis data based on the text information input by the user 101, and inputs the generated analysis data to the model execution unit 300, which has a trained model that has been retrained by the second learning unit 330. As a result, the analysis unit 340 obtains the analysis results from the model execution unit 300 and outputs them to the user 101.

[0069] Specifically, the user 101 generates text information that includes, for example, a substance name and masks the portion where the use, function, etc. of the substance having the substance name should be described, and inputs the generated text information to the analysis unit 340 .

[0070] The analysis unit 340 identifies structural information of a substance having a substance name included in the text information by referring to a database stored in the structural information storage unit 350. The analysis unit 340 converts the identified structural information into a character string according to a predetermined rule, and then inputs the character string to the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0071] The analysis unit 340 inputs the part of the text information other than the substance name into the second distributed representation generation unit 370, thereby acquiring the second distributed representation output from the second distributed representation generation unit 370. The second learning unit 330 uses a set of distributed representations including the acquired first distributed representation and the acquired second distributed representation as analysis data.

[0072] The analysis unit 340 inputs the analysis data into the model execution unit 300 , obtains tokens indicating the use, function, etc. of the substance as the analysis results, and outputs them to the user 101 .

[0073] Furthermore, the user 101 generates text information that includes, for example, the use, function, etc. of a substance and in which the portion where the substance name should be written is set as a token common to the substance name, and inputs this to the analysis unit 340 .

[0074] The analysis unit 340 inputs tokens common to the substance names included in the text information to the first distributed representation generation unit 360 , thereby acquiring the first distributed representations output from the first distributed representation generation unit 360 .

[0075] The analysis unit 340 inputs the portion of the text information other than the tokens common to the substance names to the second distributed representation generation unit 370, thereby acquiring the second distributed representations output from the second distributed representation generation unit 370. The second learning unit 330 uses a set of distributed representations including the acquired first distributed representation and the acquired second distributed representation as analysis data.

[0076] The analysis unit 340 inputs the analysis data into the model execution unit 300 to obtain tokens indicating substance names as analysis results, and outputs them to the user 101 .

[0077] (5) First Distributed Representation Generation Unit When the first distributed representation generation unit 360 receives converted structural information, which has been converted into a character string conforming to a predetermined rule, from the first learning unit 320 or the second learning unit 330 during the learning phase, the first distributed representation generation unit 360 generates and outputs first distributed representations from the converted structural information. Similarly, when the first distributed representation generation unit 360 receives converted structural information, which has been converted into a character string conforming to a predetermined rule, from the analysis unit 340 during the analysis phase, the first distributed representation generation unit 360 generates and outputs first distributed representations from the converted structural information.

[0078] Note that, examples of structural information converted into a character string conforming to a predetermined rule include SMILES information converted into a Simplified Molecular Input Line Entry System (SMILES) representation. Alternatively, examples of structural information converted into a character string conforming to a predetermined rule include identifier information converted using the InChI (International Chemical Identifier) ​​algorithm. However, in the following description of this embodiment, SMILES information is used as the converted structural information.

[0079] The first distributed representation generation unit 360 receives, for example, SMILES information acquired by referring to the structural information storage unit 350 (described later). The first distributed representation generation unit 360 generates and outputs first distributed representations that vectorize the SMILES information, and includes a Grapher as a trained first model.

[0080] The SMILES information input to the first distributed representation generation unit 360 may be the SMILES information itself, or it may be a graph generated based on the SMILES information (a graph in which molecules are nodes and bonds between molecules are edges).

[0081] There are various steps that the first distributed representation generation unit 360 takes to generate a first distributed representation based on structural information of a substance, and the above-mentioned flow of "substance name (or structural information) → SMILES information → (graph) → first distributed representation" is just one example.

[0082] (6) Second Distributed Representation Generation Unit When text information describing a substance (portion other than the substance name) is input from the first learning unit 320 or the second learning unit 330 during the learning phase, the second distributed representation generation unit 370 generates and outputs a second distributed representation from the text information. Similarly, when text information describing a substance (portion other than the substance name) is input from the analysis unit 340 during the analysis phase, the second distributed representation generation unit 370 generates and outputs a second distributed representation from the text information.

[0083] The second distributed representation generation unit 370 generates and outputs second distributed representations obtained by vectorizing text information, and includes BERT as a second model. BERT is an abbreviation for Bidirectional Encoder Representations from Transformers. However, the second model included in the second distributed representation generation unit 370 is not limited to BERT, and may be a model corresponding to document data used for training or relearning the model included in the model execution unit 300. Alternatively, the second model included in the second distributed representation generation unit 370 may be a model corresponding to the domain in which the trained model included in the model execution unit 300 is used. For example, the second distributed representation generation unit 370 may include SciBERT as the second model.

[0084] (7) Model Execution Unit: In the learning phase, when the model execution unit 300 receives first learning data from the first learning unit 320, it uses the first learning data to learn a model and generate a general-purpose trained model.

[0085] When the model execution unit 300 receives second learning data from the second learning unit 330 in the learning phase, the model execution unit 300 re-learns the general-purpose trained model through supervised learning using the second learning data.

[0086] When the model execution unit 300 receives analysis data from the analysis unit 340 in the analysis phase, the model execution unit 300 outputs the analysis results.

[0087] The model execution unit 300 is a model that is learned or re-learned by the first learning unit 320 and the second learning unit 330, and has a BERT as a model that outputs analysis results when analysis data is input by the analysis unit 340.

[0088] <Database Stored in Structural Information Storage Unit> A specific example of the database stored in the structural information storage unit 350 will be described. Fig. 4 is a first diagram showing an example of the database stored in the structural information storage unit. As shown in Fig. 4, the database 400 includes information items such as "substance name", "structural information", and "SMILES information".

[0089] The "Substance Name" field stores the name of a substance. The name of a substance stored in the "Substance Name" field may be a common name or a commonly used name. Alternatively, it may be a conventional identification number that identifies a substance, such as a CAS number.

[0090] "Structural information" stores structural information indicating the molecular structure of a substance having the substance name stored in "Substance name."

[0091] "SMILE information" stores SMILE information obtained by converting the structural information stored in "structural information" into a SMILE expression.

[0092] 4 illustrates the database 400 in which "substance names", "structural information", and "SMILES information" are associated with each other, but the databases stored in the structural information storage unit are not limited to this. For example, the structural information storage unit 350 may separately store a database in which "substance names" are associated with "structural information", and a database in which "structural information" is associated with "SMILES information".

[0093] <Details of Processing by Each Unit of the Processing Device> Details of processing by each unit of the processing device 100 (here, the first learning unit 320, the second learning unit 330, and the analysis unit 340) will be described.

[0094] (1) Details of Processing by the First Learning Unit 320 (1-1) Flow of Processing by the First Learning Unit 320 FIG. 5 is a flowchart showing the flow of processing by the first learning unit.

[0095] In step S501, the first learning unit 320 extracts text information to be used for learning from the document data.

[0096] In step S502, the first learning unit 320 identifies structural information for the substance names included in the document data.

[0097] In step S503, the first learning unit 320 converts the identified structural information into SMILES information and inputs it to the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0098] In step S504, the first learning unit 320 obtains a second distributed representation by inputting text information (text information that does not include a substance name, or the part of text information that includes a substance name other than the substance name) into the second distributed representation generation unit 370.

[0099] In step S505, the first learning unit 320 stores a set of distributed representations including the acquired first distributed representation and second distributed representation in the first training data storage unit 321 as first training data.

[0100] In step S506, the first learning unit 320 determines whether a predetermined amount of first learning data has been stored in the first learning data storage unit 321. If it is determined in step S506 that the predetermined amount of first learning data has not been stored (NO in step S506), the process returns to step S501. On the other hand, if it is determined in step S506 that the predetermined amount of first learning data has been stored (YES in step S506), the process proceeds to step S507.

[0101] In step S507, the first learning unit 320 uses the first learning data to learn a model and generate a general-purpose trained model.

[0102] (1-2) Specific Example of Processing by the First Learning Unit 320 Fig. 6 is a diagram showing a specific example of processing by the first learning unit 320. In Fig. 6, reference numeral 610 denotes an example of text information extracted from document data by the first learning unit 320 as text information to be used for learning.

[0103] In FIG. 6, reference numeral 620 denotes an example of a token generated by the first learning unit 320 based on the extracted text information.

[0104] As shown in Figure 6, the first learning unit 320 identifies the substance name ("Acetone") in the extracted text information, and by referring to the database 400, identifies structural information indicating the molecular structure of "Acetone".

[0105] As shown in FIG. 6, the first learning unit 320 generates SMILES information by converting the structural information indicating the identified molecular structure of "Acetone" into a SMILES expression.

[0106] As shown in FIG. 6 , the first learning unit 320 inputs the generated SMILES information to the first distributed representation generation unit 360 , thereby acquiring first distributed representations 631 from the first distributed representation generation unit 360 .

[0107] As shown in Figure 6, the first learning unit 320 inputs tokens other than the substance name from the extracted text information into the second distributed representation generation unit 370, thereby obtaining second distributed representations 632 to 638 output from the second distributed representation generation unit 370.

[0108] As a result, the first training unit 320 can generate, as first training data, a set of embedded representations including the acquired first embedded representation 631 and second embedded representations 632 to 638. Note that, as shown in the example of Fig. 6, when the text information (reference numeral 610) is a sentence including a substance name, the first training unit 320 generates a set of embedded representations corresponding to the word order of the original sentence, in which: the first embedded representation 631 is placed at a position corresponding to the position of the substance name, and the second embedded representations 632 to 638 are placed at positions corresponding to the positions of parts other than the substance name.

[0109] (2) Details of processing by the second learning unit 330 (2-1-1) Processing flow 1 by the second learning unit 330 Figure 7A is a first flowchart showing the processing flow by the second learning unit, and is a flowchart for re-learning a general-purpose trained model so that the use, function, etc. of a substance having the substance name can be analyzed from the substance name.

[0110] In step S701, the second learning unit 330 extracts text information including substance names from the document data as text information to be used for learning.

[0111] In step S702, the second learning unit 330 performs masking on a portion of the text information other than the substance name (the portion describing the substance's use, function, etc.).

[0112] In step S703, the second learning unit 330 identifies structural information for the substance name included in the text information.

[0113] In step S704, the second learning unit 330 converts the identified structural information into SMILES information and inputs it to the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0114] In step S705, the second learning unit 330 inputs the text information (the part other than the substance name, which is partially masked) into the second distributed representation generation unit 370, thereby obtaining the second distributed representation output from the second distributed representation generation unit 370.

[0115] In step S706, the second learning unit 330 identifies a part of the text information other than the substance name (the part describing the substance's use, function, etc.).

[0116] In step S707, the second learning unit 330 uses a set of distributed representations including the first distributed representation obtained in step S704 and the second distributed representation obtained in step S705 as input data. The second learning unit 330 uses the portion identified in step S706 (the portion describing the substance's use, function, etc.) as correct answer data. Furthermore, the second learning unit 330 stores second training data in which the input data and the correct answer data are associated with each other in the second training data storage unit 331.

[0117] In step S708, the second learning unit 330 determines whether a predetermined amount of second learning data has been stored in the second learning data storage unit 331. If it is determined in step S708 that the predetermined amount of second learning data has not been stored (NO in step S708), the process returns to step S701. On the other hand, if it is determined in step S708 that the predetermined amount of second learning data has been stored (YES in step S708), the process proceeds to step S709.

[0118] In step S709, the second learning unit 330 performs supervised learning using the second learning data, and re-learns the general-purpose trained model.

[0119] (2-1-2) Processing flow 2 by the second learning unit 330 Figure 7B is a second flowchart showing the processing flow by the second learning unit, and is a flowchart for re-learning a general-purpose trained model so that the substance name of the substance can be analyzed from the substance's use, function, etc.

[0120] In step S711, the second learning unit 330 extracts text information including substance names from the document data as text information to be used for learning.

[0121] In step S712, the second learning unit 330 performs masking processing on the substance names included in the text information using tokens that are common to the substance names.

[0122] In step S713, the second learning unit 330 inputs tokens common to the substance names to the first distributed representation generation unit 360, thereby acquiring the first distributed representations output from the first distributed representation generation unit 360.

[0123] In step S714, the second learning unit 330 inputs the part of the text information other than the substance name into the second distributed representation generation unit 370, thereby obtaining the second distributed representation output from the second distributed representation generation unit 370.

[0124] In step S715, the second learning unit 330 identifies the substance name contained in the text information.

[0125] In step S716, the second learning unit 330 sets as input data a set of distributed representations including the first distributed representation obtained in step S713 and the second distributed representation obtained in step S714. The second learning unit 330 sets as supervised data the substance name identified in step S715. Furthermore, the second learning unit 330 stores second training data in which the input data and the supervised data are associated with each other in the second training data storage unit 331.

[0126] In step S717, the second learning unit 330 determines whether a predetermined amount of second learning data has been stored in the second learning data storage unit 331. If it is determined in step S717 that the predetermined amount of second learning data has not been stored (NO in step S717), the process returns to step S711. On the other hand, if it is determined in step S717 that the predetermined amount of second learning data has been stored (YES in step S717), the process proceeds to step S718.

[0127] In step S718, the second learning unit 330 performs supervised learning using the second learning data, and re-learns the general-purpose trained model.

[0128] (2-2-1) Specific Example 1 of Processing by the Second Learning Unit 330 Fig. 8A is a first diagram showing a specific example of processing by the second learning unit 330. In Fig. 8A, reference numeral 810 denotes an example of text information extracted from document data by the second learning unit 330 as text information to be used for learning.

[0129] In FIG. 8A, reference numeral 820 denotes an example of text information after masking, in which part of the extracted text information other than the substance name ("additive") is masked.

[0130] In FIG. 8A, reference numeral 830 denotes an example of a token generated by the second learning unit 330 based on the text information after the masking process.

[0131] As shown in Figure 8A, the second learning unit 330 identifies the substance name ("1-butanol") in the extracted text information and identifies structural information indicating the molecular structure of "1-butanol" by referring to the database 400.

[0132] As shown in FIG. 8A, the second learning unit 330 converts the structural information indicating the molecular structure of the identified "1-butanol" into a SMILES expression by referring to the database 400, and generates SMILES information.

[0133] As shown in FIG. 8A , the second learning unit 330 inputs the generated SMILES information to the first distributed representation generation unit 360 to obtain the first distributed representation 841 output from the first distributed representation generation unit 360 .

[0134] 8A, the second learning unit 330 inputs tokens for the portion of the extracted text information other than the substance name to the second distributed representation generation unit 370, thereby acquiring second distributed representations 842 to 849 output from the second distributed representation generation unit 370. Note that the second distributed representation 843 output from the second distributed representation generation unit 370 by inputting some of the masked tokens ("[MASK]") to the second distributed representation generation unit 370 is assumed to be blank.

[0135] The second learning unit 330 uses a set of distributed representations including the acquired first distributed representation 841 and second distributed representations 842 to 849 as input data.

[0136] 8A, the second learning unit 330 identifies a part ("additive") of the extracted text information other than the substance name. The second learning unit 330 sets the identified part ("additive") as the correct answer data.

[0137] This allows the second learning unit 330 to generate second learning data in which the input data and the correct answer data are associated with each other.

[0138] (2-2-2) Specific Example 2 of Processing by the Second Learning Unit 330 Fig. 8B is a second diagram showing a specific example of processing by the second learning unit 330. In Fig. 8B, reference numeral 810 denotes an example of text information extracted from document data by the second learning unit 330 as text information to be used for learning.

[0139] 8B, reference numeral 821 denotes an example of text information after masking. The second learning unit 330 identifies a substance name ("1-butanol") in the extracted text information. The second learning unit 330 then masks the identified substance name ("1-butanol") using a token ("[CHEM]") that is common to substance names, thereby generating text information after masking.

[0140] In FIG. 8B, reference numeral 831 denotes an example of a token generated by the second learning unit 330 based on the text information after the masking process.

[0141] As shown in Figure 8B, the second learning unit 330 obtains the first distributed representation 861 output from the first distributed representation generation unit 360 by inputting a token common to the substance name ("[CHEM]") into the first distributed representation generation unit 360.

[0142] As shown in Figure 8B, the second learning unit 330 inputs tokens other than the substance name from the extracted text information into the second distributed representation generation unit 370, thereby obtaining second distributed representations 862 to 869 output from the second distributed representation generation unit 370.

[0143] The second learning unit 330 uses a set of distributed representations including the acquired first distributed representation 861 and second distributed representations 862 to 869 as input data.

[0144] As shown in FIG. 8B, the second learning unit 330 identifies the substance name (“1-butanol”) contained in the extracted text information and sets it as correct answer data.

[0145] This allows the second learning unit 330 to generate second learning data in which the input data and the correct answer data are associated with each other.

[0146] (3) Details of Processing by Analysis Unit 340 (3-1-1) Processing Flow 1 by Analysis Unit 340 Figure 9A is a first flowchart showing the processing flow by the analysis unit, and is a flowchart for analyzing the use, function, etc. of a substance having a substance name from the substance name.

[0147] In step S901, the analysis unit 340 accepts input of text information generated by the user 101. It is assumed that the text information generated by the user 101 in step S901 is text information that includes a substance name and has been masked to remove portions that should describe the use, function, etc. of the substance having the substance name.

[0148] In step S902, the analysis unit 340 identifies structural information for the substance name included in the text information.

[0149] In step S903, the analysis unit 340 converts the identified structural information into SMILES information and inputs it to the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0150] In step S904, the analysis unit 340 inputs the text information (the part other than the substance name, which is partially masked) into the second distributed representation generation unit 370, thereby obtaining the second distributed representation output from the second distributed representation generation unit 370.

[0151] In step S905, the analysis unit 340 sets a set of distributed representations including the first distributed representation obtained in step S903 and the second distributed representation obtained in step S904 as analysis data. The analysis unit 340 inputs the analysis data into the trained model that has been retrained and is held by the model execution unit 300. As a result, the analysis unit 340 obtains tokens that indicate the use, function, etc. of the substance as analysis results.

[0152] In step S906, the analysis unit 340 outputs the acquired use, function, etc. of the substance to the user 101.

[0153] In step S907, the analysis unit 340 determines whether or not to continue the process, and if it is determined that the process should be continued (YES in step S907), the process returns to step S901.

[0154] On the other hand, if it is determined in step S907 that the processing is not to be continued (NO in step S907), the processing by the analysis unit 340 ends.

[0155] (3-1-2) Processing Flow 2 by the Analysis Unit 340 FIG. 9B is a second flowchart showing the processing flow by the analysis unit, which is a flowchart for analyzing the substance name of a substance from the substance's use, function, etc.

[0156] In step S911, the analysis unit 340 accepts input of text information generated by the user 101. The text information generated by the user 101 in step S911 is assumed to be text information in which a token common to substance names ("[CHEM]") is written in the portion where the substance name should be written. The text information generated by the user 101 in step S911 is assumed to be text information in which the use, function, etc. of the substance having the substance name is written in the portion other than the portion where the substance name should be written.

[0157] In step S912, the analysis unit 340 inputs a token (“[CHEM]”) common to the substance names contained in the text information into the first distributed representation generation unit 360, thereby obtaining the first distributed representation output from the first distributed representation generation unit 360.

[0158] In step S913, the analysis unit 340 inputs the parts of the text information other than the token ("[CHEM]") common to the substance names to the second distributed representation generation unit 370, thereby obtaining the second distributed representation output from the second distributed representation generation unit 370.

[0159] In step S914, the analysis unit 340 sets a set of distributed representations including the first distributed representation obtained in step S912 and the second distributed representation obtained in step S913 as analysis data. The analysis unit 340 inputs the analysis data into the trained model that has been retrained and is held by the model execution unit 300. As a result, the analysis unit 340 obtains a token indicating a substance name as the analysis result.

[0160] In step S915, the analysis unit 340 outputs the substance name identified by the acquired analysis result to the user 101.

[0161] In step S916, the analysis unit 340 determines whether or not to continue the process, and if it determines to continue the process (YES in step S916), the process returns to step S911.

[0162] On the other hand, if it is determined in step S916 that the processing is not to be continued (NO in step S916), the processing by the analysis unit 340 ends.

[0163] (3-2-1) Specific Example 1 of Processing by the Analysis Unit 340 Fig. 10A is a first diagram showing a specific example of processing by the analysis unit. In Fig. 10A, reference numeral 1010 denotes text information generated by the user 101. As indicated by reference numeral 1010, the text information generated by the user 101 includes a substance name, and masking has been performed on the portion where the use of the substance having that substance name should be described.

[0164] In FIG. 10A, reference numeral 1020 denotes an example of a token generated by the analysis unit 340 based on text information.

[0165] As shown in FIG. 10A, the analysis unit 340 identifies the substance name ("Propylene carbonate") in the text information, and by referring to the database 400, identifies the structural information indicating the molecular structure of "Propylene carbonate."

[0166] As shown in FIG. 10A, the analysis unit 340 generates SMILES information by converting the structural information indicating the molecular structure of the identified "Propylene carbonate" into a SMILES expression.

[0167] As shown in FIG. 10A , the analysis unit 340 inputs the generated SMILES information to the first distributed representation generation unit 360 , thereby acquiring the first distributed representation 1031 output from the first distributed representation generation unit 360 .

[0168] 10A , the analysis unit 340 acquires second distributed representations 1032 to 1037 by inputting tokens other than the substance name from the text information to the second distributed representation generation unit 370. Note that, as a result of inputting a masked portion of the tokens ("[MASK]") to the second distributed representation generation unit 370, the second distributed representation 1035 output from the second distributed representation generation unit 370 is blank.

[0169] As a result, the analysis unit 340 can generate a set of distributed representations including the first distributed representation 1031 and the second distributed representations 1032 to 1037 as analysis data. The analysis unit 340 inputs the analysis data into the trained model that has been retrained and is held by the model execution unit 300. As a result, the analysis unit 340 obtains, for example, "electrolyte" as an analysis result, and by outputting the analysis result, the user 101 can recognize the use of propylene carbonate (that it is used in an electrolyte).

[0170] (3-2-2) Specific Example 2 of Processing by the Analysis Unit 340 FIG. 10B is a second diagram showing a specific example of processing by the analysis unit. In FIG. 10B, reference numeral 1011 denotes text information generated by the user 101. As indicated by reference numeral 1011, the text information generated by the user 101 includes a token ("[CHEM]") common to substance names in a portion where the substance name should be written. As indicated by reference numeral 1011, the text information generated by the user 101 includes information about the use, function, etc. of a substance having the substance name in a portion other than the portion where the substance name should be written.

[0171] In FIG. 10B, reference numeral 1021 denotes an example of a token generated by the analysis unit 340 based on the text information.

[0172] 10B , the analysis unit 340 inputs a token common to the substance names ("[CHEM]") to the first distributed representation generation unit 360, thereby obtaining a first distributed representation 1041 output from the first distributed representation generation unit 360. The analysis unit 340 inputs tokens other than the token common to the substance names ("[CHEM]") to the second distributed representation generation unit 370, thereby obtaining second distributed representations 1042 to 1046 output from the second distributed representation generation unit 370.

[0173] As a result, the analysis unit 340 can generate a set of distributed representations including the first distributed representation 1041 and the second distributed representations 1042 to 1046 as analysis data. The analysis unit 340 inputs the analysis data into the trained model that has been retrained and is held by the model execution unit 300. As a result, the analysis unit 340 obtains, for example, "propylene carbonate" as the analysis result, and by outputting the analysis result, the user 101 can recognize the substance name (for example, propylene carbonate) of the substance used in the electrolyte solution.

[0174] <Summary> As is clear from the above description, the processing method according to the first embodiment includes: a first output step of outputting first distributed representations based on structural information of a substance having a substance name included in the document data; a second output step of outputting second distributed representations based on text information describing a substance having a substance name included in the document data; and a learning step of learning a model using the first distributed representations and the second distributed representations. Note that learning a model here includes relearning a model that has already been learned.

[0175] In this way, the processing method according to the first embodiment uses the first distributed representation and the second distributed representation when, for example, training a model through pre-training using a large set of sentences to generate a general-purpose trained model, or when relearning a general-purpose trained model through supervised learning intended for various analysis purposes.

[0176] As a result, when analyzing document data containing text information describing substances, the trained model trained (or retrained) by the processing method according to the first embodiment: - It is no longer necessary to analyze structural information and text information separately, as has been done in the past. - In other words, in the past, structural information indicating the molecular structure of a substance and text information describing the substance (e.g., its use, function, etc.) were treated as a single entity, without any correlation between them. In contrast, the trained model makes it possible to perform analysis by taking into account, for example, information on the semantic similarity of the use or function and information on the structural similarity of the substances.

[0177] As a result, according to the trained model, when analyzing a document containing text information describing a substance, it is possible to input a first distributed representation based on structural information and a second distributed representation based on text information, thereby improving the accuracy of analysis.

[0178] Second Embodiment In the above-described first embodiment, the second learning unit 330 performs relearning on the trained model held by the model executing unit 300. However, the target of relearning by the second learning unit 330 is not limited to the trained model held by the model executing unit 300.

[0179] For example, the re-learning may be configured to target any one of the trained first model possessed by the first distributed representation generation unit 360, the trained second model possessed by the second distributed representation generation unit 370, or the trained model possessed by the model execution unit 300. The second embodiment will be described below, focusing on the differences from the first embodiment.

[0180] <Functional Configuration of Processing Device> The functional configuration of the processing device 100 according to the second embodiment will be described. Fig. 11 is a second diagram showing an example of the functional configuration of the processing device. Differences from the functional configuration described using Fig. 3 in the first embodiment above include the following: The functions of the second learning unit 1110, the first distributed representation generation unit 1120, the second distributed representation generation unit 1130, the model execution unit 1140, and the analysis unit 1150 in Fig. 11 are different from the functions of the second learning unit 330, the first distributed representation generation unit 360, the second distributed representation generation unit 370, the model execution unit 300, and the analysis unit 340 in Fig. 3.

[0181] (1) Second Learning Unit In the learning phase, the second learning unit 1110 performs masking on part of the text information included in the document data notified by the data receiving unit 310, and generates second learning data in which the input data and the correct answer data are associated with each other. The second learning unit 1110 stores the generated second learning data in the second learning data storage unit 331.

[0182] Specifically, in order to be able to analyze the uses, functions, etc. of a substance having the substance name from the substance name, the second learning unit 1110 performs a masking process on part of the text information other than the substance name (the part describing the substance's uses, functions, etc.).

[0183] The second learning unit 1110 identifies a substance name included in the text information, and for the identified substance name, identifies structural information of the substance having the identified substance name by referring to a database stored in the structural information storage unit 350. The second learning unit 1110 converts the identified structural information into a character string (e.g., a SMILES expression) that follows a predetermined rule.

[0184] The second learning unit 1110 receives as input data text information in which substance names have been converted into SMILES representations and a portion of the text information other than the substance names (portions describing the substance's uses, functions, etc.) has been masked. The second learning unit 1110 receives as input data the portion of the text information other than the substance names (portions describing the substance's uses, functions, etc.).

[0185] The second learning unit 1110 stores second learning data in which the input data and the correct answer data are associated with each other in the second learning data storage unit 331 .

[0186] Furthermore, the second learning unit 1110 performs a masking process on the substance name of the text information using tokens that are common to the substance name so that the substance name can be analyzed from the substance's use, function, etc. The second learning unit 1110 uses the text information that has been masked using tokens that are common to the substance name as input data.

[0187] The second learning unit 1110 identifies the substance name included in the text information and sets the identified substance name as correct answer data.

[0188] The second learning unit 1110 stores second learning data in which the input data and the correct answer data are associated with each other in the second learning data storage unit 331 .

[0189] In the learning phase, the second learning unit 1110 inputs input data of the second learning data stored in the second learning data storage unit 331 to the first distributed representation generation unit 1120 and the second distributed representation generation unit 1130. The second learning unit 1110 inputs the correct answer data of the second learning data stored in the second learning data storage unit 331 to the model execution unit 1140. This allows the second learning unit 1110 to re-learn, through supervised learning, the following: - the trained first model held by the first distributed representation generation unit 1120, - the trained second model held by the second distributed representation generation unit 1130, and - the general-purpose trained model held by the model execution unit 1140.

[0190] Note that by performing relearning through the supervised learning, the trained first model included in the first distributed representation generation unit 1120 is able to output first distributed representations appropriate for the context. Specifically, the trained first model is able to arrange first distributed representations based on structural information of substances with different substance names included in sentences with similar contexts, in close proximity to each other. Note that "two first distributed representations are arranged in close proximity to each other" refers to the positions (distance, orientation, etc.) of the two first distributed representations being close to each other in a space (e.g., a vector space) representing the distributed representations. For example, "two first distributed representations being arranged in close proximity to each other" includes arranging the two first distributed representations in positions that increase the cosine similarity between the two first distributed representations, or arranging the two first distributed representations in positions that reduce the Euclidean distance between the two first distributed representations, etc.

[0191] By performing re-learning through the above-described supervised learning, the trained model possessed by the model execution unit 1140 becomes able to output tokens indicating the use, function, etc. of a substance as the analysis result.

[0192] Alternatively, by performing re-learning using the above-described supervised learning, the trained model possessed by the model execution unit 1140 can output a token indicating the substance name as the analysis result.

[0193] (2) First Distributed Representation Generation Unit When the first distributed representation generation unit 1120 receives SMILES information from the input data from the second learning unit 1110 during the learning phase, the first distributed representation generation unit 1120 generates a first distributed representation from the SMILES information and outputs the first distributed representation to the model execution unit 1140. The first distributed representation generation unit 1120 has Grapher as a trained first model that generates a first distributed representation from the SMILES information. When the first distributed representation generation unit 1120 outputs the first distributed representation to the model execution unit 1140, the model execution unit 1140 calculates a loss and updates the model parameters of Grapher based on the loss. In other words, re-learning is performed on the trained first model.

[0194] In the analysis phase, when the SMILES information of the analysis data is input from the analysis unit 1150 , the first distributed representation generation unit 1120 generates a first distributed representation from the SMILES information and outputs it to the model execution unit 1140 .

[0195] (3) Second Distributed Representation Generation Unit When text information of the input data is input from the second learning unit 1110, the second distributed representation generation unit 1130 generates a second distributed representation from the text information and outputs the second distributed representation to the model execution unit 1140. The second distributed representation generation unit 1130 has a BERT as a trained second model that generates a second distributed representation from the text information. In response to the second distributed representation generation unit 1130 outputting the second distributed representation to the model execution unit 1140, the model execution unit 1140 calculates a loss and updates the model parameters of the BERT based on the loss. In other words, re-learning is performed on the trained second model.

[0196] In the analysis phase, when the second distributed representation generation unit 1130 receives text information from the analysis data from the analysis unit 1150 , the second distributed representation generation unit 1130 generates second distributed representations from the text information and outputs them to the model execution unit 1140 .

[0197] (4) Model Execution Unit In the learning phase, the model execution unit 1140 receives the first distributed representations output from the first distributed representation generation unit 1120 and the second distributed representations output from the second distributed representation generation unit 1130. When the first distributed representations and the second distributed representations are input, the general-purpose trained model outputs output data.

[0198] The model execution unit 1140 has BERT as a general-purpose trained model. When a part of the text information other than the substance name (the part describing the substance's use, function, etc.) is input as correct answer data from the second learning unit 1110, the model execution unit 1140 calculates the loss between: the output data (converted into tokens) output from BERT, and the part of the text information other than the substance name (the part describing the substance's use, function, etc.) input from the second learning unit 1110, and updates, according to the calculated loss, any of the following: the model parameters of the trained first model (Grapher), or the model parameters of the trained second model (BERT), or the model parameters of the trained model (BERT).

[0199] When the model execution unit 1140 receives the name of a substance contained in text information as correct answer data from the second learning unit 1110, it calculates the loss between: - the output data (converted into tokens) output from BERT; and - the name of the substance contained in the text information input from the second learning unit 1110, and updates either: - the model parameters of the trained first model (Grapher), or - the model parameters of the trained second model (BERT), or - the model parameters of the trained model (BERT), depending on the calculated loss.

[0200] In the analysis phase, in response to the analysis unit 1150 inputting analysis data to the first distributed representation generation unit 1120 and the second distributed representation generation unit 1130, the model execution unit 1140: - inputs the first distributed representation output from the first distributed representation generation unit 1120; - inputs the second distributed representation output from the second distributed representation generation unit 1130; and - outputs an analysis result based on the input first distributed representation and second distributed representation.

[0201] (5) Analysis Unit 1150 In the analysis phase, the analysis unit 1150 generates analysis data based on the text information input by the user 101, and inputs the analysis data to the first distributed representation generation unit 1120 and the second distributed representation generation unit 1130. As a result, the analysis unit 1150 obtains the analysis results from the model execution unit 1140 and outputs them to the user 101.

[0202] Specifically, when analyzing the use, function, etc. of a substance having a substance name from the substance name, the user 101 inputs, for example, text information that includes the substance name and masks the portion where the use, function, etc. of the substance having the substance name should be written, to the analysis unit 1150. The analysis unit 1150 accepts the input text information.

[0203] The analysis unit 1150 identifies a substance name included in the text information, and for the identified substance name, identifies structural information of the substance having the substance name by referring to the database stored in the structural information storage unit 350. The analysis unit 1150 converts the identified structural information into a character string (for example, a SMILES expression) that follows a predetermined rule.

[0204] The analysis unit 1150 converts the substance name into a SMILES representation, and generates text information in which the portions that should describe the use, function, etc. of the substance having the substance name are masked, as data for analysis.

[0205] On the other hand, when analyzing the substance name of a substance from its use, function, etc., the user 101, for example, writes a token common to the substance name in the portion where the substance name should be written, and inputs text information describing the use, function, etc. of the substance having the substance name to the analysis unit 1150. The analysis unit 1150 accepts the input text information and uses it as data for analysis.

[0206] The analysis unit 1150 inputs tokens common to the SMILES information or substance names included in the analysis data to the first distributed representation generation unit 1120. The analysis unit 1150 inputs parts other than the SMILES information or parts other than the substance names (including masked parts) included in the analysis data to the second distributed representation generation unit 1130. As a result, the analysis unit 1150 obtains the analysis results from the model execution unit 1140 and outputs them to the user 101.

[0207] <Details of Processing by Each Unit of Processing Apparatus> Details of processing by each unit (here, the second learning unit 1110) of the processing apparatus 100 according to the second embodiment will be described.

[0208] (1) Processing flow 1 by the second learning unit 1110 Figure 12A is a third flowchart showing the processing flow by the second learning unit, and is a flowchart for re-learning a trained model so that the use, function, etc. of a substance having the substance name can be analyzed from the substance name.

[0209] In step S701, the second learning unit 1110 extracts text information including substance names from the document data as text information to be used for learning.

[0210] In step S702, the second learning unit 1110 performs masking on a part of the text information other than the substance name (the part describing the substance's use, function, etc.).

[0211] In step S1201, the second learning unit 1110 identifies structural information of a substance having a substance name included in text information, and converts the structural information into a SMILES representation to generate SMILES information.

[0212] In step S1202, the second learning unit 1110 identifies a part of the text information other than the substance name (the part describing the substance's use, function, etc.).

[0213] In step S1203, the second learning unit 1110 takes as input data the text information that has been subjected to the masking process in step S702 and that includes the SMILES information generated in step S1201. The second learning unit 1110 takes as correct answer data a portion of the text information other than the substance name (a portion describing the substance's use, function, etc.) identified in step S1202. Furthermore, the second learning unit 1110 generates second learning data in which the input data and the correct answer data are associated with each other, and stores the second learning data in the second learning data storage unit 331.

[0214] In step S708, the second learning unit 1110 determines whether a predetermined amount of second learning data has been stored in the second learning data storage unit 331. If it is determined in step S708 that the predetermined amount of second learning data has not been stored (NO in step S708), the process returns to step S701. On the other hand, if it is determined in step S708 that the predetermined amount of second learning data has been stored (YES in step S708), the process proceeds to step S1204.

[0215] In step S1204, the second learning unit 1110 inputs input data of the second training data to the first distributed representation generation unit 1120 and the second distributed representation generation unit 1130, and inputs the correct answer data of the second training data to the model execution unit 1140. As a result, the second learning unit 1110 re-learns either the trained first model held by the first distributed representation generation unit 1120, the trained second model held by the second distributed representation generation unit 1130, or the trained model held by the model execution unit 1140.

[0216] (2) Processing flow 2 by the second learning unit 1110 Figure 12B is a fourth flowchart showing the processing flow by the second learning unit, and is a flowchart for re-learning a trained model so that the substance name of the substance can be analyzed from the substance's use, function, etc.

[0217] In step S711, the second learning unit 1110 extracts text information including substance names from the document data as text information to be used for learning.

[0218] In step S712, the second learning unit 1110 performs masking on the substance names included in the text information using tokens that are common to the substance names.

[0219] In step S1211, the second learning unit 1110 identifies the substance name contained in the text information.

[0220] In step S1212, the second learning unit 1110 uses the text information that was masked in step S712 as input data. The second learning unit 1110 uses the substance names included in the text information identified in step S1211 as correct answer data. Furthermore, the second learning unit 1110 generates second learning data in which the input data and the correct answer data are associated with each other, and stores the second learning data in the second learning data storage unit 331.

[0221] In step S717, the second learning unit 1110 determines whether a predetermined amount of second learning data has been stored in the second learning data storage unit 331. If it is determined in step S717 that the predetermined amount of second learning data has not been stored (NO in step S708), the process returns to step S711. On the other hand, if it is determined in step S717 that the predetermined amount of second learning data has been stored (YES in step S717), the process proceeds to step S1213.

[0222] In step S1213, the second learning unit 1110 inputs the input data of the second training data to the first distributed representation generation unit 1120 and the second distributed representation generation unit 1130, and inputs the correct answer data of the second training data to the model execution unit 1140. As a result, the second learning unit 1110 re-learns either the trained first model held by the first distributed representation generation unit 1120, the trained second model held by the second distributed representation generation unit 1130, or the trained model held by the model execution unit 1140.

[0223] <Summary> As is clear from the above description, the processing method according to the second embodiment is configured so that the outputs of the trained first model and the trained second model are input to the trained model.

[0224] The processing method according to the second embodiment, based on the configuration, performs retraining of either a trained first model that generates a first distributed representation, a trained second model that generates a second distributed representation, or a trained model to which the outputs of the trained first model and the trained second model are input. The trained first model that generates a first distributed representation is retrained so that first distributed representations based on structural information of substances with different substance names contained in sentences with similar contexts are positioned near each other.

[0225] As a result, the processing method according to the second embodiment makes it possible to generate first and second distributed representations according to various analysis purposes. As a result, a trained model that has been retrained using the processing method according to the second embodiment can further improve the analytical accuracy when analyzing documents that include text information describing substances.

[0226] In the first and second embodiments, the case where appropriate SMILES information is generated by converting structural information into a SMILES representation has been described. The appropriate SMILES information here refers to SMILES information that, when input to the first distributed representation generation unit 360, causes the first distributed representation generation unit 360 to output an appropriate first distributed representation.

[0227] However, depending on the substance, when the structural information is converted into a SMILES representation, appropriate SMILES information may not be generated.

[0228] For example, when structural information of substances with substance names = "Graphite" or "Graphene" is converted into a SMILES representation, it becomes a monatomic substance. Similarly, when structural information of substances such as inorganic materials is converted into a SMILES representation, it also becomes a monatomic substance. For this reason, when such substances are input to the first distributed representation generation unit 360, an appropriate first distributed representation is not output by the first distributed representation generation unit 360.

[0229] Materials with abnormal valence (eg, a carbon with five bonds instead of the usual four) also do not generate adequate SMILES information.

[0230] Therefore, in the third embodiment, the processing device 100 performs exceptional processing for such exceptional substances (substances that become monatomic when structural information is converted into a SMILES representation or substances with abnormal valence). Specifically, when the substance name of such a substance is included in the text information, the processing device 100 generates a second distributed representation instead of a first distributed representation. The third embodiment will be described below, focusing on the differences from the first and second embodiments.

[0231] <Database Stored in Structural Information Storage Unit> A specific example of the database stored in the structural information storage unit 350 will be described. FIG. 13 is a second diagram showing an example of the database stored in the structural information storage unit. The difference from the database 400 shown in FIG. 4 is that the database 1300 shown in FIG. 13 includes "Graphene" in the "Substance Name". As shown in FIG. 13, when the substance name is "Graphene", converting the structural information into the SMILES representation results in a single atom ("C"). Therefore, when "Graphene" is included in the text information, the processing device 100 performs exceptional processing.

[0232] <Specific Example of Processing by Second Learning Unit 330> An exceptional process performed in the processing device 100 will be described. Here, an exceptional process performed when the second learning unit 330 generates second learning data will be described. FIG. 14 is a third diagram showing a specific example of processing by the second learning unit. In FIG. 14, reference numeral 1410 denotes an example of text information extracted from document data by the second learning unit 330 as text information to be used for learning.

[0233] In FIG. 14, reference numeral 1420 denotes an example of text information after masking, in which part of the extracted text information other than the substance name ("batteries") is masked.

[0234] In FIG. 14, reference numeral 1430 denotes an example of a token generated by the second learning unit 330 based on the text information after the masking process.

[0235] As shown in FIG. 14, the second learning unit 330 identifies the substance name ("Graphene") in the extracted text information, and by referring to the database 1300, identifies structural information indicating the molecular structure of "Graphene."

[0236] As shown in FIG. 14, the second learning unit 330 converts the structural information indicating the molecular structure of the identified "Graphene" into a SMILES expression by referring to the database 1300, and generates SMILES information.

[0237] Here, the second learning unit 330 determines that exceptional processing should be performed because the generated SMILES information is a monoatomic one. Specifically, as shown in FIG. 14 , the second learning unit 330 inputs the substance name="Graphene" to the second distributed representation generation unit 370 instead of inputting the SMILES information to the first distributed representation generation unit 360. As a result, the second learning unit 330 acquires the second distributed representation 1441 output from the second distributed representation generation unit 370.

[0238] 14, the second learning unit 330 inputs tokens for the portion of the extracted text information other than the substance name to the second distributed representation generation unit 370, thereby acquiring second distributed representations 1442 to 1449 output from the second distributed representation generation unit 370. Note that the second distributed representation 1448 output from the second distributed representation generation unit 370 by inputting some of the masked tokens ("[MASK]") to the second distributed representation generation unit 370 is assumed to be blank.

[0239] The second learning unit 330 uses the acquired second distributed representations 1441 to 1449 as input data.

[0240] 14, the second learning unit 330 identifies a part of the extracted text information other than the substance name ("batteries"). The second learning unit 330 sets the identified part ("batteries") as the correct answer data.

[0241] This allows the second learning unit 330 to generate second learning data in which the input data and the correct answer data are associated with each other.

[0242] <Summary> As is clear from the above explanation, the processing method according to the third embodiment: If a substance having a substance name included in the text information is not an exceptional substance, a first distributed representation based on the structural information of the substance is obtained, and if the substance is an exceptional substance, a second distributed representation based on the substance name of the substance is obtained. Exceptional substances include substances that become monatomic when the structural information is converted into a SMILES representation or substances with anomalous valence. For parts of the text information other than the substance name, a second distributed representation based on the parts is obtained.

[0243] As a result, according to the processing method of the third embodiment, even if the text information contains the substance name of an exceptional substance, the processing device 100 can appropriately execute processing by the first learning unit, processing by the second learning unit, and processing by the analysis unit.

[0244] Fourth Embodiment In the above embodiments, the trained model that is trained by the first learning unit 320 and re-trained by the second learning unit 330 and that is used when the analysis unit 340 performs analysis has been described as being arranged within the processing device 100. However, the trained model may also be arranged as generative AI (artificial intelligence) within a server device connected via a network, for example. The following describes the fourth embodiment, focusing on differences from the above embodiments.

[0245] <System Including Processing Device> An example configuration of a system including a processing device according to the fourth embodiment will be described. Fig. 15 is a second diagram showing an example configuration of a system including a processing device. As shown in Fig. 15, a system 20 includes a processing device 1500 according to the fourth embodiment, a server device 110, and a server device 1510. In the system 20, the processing device 1500, the server device 110, and the server device 1510 are connected via a network 130.

[0246] The server device 110 has already been described in the first embodiment, and therefore will not be described here.

[0247] The server device 1510 functions as an analysis service providing unit 1511 and has a generation AI 1512. The server device 1510 performs learning or relearning on the generation AI 1512 based on instructions from the processing device 1500. The server device 1510 performs analysis based on instructions from the processing device 1500 and transmits the analysis results to the server device 1510.

[0248] The processing device 1500 is a device operated by the user 101. In the learning phase, the processing device 1500 generates first learning data or second learning data. The processing device 1500 transmits the generated first learning data or second learning data to the server device 1510, thereby instructing the server device 1510 to learn or re-learn the generated AI 1512.

[0249] In the analysis phase, the processing device 1500 generates analysis data and sends the generated analysis data to the server device 1510, thereby instructing the server device 1510 to perform analysis by the generation AI 1512.

[0250] The processing device 1500 obtains the analysis results analyzed by the generation AI 1512 from the server device 1510 and outputs them to the user 101.

[0251] <Functional Configuration of Processing Device> The functional configuration of the processing device 1500 will be described. FIG. 16 is a third diagram illustrating an example of the functional configuration of the processing device. As in the first embodiment, a processing program is installed in the processing device 1500, and by executing the processing program, the processing device 1500 functions as a data transmission / reception unit 1610, a first learning unit 320, a second learning unit 330, and an analysis unit 1620. The processing device 1500 functions as a first distributed representation generation unit 360 and a second distributed representation generation unit 370. Of the functional units illustrated in FIG. 16, the first learning unit 320, the second learning unit 330, the first distributed representation generation unit 360, and the second distributed representation generation unit 370 have already been described, and therefore their description will be omitted here.

[0252] The data transmitting / receiving unit 1610 receives document data provided by the server device 110. The data receiving unit 310 notifies the first learning unit 320 or the second learning unit 330 of the received document data.

[0253] The data transmission / reception unit 1610 transmits the first learning data generated by the first learning unit 320 to the server device 1510 and instructs the server device 1510 to use the first learning data to train the generated AI 1512.

[0254] The data transmission / reception unit 1610 transmits the second learning data generated by the second learning unit 330 to the server device 1510 and instructs the server device 1510 to re-learn the generated AI 1512 using the second learning data.

[0255] The data transmission / reception unit 1610 transmits the analysis data generated by the analysis unit 1620 to the server device 1510, and instructs the server device 1510 on the analysis content to be performed using the analysis data.

[0256] Furthermore, the data transmission / reception unit 1610 receives the analysis results transmitted from the server device 1510 in response to an instruction to the server device 1510 regarding the analysis content, and notifies the analysis unit 1620 of the results.

[0257] The analysis unit 1620 generates analysis data based on the text information input by the user 101. The analysis unit 1620 generates an analysis instruction describing the analysis content, and notifies the data transmission / reception unit 1610 of the instruction together with the analysis data.

[0258] The analysis unit 1620 outputs the analysis result notified by the data transmission / reception unit 1610 in response to the notification of the analysis instruction to the user 101 .

[0259] <Details of Processing by the Analysis Unit of the Processing Device> A description will be given of the flow of processing by the analysis unit 1620 of the processing device 1500. Fig. 17 is a second flowchart showing the flow of processing by the analysis unit.

[0260] In step S 1701 , the analysis unit 1620 receives input of text information generated by the user 101 .

[0261] In step S1702, if the text information includes a substance name, the analysis unit 1620 identifies structural information of the substance having the substance name and obtains a first distributed representation based on the identified structural information. If the text information includes a token common to the substance name, the analysis unit 1620 obtains a first distributed representation based on the token common to the substance name.

[0262] In step S1703, the analysis unit 1620 obtains a second distributed representation based on the part of the text information other than the substance name.

[0263] In step S1704, the analysis unit 1620 generates an analysis instruction describing the analysis content, and transmits the analysis instruction together with the analysis data to the server device 1510. Note that the analysis unit 1620 sets a set of distributed representations including the first distributed representation obtained in step S1701 and the second distributed representation obtained in step S1703 as the analysis data.

[0264] In step S 1705 , the analysis unit 1620 acquires the analysis result from the server device 1510 and outputs it to the user 101 .

[0265] In step S1706, the analysis unit 1620 determines whether or not to continue the analysis, and if it is determined that the analysis should be continued (YES in step S1706), the process returns to step S1701.

[0266] On the other hand, if it is determined in step S1706 that the analysis is to be ended (NO in step S1706), the processing by the analysis unit 1620 ends.

[0267] <Summary> As is clear from the above description, the processing method according to the fourth embodiment performs learning or relearning on a generation AI located in a server device connected via a network, using a set of distributed representations including a first distributed representation based on structural information of a substance and a second distributed representation based on text information describing the substance. Analysis data including the first distributed representation based on structural information of the substance or the second distributed representation based on text information describing the substance, together with an analysis instruction describing the analysis content, is transmitted to the server device, thereby obtaining analysis results.

[0268] In this way, according to the processing method of the fourth embodiment, similar learning or relearning can be performed on the generated AI located in the server device connected via a network. In other words, according to the fourth embodiment, it is possible to enjoy the same effects as in the first embodiment.

[0269] Other Embodiments In the above embodiments, the first training unit 320 trains a model using a set of embedded representations corresponding to the word order of the original sentence as the first training data. However, the first training data used by the first training unit 320 for model training is not limited to this. For example, a set of embedded representations consisting of pairs of first embedded representations based on structural information of a substance and second embedded representations based on a description of the use or function of the substance may be used as the first training data.

[0270] In the above embodiments, the analysis unit 340 has been described as acquiring a token indicating the name of a substance as an example of an analysis result, but the analysis unit 340 may be configured to acquire tokens indicating the names of related substances related to the substance. Related substances related to the substance include, for example, an alternative material that substitutes for a material containing the substance, a composite material that contains the substance, and a solvent that turns a material containing the substance into a solution.

[0271] In addition, when obtaining a token indicating the substance name of the related substance as the analysis result, the analysis unit 340 may input analysis data including a first distributed representation based on the substance name of the substance to the model execution unit 300. In other words, the analysis unit 340 may analyze related substances related to the substance from the substance.

[0272] Alternatively, when obtaining tokens indicating the substance names of related substances as analysis results, the analysis unit 340 may input analysis data including second distributed representations based on text information including the use, function, etc. of the substance to the model execution unit 300. In other words, the analysis unit 340 may analyze related substances related to the substance from the use, function, etc. of the substance.

[0273] Alternatively, when obtaining a token indicating the substance name of a related substance as an analysis result, the analysis unit 340 may input a set of distributed representations including both: a first distributed representation based on the substance name of the substance; and a second distributed representation based on text information including the use, function, etc. of the substance, as analysis data to the model execution unit 300. In other words, the analysis unit 340 may analyze related substances related to a substance from the substance, its use, function, etc.

[0274] In the above embodiments, the masking process is performed before the first or second distributed representation is acquired, but the processing order is not limited to this. For example, after the first distributed representation is acquired, part of the first distributed representation may be replaced with a first distributed representation based on a token common to the substance name. Alternatively, after the second distributed representation is acquired, part of the second distributed representation may be left blank.

[0275] In the above embodiments, a method using mask processing (MLM: Mask Language Model) has been described as a method for performing supervised learning. However, the method for performing supervised learning is not limited to this. For example, supervised learning may be performed using Next Sentence Prediction (NSP). Alternatively, supervised learning may be performed using Causal Language Model (CLM). Alternatively, supervised learning may be performed by preparing other correct answer data. The other correct answer data referred to here includes, for example, data annotated with text information indicating usage, function, etc.

[0276] In each of the above embodiments, a part of the text information other than the substance name (a part describing the substance's use, function, etc.) is used as the correct answer data. However, the correct answer data is not limited to the part of the text information other than the substance name, and may be a second distributed representation based on the part of the text information other than the substance name.

[0277] Similarly, in each of the above embodiments, the substance name included in the text information is used as the correct answer data. However, the correct answer data is not limited to the substance name included in the text information, and may be structural information of a substance having the substance name, or SMILES information obtained by converting the structural information into a SMILES representation. Alternatively, the correct answer data may be a graph based on the SMILES information. Alternatively, the correct answer data may be a first distributed representation based on the SMILES information or a graph.

[0278] In the above embodiments, when masking substance names, the masking is performed using tokens that are common to the substance names, but the method of masking substance names is not limited to this. However, using tokens that are common to the substance names has the advantage of reducing the time required for learning.

[0279] In the above embodiments, examples have been given of cases in which the analysis results are output as the substance name, substance use, function, etc. However, the analysis results to be output are not limited to these, and for example, an outline of the input document data may be output. In this case, for example, by adding a configuration for calculating the similarity of each document data based on the outline of each document data, the processing device 100 can function as a search device that searches for similar document data.

[0280] In each of the above embodiments, the structural information of a substance having a substance name included in text information is identified by referring to a database. However, the method of identifying the structural information is not limited to this. For example, in the case of a document that includes structural information indicating the molecular structure of a substance, such as a paper, magazine, or patent document in the field of chemistry, the structural information included in the document may be used to identify the structural information of the substance. In this case, the processing device 100 analyzes a document that includes structural information of a substance and text information describing the substance using a single trained model.

[0281] In the above embodiments, the structural information of a substance having a substance name included in text information is identified from the substance name. However, if the substance having the substance name is a polymer, the processing device 100 identifies the structural information of the monomer instead of the polymer. In this way, the processing device 100 is assumed to have a database suitable for executing the processing method described in the above embodiments.

[0282] In the above embodiments, the processing is performed by one processing device 100 (or processing device 1500), but the processing may be performed by a plurality of processing devices.

[0283] Although the embodiments have been described above, it will be understood that various changes in form and details can be made without departing from the spirit and scope of the claims.

[0284] This application claims priority based on Japanese Patent Application No. 2024-029780, filed on February 29, 2024, the entire contents of which are incorporated herein by reference.

[0285] 100: Processing device 300: Model execution unit 310: Data receiving unit 320: First learning unit 330: Second learning unit 340: Analysis unit 360: First distributed representation generation unit 370: Second distributed representation generation unit 400: Database 1110: Second learning unit 1120: First distributed representation generation unit 1130: Second distributed representation generation unit 1140: Model execution unit 1500: Processing device 1510: Server device 1512: Generation AI 1610: Data transmission / reception unit 1620: Analysis unit

Claims

1. A processing method comprising: a first output step of outputting a first distributed representation based on structural information of a substance having a substance name included in a document; a second output step of outputting a second distributed representation based on text information describing a substance having a substance name included in the document; and a learning step of learning a model using the first distributed representation and the second distributed representation.

2. The processing method according to claim 1, wherein in the learning step, learning the model includes relearning a model that has already been learned.

3. The processing method described in claim 1, wherein the text information is a sentence including a substance name, and in the learning process, learning the model includes learning the model using a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output process and the second output process.

4. The processing method described in claim 3, wherein, when the text information is a sentence including a substance name, a set of embedded representations is generated in which a first embedded representation is placed at a position corresponding to the position of the substance name and a second embedded representation is placed at a position corresponding to the position of the sentence other than the substance name.

5. The processing method of claim 3, wherein the text information is a sentence including a substance name, and in the learning step, relearning the model includes: masking a portion of the text information other than the substance name; and performing supervised learning on the text information after the masking step, using as input data a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output step and the second output step, and using as correct data the portion of the text information other than the substance name.

6. The processing method of claim 2, wherein the text information is a sentence including a substance name, and in the learning step, relearning the trained model includes: masking the substance name of the text information using a common token; and performing supervised learning on the text information after the masking step using a set of distributed representations including a first distributed representation and a second distributed representation output by executing the first output step and the second output step as input data, and using the substance name of the text information as correct answer data.

7. The processing method of claim 1, wherein the text information is a sentence including a substance name, and in the learning process, learning the model includes relearning either a trained first model that outputs a first distributed representation, or a trained second model that outputs a second distributed representation, or the trained model to which the outputs of the trained first model and the trained second model are input.

8. The processing method described in claim 7, wherein in the learning process, the trained first model that outputs the first distributed representation is trained so that first distributed representations based on structural information of substances with different substance names that are respectively included in sentences with similar contexts are placed close to each other.

9. The processing method according to claim 1, further comprising a third output step of inputting a first distributed representation based on structural information of a substance into a trained model trained in the training step, thereby outputting a token indicating the use or function of the substance.

10. The processing method of claim 1, further comprising a third output step of inputting a second distributed representation based on text information describing the use or function of a substance into a trained model trained in the training step, thereby outputting a token indicating the substance name of the substance.

11. The processing method of claim 1, further comprising a third output step of inputting either or both of a first distributed representation based on structural information of a substance and a second distributed representation based on text information describing the use or function of the substance into a trained model trained in the training step, thereby outputting tokens indicating the names of related substances related to the substance.

12. A processing method according to any one of claims 1 to 11, comprising an identification step of identifying structural information of a substance having a substance name or the substance name of a substance having structural information by referring to a database in which substance names are associated with structural information of substances having said substance names.

13. The processing method described in claim 12, wherein if a substance having a substance name contained in the document is a substance that becomes a single atom when the structural information of the substance is converted into a string according to predetermined rules, or if a substance having a substance name contained in the document is a substance with an anomalous valence, the second output step outputs the second distributed representation based on the substance name contained in the document.

14. The method of any one of claims 1 to 11, wherein the documents include papers, journals, and patent documents.

15. A processing device (100, 1500) having a control unit (200), wherein the control unit (200) outputs a first distributed representation based on structural information of a substance having a substance name included in a document, outputs a second distributed representation based on text information describing a substance having a substance name included in the document, and trains a model using the first distributed representation and the second distributed representation.

16. The processing device (100, 1500) according to claim 15, wherein the control unit (200) outputs a token indicating the use or function of the substance by inputting a first distributed representation based on structural information of the substance into a trained model.

17. The processing device (100, 1500) of claim 15, wherein the control unit (200) outputs a token indicating the substance name of the substance by inputting a second distributed representation based on text information describing the use or function of the substance into a trained model.

18. A processing program for causing a control unit (200) to execute the following steps: a first output step of outputting a first distributed representation based on structural information of a substance having a substance name included in a document; a second output step of outputting a second distributed representation based on text information describing a substance having a substance name included in the document; and a learning step of learning a model using the first distributed representation and the second distributed representation.

19. A processing program as described in claim 18, for causing a control unit (200) to execute a third output step of inputting a first distributed representation based on structural information of a substance into a trained model, thereby outputting a token indicating the use or function of the substance.

20. A processing program as described in claim 18, for causing the control unit (200) to execute a third output step of inputting a second distributed representation based on text information describing the use or function of a substance into a trained model, thereby outputting a token indicating the substance name of the substance.

Images