Information processing system, information processing apparatus, method, and program

Abstract

To provide an information processing system, an information processing device, a method, and a program for generating a machine learning model with improved accuracy of answers.SOLUTION: An information processing system that performs post-learning of a machine learning model includes a database information acquisition unit 304 that acquires information stored in a database storage unit 308, a learning file generation unit 305 that generates a file for post-learning on the basis of the information acquired by the database information acquisition unit 304, and a post-learning processing unit 309 that performs processing of post-learning of the machine learning model on the basis of the file.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to an information processing system, an information processing device, a method, and a program for generating a fine-tuned machine learning model in generative AI. [Background technology]

[0002] With the advancement of information processing technology, development of AI that can generate text is progressing.

[0003] For example, Japanese Patent Application Laid-Open No. 2021-64183 (Patent Document 1) discloses a technology for outputting text to be used for medication instruction based on input prescription data. Patent Document 1 provides a medication instruction support device that can efficiently and easily provide appropriate medication instruction to patients.

[0004] Recently, AI technologies have been developed that can generate text and provide answers to questions entered in natural language by users. However, because these technologies generate machine learning models based on data collected widely from the internet, there is a risk that they may learn incorrect information and provide answers. To address this issue, there are techniques called fine-tuning, which allows machine learning models to learn information from specific fields, thereby generating machine learning models specialized for specific fields. However, this requires developers to provide information sources and perform fine-tuning one by one, which is a significant burden.

[0005] In addition, the content of the answer may vary depending on the quality of the user's question, and the answer may not be appropriate. In other words, the answer may not be based on the user's attributes (for example, whether the user is an expert or an ordinary person), and the user may not necessarily receive an appropriate answer.

[0006] In this way, depending on the constructed machine learning model and the question, an incorrect answer may be given, which may cause a disadvantage to the user. In particular, as in Patent Document 1, in the field of medication instructions, this could lead to serious consequences.

[0007] Therefore, there is a need for further technology to easily generate machine learning models that can appropriately answer questions. Summary of the Invention [Problem to be solved by the invention]

[0008] The present invention has been made in consideration of the problems in the above-mentioned conventional technology, and aims to provide an information processing system, an information processing device, a method, and a program for generating a machine learning model with improved answer accuracy. [Means for solving the problem]

[0009] That is, according to the present invention, An information processing system for post-learning a machine learning model, An acquisition means for acquiring information stored in the database; a generation means for generating a file for post-learning based on the information acquired by the acquisition means; a processing means for performing post-learning of a machine learning model based on the file; An information processing system is provided, including: [Effects of the Invention]

[0010] According to the present invention, it is possible to provide an information processing system, an information processing device, a method, and a program for generating a machine learning model with improved answer accuracy. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram showing a schematic hardware configuration of an entire information processing system according to an embodiment of the present invention. [Figure 2]FIG. 2 is a diagram showing the hardware configuration included in the information processing apparatus of the present embodiment. [Figure 3] FIG. 2 is a block diagram of software included in the information processing system of the present embodiment. [Figure 4] FIG. 4 is a diagram showing an example of information stored in the embodiment. [Figure 5] 4 is a flowchart showing a process executed by the information processing system of the present embodiment. [Figure 6] FIG. 10 is a diagram showing an example of an answer screen displayed in the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0012] The present invention will be described below with reference to embodiments, but the present invention is not limited to the embodiments described below. In the drawings referred to below, the same reference numerals will be used for common elements, and their description will be omitted as appropriate.

[0013] FIG. 1 is a diagram showing a schematic hardware configuration of an information processing system 100 according to this embodiment. As shown in FIG. 1, the information processing system 100 according to this embodiment is configured to include a plurality of information processing devices 110. In the example shown in FIG. 1, the information processing system 100 includes various information processing devices 110, such as a development environment terminal 110a, a development environment server 110b, a database server 110c, a text generation service providing server 110d, and a user terminal 110e. The plurality of information processing devices 110 can communicate with each other via a network 120 such as the Internet or a LAN. The information processing devices 110 may be connected to the network 120 either wired or wirelessly.

[0014] The development environment terminal 110a is a terminal used in the development environment of this embodiment. Here, the development environment of this embodiment refers to an environment in which processing for building a machine learning model specialized for a specific field is performed. The development environment terminal 110a can be in the form of a personal computer, for example. The development environment terminal 110a of this embodiment is used by a developer to communicate with a development environment server 110b via a network 120 and perform various processes.

[0015] The development environment server 110b is a server used in the development environment and can take the form of, for example, a server computer. The development environment server 110b of this embodiment can communicate with other server devices via the network 120 and perform processing to build a machine learning model.

[0016] In this embodiment, the database server 110c is a server that provides various databases and may take the form of, for example, a server computer. The database server 110c of this embodiment can store databases specialized in specific fields for post-learning (also referred to as fine tuning) of a general-purpose machine learning model and provide them to the development environment server 110b. The database server 110c can store, for example, a database related to medicines, law, literature, art, or economics. However, this is merely an example and does not limit the embodiment, and other databases may be stored. Note that the information processing system 100 of this embodiment may include multiple database servers 110c.

[0017] The text generation service providing server 110d is a server that provides a service for generating text in this embodiment, and may take the form of, for example, a server computer. The text generation service providing server 110d of this embodiment stores a machine learning model and can provide a user with text generated in an interactive format. The text generation service providing server 110d of this embodiment can be configured as a so-called text generation AI, and the AI ​​can provide an answer to a question entered by a user based on a machine learning model. Commonly known text generation services include ChatGPT (registered trademark) provided by OpenAI (registered trademark) and Gemini (registered trademark) provided by Google (registered trademark). The text generation service providing server 110d of this embodiment may use these services or other services.

[0018] The user terminal 110e is a terminal used by a user who uses the text generation service of this embodiment, and can take the form of, for example, a smartphone terminal, a tablet terminal, or a personal computer terminal. The user can access the text generation service providing server 110d using the user terminal 110e. For example, the user can input a question about a specific field and obtain an answer from the text generation service providing server 110d of this embodiment.

[0019] Note that the configuration of the information processing device 110 included in the information processing system 100 shown in Fig. 1 is an example and does not limit the embodiment. Therefore, the information processing system 100 of this embodiment may include information processing devices 110 other than those shown in Fig. 1, or may be configured without including some of the information processing devices 110 shown in Fig. 1. Furthermore, the development environment server 110b, database server 110c, and text generation service providing server 110d in this embodiment do not have to be in the form of server computers, and may be in the form of, for example, a cloud computing service.

[0020] Next, the hardware configuration of each information processing device 110 of this embodiment will be described. Fig. 2 is a diagram showing the hardware configuration included in the information processing device 110 of this embodiment. The information processing device 110 is configured to include a CPU 210, a RAM 220, a ROM 230, a storage device 240, a communication I / F 250, a display 260, and an input device 270, and each piece of hardware is connected via a bus. Note that the development environment server 110b, the database server 110c, and the text generation service providing server 110d of this embodiment may be configured not to include the display 260 or the input device 270.

[0021] The CPU 210 is a device that executes programs that control the operation of the information processing device 110 and performs predetermined processing. The RAM 220 is a volatile storage device that provides an execution space for the programs executed by the CPU 210, and is used for storing and extracting programs and data. The ROM 230 is a non-volatile storage device that stores programs executed by the CPU 210, firmware, etc.

[0022] The storage device 240 is a readable / writable nonvolatile storage device that stores an OS, various software, setting information, various data, etc. that operate the information processing device 110. Examples of the storage device 240 include a hard disk drive (HDD) and a solid state drive (SSD).

[0023] The communication I / F 250 connects the information processing device 110 to the network 120, enabling communication with other devices via the network 120. The communication via the network 120 may be either wired communication or wireless communication, and various types of data can be sent and received using a predetermined communication protocol such as TCP / IP.

[0024] The display 260 is a device that displays various data, the status of the information processing device 110, etc. to the user, and examples thereof include an LCD (Liquid Crystal Display). The input device 270 is a device that allows the user to operate the information processing device 110, and examples thereof include a keyboard and a mouse. The display 260 and the input device 270 may be separate devices, or may be a device that has both functions, such as a touch panel display.

[0025] The hardware configuration included in the information processing device 110 of this embodiment has been described above. Next, the functional means executed by each piece of hardware in this embodiment will be described with reference to Fig. 3. Fig. 3 is a software block diagram included in the information processing system 100 of this embodiment.

[0026] As shown in Fig. 3, the development environment terminal 110a and user terminal 110e of this embodiment include a communication unit 301, a display unit 302, and an operation unit 303. The development environment server 110b of this embodiment includes a communication unit 301, a database information acquisition unit 304, a training file generation unit 305, a variation creation unit 306, and a log recording unit 307. The database server 110c of this embodiment includes a database storage unit 308. The text generation service providing server 110d of this embodiment includes a post-learning processing unit 309, a general-purpose model storage unit 310, a dedicated model storage unit 311, and a text generation unit 312. Each functional unit will be described in detail below.

[0027] The communication unit 301 is a means for controlling the operation of the communication I / F 250 and performing communication between the information processing devices 110. The communication unit 301 constitutes the communication means in this embodiment. The communication unit 301 in this embodiment can, for example, receive various data from other information processing devices 110 and transmit various data to other information processing devices 110.

[0028] The display unit 302 is a means for controlling the operation of the display 260 and displaying various types of information. The display unit 302 constitutes the display means in this embodiment. The display unit 302 in this embodiment displays various types of information on the display 260, allowing developers and users to recognize the information. The display unit 302 of the development environment terminal 110a can display, for example, information related to the process of generating a machine learning model. Furthermore, the display unit 302 of the user terminal 110e can display, for example, answers to questions asked by a user in a text generation service.

[0029] The operation unit 303 is a means for accepting operations via the input device 270. The operation unit 303 constitutes the operation means in this embodiment. The operation unit 303 in this embodiment can accept operations performed on the screen displayed on the display 260 by the display unit 302. The operation unit 303 of the development environment terminal 110a can accept, for example, an operation for performing processing to generate a machine learning model. Furthermore, the operation unit 303 of the user terminal 110e can accept, for example, an operation for inputting a question when using a text generation service.

[0030] The database information acquisition unit 304 is a means for acquiring database information from the database server 110c. The database information acquisition unit 304 constitutes an acquisition means in this embodiment. The database information acquisition unit 304 in this embodiment can acquire information on a database that serves as a teacher for post-learning (fine-tuning) a general-purpose machine learning model.

[0031] The learning file generation unit 305 is a means for generating a learning file for post-learning. The learning file generation unit 305 constitutes the generation means in this embodiment. The learning file generation unit 305 in this embodiment can create a learning file for post-learning based on the information acquired by the database information acquisition unit 304 and the variations of the question created by the variation creation unit 306. The created learning file can be sent to the text generation service providing server 110d.

[0032] The variation creation unit 306 is configured as part of the training file generation unit 305 and is means for creating variations of questions when post-training the machine learning model. The variation creation unit 306 constitutes the creation means in this embodiment. The variation creation unit 306 in this embodiment uses a general-purpose large-scale language model to create variations of questions assuming questioners with different levels of knowledge in a specific field. In the post-training in this embodiment, by taking into account multiple variations of questions, it is possible to generate a machine learning model that can answer the same question even if the question is asked in different ways. This makes it possible to provide the same answer tailored to the level of the questioner, whether the questioner (user) is an expert or an ordinary person.

[0033] The log recording unit 307 is a means for controlling the operation of the storage device 240 of the development environment server 110b and recording a log of the process of post-learning a machine learning model in the storage device 240. The log recording unit 307 constitutes the recording means in this embodiment. Here, the log information recorded in the log recording unit 307 in this embodiment will be described with reference to FIG. 4(a).

[0034] Fig. 4 is a diagram showing an example of information stored in this embodiment, and Fig. 4(a) shows an example of log information recorded in the log recording unit 307. As shown in Fig. 4(a), the log recording unit 307 records a table in which an ID (Log ID) for identifying a log, the date and time (Date) when the post-learning process was performed, the result (Result) of the post-learning process, and the model name (Model name) of the machine learning model generated by the post-learning process are associated with each other. Note that if the post-learning process is unsuccessful for some reason, "failure" is recorded in the "Result" field, and the "Model name" field is blank.

[0035] Returning to FIG. 3 for the explanation, the database storage unit 308 is a means for controlling the operation of the storage device 240 of the database server 110c and storing the database. The database storage unit 308 constitutes a first storage means in this embodiment. The database storage unit 308 in this embodiment can store a database specialized in a specific field that serves as a teacher for post-learning. The database stored in the database storage unit 308 is open to the public in an accessible state. The database storage unit 308 can store, for example, a database related to drugs.

[0036] The post-learning processing unit 309 is a means for performing a process of post-learning a machine learning model based on the training file. The post-learning processing unit 309 constitutes the processing means in this embodiment. The post-learning processing unit 309 in this embodiment can perform post-learning (fine-tuning) by applying the training file to a general-purpose machine learning model. This makes it possible to generate a machine learning model specialized for a specific field (hereinafter referred to as a dedicated model).

[0037] The generic model storage unit 310 controls the operation of the storage device 240 of the text generation service providing server 110d and stores generic machine learning models (hereinafter referred to as generic models) provided by the text generation service. The generic model storage unit 310 constitutes a second storage unit in this embodiment. Generic models can be used to generate answers to questions that do not require highly specialized answers. Generic models may also be used by the variation creation unit 306 to create variations of multiple question sentences.

[0038] The dedicated model storage unit 311 is a means for controlling the operation of the storage device 240 of the text generation service providing server 110d and storing the dedicated models generated by the post-learning processing unit 309. The dedicated model storage unit 311 constitutes a third storage means in this embodiment. Here, the dedicated models stored in the dedicated model storage unit 311 in this embodiment will be described with reference to FIG. 4(b).

[0039] As shown in FIG. 4(b), the dedicated model storage unit 311 stores an ID (Model ID) for identifying a dedicated model, data of the dedicated model (Model data), and the date and time when the dedicated model was generated, in association with each other. The dedicated model storage unit 311 of this embodiment can store dedicated models generated by a post-learning process, and can continue to store previously generated generative models even when a new post-learning process is performed. For example, as shown in FIG. 4(b), the latest dedicated model M004 was generated on August 2, 2024, and stored in the dedicated model storage unit 311. However, dedicated models M001, M002, and M003 generated before that are also stored in the dedicated model storage unit 311. In this way, by storing dedicated models generated in the past, even if incorrect information is learned, the text generation service can be provided using past dedicated models, thereby enabling highly accurate answers. Note that the dedicated model used in the text generation service may be selected by the developer or user.

[0040] Returning to FIG. 3 for the explanation, the text generation unit 312 is a means for generating text that answers a question. The text generation unit 312 constitutes the text generation means in this embodiment. The text generation unit 312 in this embodiment can input a question received from a user into a machine learning model and generate text that answers the question. The generated text is transmitted to the user terminal 110e and can be viewed by the user.

[0041] 3 correspond to functional means realized by causing each piece of hardware to function as a result of the CPU 210 executing the program of this embodiment. Furthermore, the functional means shown in each embodiment may be realized entirely by software, or some or all of them may be implemented as hardware that provides equivalent functions.

[0042] Furthermore, the above-mentioned functional means do not necessarily all need to be included in the configuration shown in Fig. 3. For example, the development environment terminal 110a may be configured to include functional means included in the development environment server 110b. Also, for example, the development environment server 110b may be configured to include the post-learning processing unit 309. In still other embodiments, the functional means may be realized by cooperation between multiple information processing devices 110.

[0043] Next, the processes executed by the various functional means described above will be described with reference to Fig. 5. Note that, in the following embodiment, an example will be described in which information about a drug is acquired from a database and a dedicated model is generated through post-training, but this is not intended to limit the embodiment in any way.

[0044] 5 is a flowchart showing the processing executed by the information processing system 100 of this embodiment. The information processing system 100 starts the processing from step S1000.

[0045] 5 can be started at any timing. For example, the information processing system 100 may be configured to start the processing periodically, such as at a predetermined time, a predetermined date, or a predetermined day of the week. The information processing system 100 may also start the processing when a developer performs an operation to start the processing. Furthermore, the information processing system 100 may detect an update to the database by, for example, monitoring the database in the database storage unit 308 or detecting differences in data within the database, and start the processing when the database is updated.

[0046] In step S1001, the development environment server 110b accesses the database server 110c. In this embodiment, the development environment server 110b can access the database server 110c through the operation of the development environment terminal 110a by the developer. The processing of step S1001 allows the development environment server 110b to establish a connection with the database server 110c.

[0047] In step S1002, the database information acquisition unit 304 acquires data related to keywords from the database storage unit 308 of the database server 110c. Here, for example, consider a case where post-learning related to drugs is performed and the keyword to be searched is "cancer." When the development environment terminal 110a accepts input of "cancer" as a keyword, the database information acquisition unit 304 searches the database in the database storage unit 308 for data related to "cancer." Then, the database information acquisition unit 304 acquires the data extracted by the search. Because the database contains data related to drugs, it is possible to extract data on anticancer drugs, for example.

[0048] Then, in step S1003, the variation creation unit 306 creates variations in the expression of the answer. The creation of variations can be performed using, for example, a generic model. The variation creation unit 306 of this embodiment can create variations of questioners with different levels of knowledge about medications using the generic model. For example, the variation creation unit 306 can create multiple question variations, such as questions from questioners with specialized knowledge about medications, such as pharmacists and medical professionals, questions from ordinary people, such as patients who take medications, and questions from children who do not have sufficient knowledge about medications. In other words, the variation creation unit 306 of this embodiment can create variations in the linguistic expression of the question.

[0049] Next, in step S1004, the training file generation unit 305 associates the keywords acquired in step S1002 with the variations created in step S1003 and stores them in a JSON-formatted file. The JSON format is a common data description format used when exchanging data, and is characterized by its small data volume and high readability. The file generated in step S1004 is not limited to JSON format and can be in any format. In step S1004, the keywords can be stored in a number of files corresponding to the number of variations created in step S1003.

[0050] In the next step S1005, the process branches depending on whether or not there are additional keywords. Additional keywords are keywords to be searched in addition to the keywords searched in step S1002. For example, if data related to the keyword "cancer" is acquired in step S1002, the process branches in step S1005 depending on whether or not keywords other than "cancer" are set as search targets. For example, if "adverse drug reaction" is set as an additional keyword related to a drug other than "cancer" (YES), the process proceeds to step S1006.

[0051] In step S1006, similar to step S1002, the database information acquisition unit 304 searches the database in the database storage unit 308 for data related to "adverse drug reaction." Then, the database information acquisition unit 304 acquires the data extracted by the search. Because the database contains data related to drugs, for example, data related to adverse drug reactions of anticancer drugs can be extracted in addition to "cancer," which is the keyword to be searched in step S1002.

[0052] Thereafter, in step S1007, the training file generation unit 305 stores the acquired data in a JSON format file, similar to step S1004. In step S1007, for example, the acquired data may be temporarily pooled as a preliminary process before writing out a JSON format file, which will be described later.

[0053] Thereafter, the process returns to step S1005 and the above process is repeated for all additional keywords.

[0054] In step S1005, if there are no additional keywords related to drugs other than "cancer" (NO), that is, if no additional keywords have been set or if steps S1006 to S1007 have been processed for all additional keywords, proceed to step S1008.

[0055] In step S1008, the training file generation unit 305 generates a file for post-learning. The file for post-learning can be created based on a file that associates the acquired keywords with the variations created in step S1003 and stores them in JSON format. In this embodiment, the file for post-learning can be configured as a file in JSONL format, for example. The JSONL format is a JSON format in which each unit is written on one line, and has the advantage of reducing sparsity. Note that the JSONL format file is just an example and does not particularly limit the embodiment, and a file in another format may also be used. The generated file for post-learning is transmitted to the text generation service providing server 110d via the communication unit 301.

[0056] Then, in step S1009, the post-learning processing unit 309 of the text generation service providing server 110d performs post-learning on the general-purpose model. The post-learning processing unit 309 of this embodiment can post-learn the general-purpose model based on the post-learning file generated and transmitted in step S1008, and construct a dedicated model. The results of learning by the post-learning processing unit 309 can be transmitted to the development environment server 110b.

[0057] Next, in step S1010, the log recording unit 307 records the execution result of the post-learning as a log. This allows the result of the post-learning to be recorded in the table shown in Fig. 4(a), and the developer can understand the result of the post-learning process.

[0058] In step S1011, the post-learning is completed and the output dedicated model is stored in the dedicated model storage unit 311. When storing the dedicated model in the dedicated model storage unit 311 in step S1011, the previously stored dedicated model is not deleted, but is maintained together with the newly output dedicated model in the dedicated model storage unit 311. As a result, even if the new dedicated model is not appropriate, a highly accurate answer can be obtained by using the previous dedicated model.

[0059] Thereafter, in step S1012, the information processing system 100 ends the process.

[0060] 5, the information processing system 100 of this embodiment can create a dedicated machine learning model that has learned specialized information. In particular, in this embodiment, by acquiring information from an external database, post-learning can be performed automatically, eliminating the need to manually input data related to learning.

[0061] The dedicated model generated by the process of Fig. 5 may be verified by the developer. If any defect is discovered during the verification process by the developer, the newly generated dedicated model may be discarded, and the previous dedicated model may continue to be usable.

[0062] The generated dedicated model can be used by a user through a text generation service provided via the text generation service providing server 110d. When the text generation service providing server 110d receives a question input by a user via the user terminal 110e, the text generator 312 inputs the question text into the dedicated model and outputs an answer to the question text. The output answer is transmitted to the user terminal 110e and displayed on the display 260 of the user terminal 110e, allowing the user to view the answer.

[0063] Here, the answer screen in this embodiment will be described with reference to FIG. 6. FIG. 6 is a diagram showing an example of an answer screen displayed in this embodiment. FIG. 6(a) is an example of a screen displaying an answer to a question posed by a user with specialized knowledge about drugs when an appropriate dedicated model has been generated by post-mortem learning. FIG. 6(b) is an example of a screen displaying an answer to a question posed by a user with specialized knowledge about drugs when an inappropriate dedicated model has been generated by post-mortem learning. FIG. 6(c) is an example of a screen displaying an answer to a question posed by a user whose knowledge about drugs is not at a specialized level.

[0064] First, FIG. 6(a) will be described. FIG. 6(a) shows an example in which a user with specialized knowledge about drugs asks, "I want to know the pharmacological action of cisplatin." The text generation unit 312 inputs the input question into a dedicated model, performs processing such as morphological analysis, and generates an answer to the question. Here, since the question contains relatively specialized terms such as "pharmacological action," the text generation unit 312 can generate a highly specialized answer tailored to the user's level of knowledge, assuming that the user has specialized knowledge. Therefore, the text generation unit 312 can output and display a highly specialized answer to the question "pharmacological action of cisplatin," as shown in FIG. 6(a).

[0065] The answer display screen in this embodiment may be configured to display, for example, a "Answer with an older version" button. If the displayed answer is clearly incorrect, the user can press the "Answer with an older version" button, which allows the text generator 312 to generate an answer for the same question using a past dedicated model, thereby improving the accuracy of the answer. For example, if the dedicated model generated by the latest post-learning process has not been properly trained for some reason, an answer such as that shown in FIG. 6(b) may be displayed.

[0066] FIG. 6(b) shows an example of an answer provided using a dedicated model generated by inappropriate post-training. In the example shown in FIG. 6(b), the question "I want to know the pharmacological effects of cisplatin" is entered, as in FIG. 6(a). However, the dedicated model used to output the answer was not properly post-trained, resulting in incorrect information about cisplatin. In such a case, the user can press the "Answer with an older version" button, which causes the text generator 312 to generate an answer to the same question using a previously generated dedicated model, rather than the dedicated model used to output the answer shown in FIG. 6(b). Therefore, pressing the "Answer with an older version" button on the screen shown in FIG. 6(b) can display a screen in which an appropriate answer has been generated, such as that shown in FIG. 6(a).

[0067] Furthermore, in this embodiment, since multiple variations of question sentences are learned during post-training, answers to the same meaning can be generated using different expressions depending on the question sentence. For example, as shown in FIG. 6(c), consider a case where a user without specialized knowledge about drugs enters a question such as "What is cisplatin?" In this case, since the question does not use highly specialized words such as "pharmacological action" as shown in FIG. 6(a), the text generator 312 determines that the user does not have specialized knowledge and can generate an answer that does not use technical terms, such as the example shown in FIG. 6(c), in accordance with the user's level of knowledge.

[0068] In this way, according to the information processing system 100 of this embodiment, even if the answer has similar content, it is possible to generate text with different expressions depending on the context of the question, and to output an answer that is easy for the user (questioner) to understand.

[0069] Furthermore, in this embodiment, post-mortem learning is performed based on information from a database external to the development environment, thereby eliminating the burden on developers to input training data. Furthermore, while the above embodiment describes an example of learning from a database related to drugs, highly reliable databases are publicly available in the pharmaceutical field, and post-mortem learning based on these databases can generate dedicated machine learning models that can provide accurate answers. In particular, in the pharmaceutical field, questions about health are expected to be asked, so highly accurate answers are required, making the configuration of this embodiment effective. A text generation service that generates answers using a dedicated model post-mortem trained on drug-related data as in the described embodiment allows users (e.g., patients) to obtain highly accurate answers to drug-related questions, even in environments where pharmacists are absent (e.g., disaster areas, areas with poor transportation access, etc.).

[0070] According to the embodiments of the present invention described above, it is possible to provide an information processing system, an information processing device, a method, and a program for generating a machine learning model with improved answer accuracy.

[0071] Each function of the above-described embodiments of the present invention can be realized by a device-executable program written in C, C++, C#, Java (registered trademark), etc., and the program of this embodiment can be stored and distributed on a device-readable recording medium such as a hard disk drive, CD-ROM, MO, DVD, flexible disk, EEPROM (registered trademark), EPROM, etc., and can also be transmitted over a network in a format that can be used by other devices.

[0072] Although the present invention has been described above with reference to embodiments, the present invention is not limited to the above-described embodiments, and any embodiment that can be conceived by a person skilled in the art is included in the scope of the present invention as long as it exhibits the functions and effects of the present invention. [Explanation of symbols]

[0073] 100...Information processing systems, 110...information processing device, 110a...Development environment terminal, 110b...Development environment server, 110c...database server, 110d...Text generation service providing server, 110e...user terminal, 120...Network, 210...CPU, 220…RAM, 230...ROM, 240...Storage device, 250...Communication I / F, 260...display, 270...input device, 301…Communications Department, 302...display section, 303...Operation unit, 304...database information acquisition unit, 305...learning file generation unit, 306...Variation Creation Department, 307...log recording section, 308...database storage unit, 309...Post-learning processing unit, 310...general-purpose model storage unit, 311... Dedicated model storage area, 312...Text generation unit [Prior art documents] [Patent documents]

[0074] [Patent Document 1] Patent Publication No. 2021-64183

Claims

1. An information processing system for post-learning a machine learning model, An acquisition means for acquiring information stored in the database; a generation means for generating a file for post-learning based on the information acquired by the acquisition means; a processing means for performing post-learning of a machine learning model based on the file; An information processing system comprising:

2. A method for creating variations of questions in conversations further comprising the generating means generates the file based on the variation generated by the generating means and the information. The information processing system according to claim 1 .

3. The creation means creates the variations using a general-purpose machine learning model. The information processing system according to claim 2 .

4. The database is a database specialized in a specific field. The information processing system according to claim 1 .

5. The database is a database storing information about drugs. The information processing system according to claim 4 .

6. a storage means for storing the machine learning model created by the processing means through post-learning together with machine learning models created by the processing means in the past; The information processing system of claim 1 , further comprising:

7. An information processing device, An acquisition means for acquiring information stored in the database; a generation means for generating a file for post-learning based on the information acquired by the acquisition means; a communication means for transmitting the file generated by the generation means to an information processing device that performs post-learning; An information processing device comprising:

8. 1. A method for post-training a machine learning model, comprising: retrieving information stored in a database; generating a file for post-learning based on the information acquired in the acquiring step; performing a post-learning process for a machine learning model based on the file; A method comprising:

9. A program executed by an information processing device, The information processing device A means for retrieving information stored in a database; a generation means for generating a file for post-learning based on the information acquired by the acquisition means; a communication means for transmitting the file generated by the generation means to an information processing device that performs post-learning; A program that functions as a

Images