Information processing method, information processing device, and information processing program

Abstract

Provided is an information processing method according to which a processor executes processing for supplying a large language model with an input task causing the large language model to propose an idea in a specific technical field, such that the input task includes at least one technical field from a technical classification list, which is information pertaining to technical classifications in list form, and causing the large language model to generate an idea related to the specific technical field.

Description

Information processing method, information processing device, and information processing program

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

[0002] As a technology for supporting engineers in generating ideas, for example, Japanese Patent Application Laid-Open No. 2023-179025 is disclosed. Japanese Patent Application Laid-Open No. 2023-179025 discloses a technology for providing an idea support device, an idea support program, and an idea support method that identify ideas suitable for reference and contribute to deriving better ideas.

[0003] Additionally, in recent years, active research has been conducted into technologies related to generative AI, which generates sentences and the like by having a large-scale language model (LLM) execute natural language processing tasks.

[0004] Breakthroughs in solving new problems often come from the combination of past knowledge. Here, inspiration refers to new ideas, insights, and solutions that lead to the resolution of difficult problems. In particular, we consider a situation in which, by the time a problem becomes difficult to solve, all relevant ideas and solutions have been exhausted, and seemingly unrelated elements and ideas play an important role in solving the problem. Large-scale language models possess extensive and detailed technical information, which can potentially provide inspiration for new problems. However, simply asking a large-scale language model for ideas often results in answers that are technically mundane, making it difficult to obtain outputs that are technically insightful and deeply insightful. This is because large-scale language models' parameters are tuned to produce output that is easy for many people to understand. Furthermore, providing detailed information and data increases the risk of errors, so they are adjusted to limit answers to general ones. As a result, simply asking a large-scale language model for ideas makes it difficult to obtain answers that lead to inspiration. "Difficult" here means that the probability of obtaining an answer that leads to inspiration is low. Since the sentences generated by large-scale language models depend on probability, it cannot be said that no answer will be obtained at all, but it means that the probability of generating an answer is clearly low.

[0005] The present disclosure has been made in consideration of the above points, and aims to provide an information processing method, an information processing device, and an information processing program that can more easily obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas on a problem.

[0006] In an information processing method according to a first aspect of the present disclosure, a processor, in response to an input task of having a large-scale language model propose ideas in a specific technical field, provides the large-scale language model with at least one technical field from a technology classification list, which is information on listed technology classifications, and executes a process of having the large-scale language model generate ideas related to the specific technical field.

[0007] According to the first aspect of the present disclosure, by asking a large-scale language model that has been given a technology classification including at least one technology field for ideas, it becomes easier to obtain answers that are rich in technical suggestions and have deep insights, compared to simply asking a large-scale language model for ideas.

[0008] An information processing method according to a second aspect of the present disclosure is the information processing method according to the first aspect, wherein the processor provides the large-scale language model with at least one technical field different from the specific technical field.

[0009] According to a second aspect of the present disclosure, by asking a large-scale language model for ideas that include at least technical fields different from a specific technical field, it becomes easier to obtain answers that are rich in technical suggestions, provide deeper insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0010] An information processing method according to a third aspect of the present disclosure is the information processing method according to the first aspect, wherein the processor extracts a technology classification to be given to the large-scale language model from the technology classification list based on the input task, gives the input task to the large-scale language model so that the extracted technology classification is included, and causes the large-scale language model to generate ideas related to the specific technology field.

[0011] According to the third aspect of the present disclosure, by extracting a technology classification that includes a specific technology field and different technology fields, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration that include the technology field.

[0012] An information processing method according to a fourth aspect of the present disclosure is the information processing method according to the first aspect, wherein the processor provides the input task to a large-scale language model that extracts a technology classification, and extracts the technology classification based on a response from the large-scale language model.

[0013] According to a fourth aspect of the present disclosure, by extracting a technology classification based on a response from a large-scale language model that extracts technology classifications, even a user who is not familiar with technology classifications can generate ideas that specify an appropriate technology classification.

[0014] An information processing method according to a fifth aspect of the present disclosure is the information processing method according to the first aspect, wherein the processor provides the input task to a large-scale language model specialized for extracting technology categories, and extracts the technology categories based on a response from the large-scale language model. Specifically, the large-scale language model specialized for extracting technology categories may be a fine-tuned large-scale language model, a large-scale language model that has undergone short-shot learning on technology information to be considered, or a large-scale language model that has been trained using prompts in which the technology information to be considered is input. Furthermore, the information to be considered may be accumulated in a Retrieval Augmented Generation (RAG), and related information may be extracted from the RAG for the input task of extracting the technology category and provided to the large-scale language model, or the information to be considered may be directly added to the input task, thereby causing the large-scale language model specialized for extracting the technology category to extract the technology category.

[0015] According to a fifth aspect of the present disclosure, by extracting a technology classification based on a response from a large-scale language model specialized in extracting technology classifications, even a user who is not familiar with technology classifications can generate ideas specifying an appropriate technology classification.

[0016] An information processing method according to a sixth aspect of the present disclosure is the information processing method according to the first aspect, further comprising an idea evaluation unit that outputs an evaluation of the idea obtained by the idea generation unit to a large-scale language model.

[0017] According to a sixth aspect of the present disclosure, by extracting a technology classification based on a response from a large-scale language model that extracts technology classifications, even a user who is not familiar with technology classifications can generate ideas that specify an appropriate technology classification.

[0018] An information processing method according to a seventh aspect of the present disclosure is the information processing method according to the first aspect, wherein the processor uses the input task and the technology classification to instruct the large-scale language model to generate ideas specialized for a specific technology classification.

[0019] According to the seventh aspect of the present disclosure, by instructing the large-scale language model to generate ideas specialized for a specific technology category, it becomes easier to obtain ideas specialized for a specific technology category from the large-scale language model compared to when such an instruction is not given.

[0020] An information processing method according to an eighth aspect of the present disclosure is the information processing method according to the first aspect, wherein the large-scale language model is a model in which the technology classification list has been trained in advance.

[0021] According to an eighth aspect of the present disclosure, by asking a large-scale language model with a pre-trained technology classification list for ideas, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0022] An information processing method according to a ninth aspect of the present disclosure is an information processing method according to the eighth aspect, wherein the processor incorporates technical information from a RAG that stores technical information related to at least the technical classification list and provides the technical information to the large-scale language model.

[0023] According to the ninth aspect of the present disclosure, by obtaining answers to input tasks using a RAG having a technology classification list, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0024] An information processing method according to a tenth aspect of the present disclosure is the information processing method according to the eighth aspect, wherein the large-scale language model is a model that has been trained in advance through fine tuning to learn at least technical information related to the technology classification list.

[0025] According to a tenth aspect of the present disclosure, by asking a large-scale language model that has been fine-tuned using a technology classification list for ideas, it is possible to more easily obtain answers that are rich in technical suggestions, provide deeper insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0026] An information processing method according to an eleventh aspect of the present disclosure is the information processing method according to the eighth aspect, wherein the large-scale language model is a short-shot trained model or a model trained using prompts into which technical information related to the technology classification list is input.

[0027] According to an eleventh aspect of the present disclosure, by asking a large-scale language model that has been trained in a short-shot manner or that has been trained with prompts into which technical information related to a technology classification list is input, it is possible to more easily obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0028] An information processing method according to a twelfth aspect of the present disclosure is the information processing method according to the first aspect, wherein the large-scale language model is a model that can refer to the technology classification list.

[0029] According to a twelfth aspect of the present disclosure, by asking a large-scale language model that can be referenced in a technology classification list for ideas, it becomes easier to obtain answers that are rich in technical suggestions, provide deeper insights, and lead to inspiration, compared to when asking a large-scale language model for ideas.

[0030] An information processing method according to a thirteenth aspect of the present disclosure is the information processing method according to the first aspect, wherein the technology classification list is an International Patent Classification.

[0031] According to the thirteenth aspect of the present disclosure, by asking a large-scale language model for ideas that has been given in advance information on the International Patent Classification as a technology classification list, it becomes easier to obtain answers that are rich in technical suggestions, provide deeper insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0032] An information processing method according to a fourteenth aspect of the present disclosure is the information processing method according to the first aspect, wherein the technology classification list is a list in which existing technology classifications are subdivided using independently prepared technology classifications, or a list in which information is added to existing technology classifications.

[0033] According to a fourteenth aspect of the present disclosure, by using a list in which existing technology classifications are subdivided using a technology classification that is uniquely prepared as a technology classification list, or a list in which information is added to existing technology classifications, it is possible to more easily obtain answers that are rich in technical suggestions, provide deeper insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0034] An information processing method according to a fifteenth aspect of the present disclosure is the information processing method according to the first aspect, wherein when the processor provides the input task to the large-scale language model, the processor assigns a priority to the order of access to the technical classification and provides the input task to the large-scale language model.

[0035] According to the fifteenth aspect of the present disclosure, by appropriately prioritizing the order of access to technology classifications, it becomes possible to effectively obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration.

[0036] An information processing method according to a sixteenth aspect of the present disclosure is the information processing method according to the first aspect, wherein the input task is a task of having the large-scale language model propose ideas for solving a technical problem.

[0037] According to the sixteenth aspect of the present disclosure, a large-scale language model given a technology classification list can be made to propose ideas for solving a technical problem.

[0038] An information processing method according to a seventeenth aspect of the present disclosure is the information processing method according to the first aspect, wherein the input task is a task of having the large-scale language model propose a new idea.

[0039] According to the seventeenth aspect of the present disclosure, a large-scale language model can be made to propose new ideas when a technology classification list is given.

[0040] An information processing method according to an eighteenth aspect of the present disclosure is the information processing method according to the first aspect, wherein the number of parameters of the large-scale language model for which the idea generation unit causes ideas to be proposed is 500 billion or more.

[0041] According to the eighteenth aspect of the present disclosure, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to when the number of parameters of a large-scale language model is less than 500 billion.

[0042] An information processing device according to a nineteenth aspect of the present disclosure includes an input unit that acquires an input task that causes a large-scale language model to propose ideas in a specific technical field, and an idea generation unit that provides the large-scale language model with the input task, including at least one technical field, from a technology classification list that is information on listed technology classifications, and causes the large-scale language model to generate ideas related to the specific technical field.

[0043] According to a nineteenth aspect of the present disclosure, by asking a large-scale language model for ideas that has been given a technology classification including at least one technology field, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0044] An information processing program according to a twentieth aspect of the present disclosure provides, to a computer, an input task of having a large-scale language model propose ideas in a specific technical field, at least one technical field from a technology classification list, which is information on listed technology classifications, to the large-scale language model, and causes the large-scale language model to execute a process of generating ideas related to the specific technical field.

[0045] According to a twentieth aspect of the present disclosure, by asking a large-scale language model for ideas that has been given a technology classification including at least one technology field, it becomes easier to obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0046] According to the present disclosure, it is possible to provide an information processing method, an information processing device, and an information processing program that can more easily obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas on a problem.

[0047] 1 is a diagram showing a schematic configuration of an information processing device according to an embodiment of the disclosed technology. FIG. 1 is a block diagram showing a hardware configuration of the information processing device. FIG. 2 is a block diagram showing an example of the functional configuration of the information processing device. FIG. 3 is a block diagram showing an example of the configuration of an idea generation unit. FIG. 4 is a display example of ideas output to a large-scale language model, output to a display unit by an output unit. FIG. 5 is a diagram showing an example of a technology classification list to which physical or chemical characteristics of substances have been assigned. FIG. 6 is a block diagram showing an example of the functional configuration of an information processing device. FIG. 7 is a diagram showing ideas output to GPT-4 and the scoring results for each idea. FIG. 8 is a diagram showing ideas output to GPT-4 and the scoring results for each idea. FIG. 9 is a diagram showing ideas output to GPT-3.5 and the scoring results for each idea. FIG. 10 is a diagram showing ideas output to GPT-3.5 and the scoring results for each idea. FIG. 11 is a block diagram showing another example of the functional configuration of an information processing device. FIG. 12 is a diagram showing an example in which ideas proposed by a large-scale language model 1 are arranged in an International Patent Classification matrix. FIG. 13 is a flowchart showing the flow of information processing by an information processing device.

[0048] Before describing the embodiments of the present disclosure, the circumstances that led the present inventor to the embodiments of the present disclosure will be described.

[0049] Finding solutions to technical challenges is often difficult. For example, OLED light-extraction glass substrates for lighting, which use glass substrates with a glass scattering layer formed using a high-refractive-index glass frit paste, are the only mass-production technology (KIWI technology) that significantly improves light extraction efficiency (see Asahi Glass Research Report 62 (2012) Internet URL: https: / / www.agc.com / innovation / library / pdf / 62-04.pdf, Advanced Glass Substrate for the Enhancement of OLED Lighting Out-Coupling Efficiency Digest of Technical Papers - SID International Symposium, vol. 44, pp. 803-806 (2013)). This technology combines a lens glass material used in optical lenses with a glass film formation technology using glass frit paste for forming the front dielectric used in plasma TVs. At the time, organic EL device manufacturers and material manufacturers were all busy developing light extraction substrates, but because these technologies were localized to glass manufacturers, they ended up taking the lead in commercializing them. As this example shows, while there are many cases where a technical problem can be solved by combining existing technologies, there are also many cases where a solution cannot be reached because the existing technologies are localized or scattered and go unnoticed.

[0050] On the other hand, with the widespread use of large-scale language models, it is believed that the contents of many existing technologies are learned using large-scale language models, and it may be possible to have large-scale language models that have learned the contents of existing technologies propose ideas for solving technical issues.

[0051] However, even though a large-scale language model has learned the contents of many existing technologies, it is difficult to get it to propose the desired ideas simply by inputting prompts, which are commands that instruct the model to output ideas that will solve technical problems. The output of a large-scale language model is technically general, and does not yield output that is rich in technical suggestions and deep insights. This is because the parameters of the large-scale language model are tuned to produce output that is easy for many people to understand. Therefore, as mentioned above, an idea such as applying a glass scattering layer to the light extraction layer of an organic light-emitting diode (OLED) for lighting is a technology that is unfamiliar to many people, and is therefore unlikely to be output.

[0052] Therefore, the present inventors have conducted extensive research into a technology for effectively outputting ideas from a large-scale language model that are rich in technical suggestions, provide deep insights, and lead to inspiration for solutions to technical problems. As a result, as described below, the present inventors have devised a technology that can effectively output ideas from a large-scale language model that are rich in technical suggestions, provide deep insights, and lead to inspiration for solutions to technical problems, compared to simply inputting prompts to cause a large-scale language model to output ideas for solutions to technical problems.

[0053] An example of an embodiment of the present disclosure will be described below with reference to the drawings. The same reference numerals are used throughout the drawings to designate identical or equivalent components and parts. The dimensional proportions of the drawings are exaggerated for illustrative purposes and may differ from the actual proportions.

[0054] 1 is a diagram showing a schematic configuration of an information processing device according to this embodiment. The information processing device 10 according to this embodiment is a device used by a user who wants to obtain some kind of technical idea, and is a device for having a large-scale language model 1 propose ideas. The large-scale language model 1 is a model that performs natural language processing tasks, and is a generative AI that has been trained to output answers in natural language corresponding to input prompts.

[0055] However, because the large-scale language model 1 is a model trained based on a vast amount of information, simply providing the large-scale language model 1 with a prompt to output an idea does not necessarily mean that the large-scale language model 1 will be able to output the idea desired by the user of the information processing device 10. Without providing appropriate information to the large-scale language model 1, the ideas output by the large-scale language model 1 will tend to be mundane. For example, even if one were to ask the large-scale language model 1, "With organic EL lighting, only 20% of the emitted light can be extracted to the outside. One possible solution would be to form a highly refractive light-scattering layer on a glass substrate and then create an organic EL element on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, because an uneven surface on an organic EL device can easily short-circuit the two electrodes. Please use all your glass-related technology knowledge to propose a scattering layer that meets these conditions," it would be difficult for the large-scale language model 1, which has not been given appropriate information in advance, to provide technically suggestive, insightful, and insightful ideas for combining the lens glass material used in optical lenses and the glass frit paste used to form the front dielectric in plasma televisions, as described above.

[0056] Therefore, when causing the large-scale language model 1 to output ideas, the information processing device 10 does not simply provide the large-scale language model 1 with a prompt for idea output, but instead utilizes a technology classification list, which is information on listed technology classifications. In other words, the technology classification information is included in the prompt provided to the large-scale language model 1. This enables the information processing device 10 to cause the large-scale language model 1 to behave as an expert in a specific technology classification. The technology classification list is a list in which a description identifying a certain technical field is paired with a corresponding code. In this embodiment, the International Patent Classification (IPC) is used as the technology classification list. The IPC is a classification based on the technical content of patent documents that is used internationally. The IPC classifies various technologies in an orderly and comprehensive manner, and is information that is comprehensively organized at a certain technical depth. The technology classification list may be a list covering the entire technical field, such as CPC (Cooperative Patent Classification), FI, or F-term (Japanese Patent Classification), or may be an in-house technology classification that classifies a limited range of technology.Furthermore, the technology classification list may be an industry standard or specification such as IEEE 802.11.

[0057] The International Patent Classification is defined in multiple hierarchies, as shown in (1) to (5) below. Taking C03C 3 / 089 as an example, the explanation is as follows: (1) "Section" - indicated by one of the capital letters A to H. Example: C (2) "Class" - a section symbol followed by two additional digits. Example: C03 (3) "Subclass" - a class symbol followed by one additional capital letter. Example: C03C (4) "Main group" - a subclass symbol followed by one to three digits, a slash, and the number 00. Example: C03C 3 / 00 (5) "Subgroup" - a subdivision item below a main group. Each subgroup symbol consists of a subclass symbol followed by one to three digits of the main group, a slash, and at least two digits other than 00. Example: C03C 3 / 089

[0058] By providing a prompt containing technology classification information to a large-scale language model 1 that has learned a technology classification list, which is information on listed technology classifications, or that can refer to a technology classification list, the information processing device 10 of this embodiment can more easily obtain answers that are rich in technical suggestions, provide deep insights, and lead to inspiration, compared to simply asking a large-scale language model for ideas.

[0059] The large-scale language model 1 may be constructed on a device different from the information processing device 10, for example, on a server on the Internet, or may be constructed inside the information processing device 10. In the following description, the large-scale language model 1 is assumed to be constructed on a device different from the information processing device 10. The large-scale language model 1 is, for example, a large-scale language model of GPT-3.5 and 4 (OpenAI), Gemini Pro 1.5 (Google), or LLaMa-2 (Meta AI).

[0060] FIG. 2 is a block diagram showing the hardware configuration of the large-scale language model 1 and the information processing device 10.

[0061] 2, the large-scale language model 1 and the information processing device 10 include a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input unit 15, a display unit 16, and a communication interface (I / F) 17. Each component is connected to each other via a bus 19 so as to be able to communicate with each other.

[0062] The CPU 11 is a central processing unit that executes various programs and controls each component. That is, the CPU 11 reads the programs from the ROM 12 or the storage 14 and executes the programs using the RAM 13 as a work area. The CPU 11 controls the above components and performs various arithmetic processing in accordance with the programs recorded in the ROM 12 or the storage 14. In this embodiment, the ROM 12 or the storage 14 stores an information processing program that causes the large-scale language model 1 to propose ideas.

[0063] The ROM 12 stores various programs and various data. The RAM 13 temporarily stores programs or data as a working area. The storage 14 is configured with a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory, and stores various programs including an operating system and various data.

[0064] The input unit 15 includes a pointing device such as a mouse and a keyboard, and is used to input various types of information.

[0065] The display unit 16 is, for example, a liquid crystal display, and displays various information. The display unit 16 may be a touch panel type and function as the input unit 15.

[0066] The communication interface 17 is an interface for communicating with other devices, and uses standards such as Ethernet (registered trademark), FDDI, and Wi-Fi (registered trademark).

[0067] When executing the information processing program, the information processing device 10 uses the hardware resources described above to realize various functions. The functional configuration realized by the information processing device 10 will be described below.

[0068] FIG. 3 is a block diagram showing an example of the functional configuration of the information processing device 10.

[0069] 3, the information processing device 10 has, as its functional components, an input unit 101, a technology classification preparation unit 102, an idea generation unit 104, and an output unit 105. Each functional component is realized by the CPU 11 reading and executing an information processing program stored in the ROM 12 or the storage 14.

[0070] The input unit 101 acquires an input task to be given to the large-scale language model 1 to cause the large-scale language model 1 to propose an idea. The input task may be input by a user of the information processing device 10 using the input unit 15.

[0071] The input task acquired by the input unit 101 may include technical issues described by the user, settings of technical information to be used in proposing ideas, technology classification information, and settings of the large-scale language model 1. The technical issues may specifically include a description of the current issues the user is aware of and a description of what the user wants. The settings of technical information to be used in proposing ideas may specifically include information on a technology classification list the user wants to use and settings of specific technology classifications. The settings of the large-scale language model 1 may specifically include the type of model to be used, the number of ideas to be created, and parameters for adjusting the diversity or predictability of text generated by the large-scale language model 1.

[0072] The technology classification preparation unit 102 executes processing related to the preparation of technology classifications to be given to the large-scale language model 1. Using a technology classification list 103, which is information on listed technology classifications, the technology classification preparation unit 102 executes processing related to preparations for having the large-scale language model 1 behave as an expert in a specific technology classification.

[0073] As part of the process of preparing a technology classification, the technology classification preparation unit 102 extracts a technology classification to be given to the large-scale language model 1. For example, if a technology classification is included in an input task input by a user, the technology classification preparation unit 102 may extract the technology classification from the technology classification list 103 based on the content of the input task.

[0074] Furthermore, the technology classification preparation unit 102 may combine a technology classification extracted from all of the technology classifications in the technology classification list 103 based on an input task input by a user with a technology classification extracted by a large-scale language model based on an input task input by a user. The large-scale language model from which a technology classification is extracted based on an input task may be the same model as the large-scale language model 1, or may be a different model.

[0075] The idea generation unit 104 generates a prompt based on the input task acquired by the acquisition unit 101, provides the generated prompt to the large-scale language model 1, and obtains an output from the large-scale language model 1, thereby executing a process of having the large-scale language model 1 propose an idea. When generating the prompt, the idea generation unit 104 includes the technology classification extracted by the technology classification preparation unit 102 in the prompt.

[0076] The large-scale language model 1 from which the idea generation unit 104 proposes ideas may be prepared for each technology classification, or may be prepared to support all technology classifications. Here, "prepared" may mean instructing the large-scale language model 1 to become an expert in a specified technology classification. When a large-scale language model 1 is prepared for each technology classification, and the idea generation unit 104 specifies multiple technology classifications to have the large-scale language model 1 propose ideas, the idea generation unit 104 causes the large-scale language model 1 to propose ideas for each of the corresponding multiple technology classifications.

[0077] When a large-scale language model 1 is prepared for one or more technology classifications and ideas are proposed from each of the multiple large-scale language models 1, the idea generation unit 104 may perform processing to consolidate the multiple idea proposals from the large-scale language models 1. For example, the idea generation unit 104 may inquire of a large-scale language model 1 what it thinks about an idea proposed by another large-scale language model 1. Furthermore, the idea generation unit 104 may inquire of a large-scale language model acting as a chairperson for consolidating the multiple idea proposals what it thinks about the idea proposed by the large-scale language model 1 or which idea it thinks is superior.

[0078] 4 is a block diagram showing an example of the functional configuration of the large-scale language model 1. The large-scale language model 1 shown in FIG. 4 functions as expert LLMs 111A and 111B that act as experts in a certain technical field, and a coordinating LLM 112 that compiles ideas proposed by the expert LLMs 111A and 111B. The large-scale language model 1 also includes a Retrieval Augmented Generation (RAG) that accumulates technology-related data 113 related to technologies in a technology classification list, and technical information from the RAG may be incorporated into the prompts received by the LLMs 11A and 11B.

[0079] The expert LLMs 111A and 111B may be fine-tuned to a specific field of technology classified by a technology classification list. For example, if an idea related to glass is desired, the expert LLMs 111A and 111B may be fine-tuned to technical information related to glass.

[0080] The output unit 105 outputs the ideas that the idea generation unit 104 has caused the large-scale language model 1 to propose to the display unit 16. Fig. 5 shows an example of a display of ideas that have been output to the large-scale language model 1 by the output unit 105 to the display unit 16. Reference numeral 201 denotes an area in which an input task entered by a user of the information processing device 10 is displayed. Reference numeral 202 denotes an area in which an idea that has been output to the large-scale language model 1 by the idea generation unit 104 is displayed. By the output unit 105 displaying information on the display unit 16 in this way, the user of the information processing device 10 can check the ideas that have been output by the large-scale language model 1.

[0081] An example of an input task that a user might enter is as follows. This input task includes the technical problem described by the user and the technical information to be used in the idea proposal. "With OLED lighting, only 20% of the emitted light can be extracted. One solution is to form a highly refractive, light-scattering layer on a glass substrate and then create an OLED on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, since an uneven surface on an OLED device can easily short-circuit the electrodes. Please propose a scattering layer that meets these conditions using the technology associated with the IPC numbers listed in the attached C03B and C03C documents. Please indicate which number you used and how you created the scattering layer." In this case, a list of international patent classifications (IPCs) was pre-loaded into large-scale language model 1 as technology classification list 103 based on the input task. The international patent classifications pre-loaded into large-scale language model 1 were 306 items in C03B (dissolving, molding, and processing) and 193 items in C03C (composition, enamel, glaze, and surface treatment).

[0082] When such an input task is input, the technology classification preparation unit 102 refers to the technology classification list 103 and extracts C03B and C03C of the International Patent Classification as technology classifications. Then, the idea generation unit 104 generates a script for causing the large-scale language model 1 to propose an idea as an expert in C03B and C03C of the International Patent Classification, and provides this to the large-scale language model 1. The large-scale language model 1 outputs a response to the script, and the output unit 105 outputs the response from the large-scale language model 1 to the display unit 16 as an idea that has been proposed by the large-scale language model 1.

[0083] When the above input task is input to the information processing device 10 according to this embodiment, the large-scale language model 1 can output, for example, the following idea. Note that the following output is obtained when GPT-4.0 is used for the large-scale language model 1. 1. C03C 3 / 089: Boron-containing glass composition. A boron-rich glass layer is formed on a substrate to achieve a high refractive index. 2. C03C 17 / 42: Surface treatment with organic materials and non-metallic coatings. To ensure surface smoothness, surface treatment is performed with specific organic materials. 3. C03C 25 / 10: Precision coating technology. A high refractive index material is uniformly coated on a nanoscale to enhance surface smoothness. 4. C03C 3 / 062: Low silica content glass. A high refractive index is achieved by adjusting the silica content and adding specific additives (e.g., lanthorium). 5. C03C 3 / 093: Zinc- or zirconium-containing glass. A glass layer containing zinc or zirconium with a high refractive index is formed on a substrate. 6. C03C 17 / 34: Multiple coating layers. Multiple coating layers with different compositions are applied to simultaneously achieve a light scattering effect and surface smoothness. 7. C03B 5 / 235: Optimization of the Glass Melting Process: Specific additives are introduced during the glass melting process to increase the refractive index and improve the efficiency of the scattering layer. 8. C03C 17 / 00: Glass Surface Coating: Special surface coating techniques are used to form a high-refractive-index scattering layer, followed by a protective layer to smooth the surface. 9. C03C 12 / 00: Application of Powdered Glass: A scattering layer is formed by selecting high-refractive-index powdered glass and uniformly dispersing it on a substrate using techniques such as the Soligel method. 10. C03C 8 / 00: Enamels and Glazes: Specific enamel and glaze compositions are adjusted and applied as a scattering layer on a high-refractive-index glass substrate. These materials can improve both optical properties and surface smoothness.

[0084] Although hallucinations were observed in the output of the large-scale language model 1, compared to when technology classification information was not provided, the output contained ideas that referred to glass composition and processes, and were rich in technological suggestions and deeply insightful. Furthermore, idea "10." from the output of the large-scale language model 1 relates to the idea of ​​combining a glass film formation technology using a lens glass material used in optical lenses with a glass frit paste for forming a front dielectric in plasma televisions. In this input task, the specific technology field that the large-scale language model 1 is asked to propose is an idea regarding the materials and processes used to form a high-refractive-index internal scattering layer on a glass substrate to improve the light extraction efficiency of organic electroluminescence (EL) lighting. In this case, the specific technology field is a semiconductor light-emitting element corresponding to organic EL (F21K9), a semiconductor device or solid-state device or their components (H01L21), and an optical element characterized by the materials used (G02B1). The inclusion of technology classifications such as glass melting, molding, and processing (C03B) and glass composition, enamel, glaze, and surface treatment (C03C) is important for generating inspirational ideas. As in the above example, the idea generation unit 104 may provide only technology classifications of technology fields different from the specific technology field to the large-scale language model 1, or may include technology classifications of technology fields different from the specific technology field and some or all of the technology classifications of the above-mentioned specific technology field. In another example, when it is desired to output a synthesis method for halogenating a hydrocarbon compound, the specific technology field is CO7C17 (production of halogenated hydrocarbons).

[0085] In the above example, IPC numbers corresponding to specific technical fields are divided into subclasses, but this is not limited to this. For example, IPC numbers corresponding to specific technical fields may be divided into main groups, or into subgroups, or, if multiple specific technical fields are established, these may be combined. IPC numbers corresponding to specific technical fields can be determined depending on the technical scope of the idea to be generated and the granularity of classification for each technical field (e.g., the number of main groups and subgroups). The same applies to FI and other technical classifications.

[0086] When generating ideas, the idea generation unit 104 may instruct the large-scale language model 1 to generate ideas specialized in a specific technical field from the input task and the technology classification. In the above example, the idea generation unit 104 may instruct the large-scale language model 1 to generate ideas specialized in a semiconductor light-emitting element corresponding to organic electroluminescence (F21K9), a semiconductor device or solid-state device or a component thereof (H01L21), or an optical element characterized by the materials used (G02B1). Even if the idea generation unit 104 provides only the technology classification of a specific technical field to the large-scale language model 1, it is possible for the idea generation unit 104 to generate ideas that are related to the specific technical field, are rich in technical suggestions, and have deep insights.

[0087] As a comparative example, we will show an example of querying large-scale language model 1 for ideas without providing it with a technology classification in advance and without specifying an international patent classification in the input task. Note that the output below is from large-scale language model 1 using GPT-4.0. The input task was, "With organic EL lighting, only 20% of the emitted light can be extracted to the outside. One possible solution is to form a highly refractive, light-scattering layer on a glass substrate and then create an organic EL element on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, because in organic EL devices, if the base is uneven, the two electrodes will easily short-circuit. Please use all the glass-related technology you know to propose a scattering layer that meets these conditions." The response from large-scale language model 1 was as follows: "In order to improve the light extraction efficiency of OLED lighting, the design of the light-scattering layer is important. To propose a scattering layer that meets the requirements in your question, namely a high refractive index (1.9 or higher) and surface smoothness, the following technologies should be considered. 1. Selection of High-Refractive Index Materials Titanium Oxide (TiO2): Titanium oxide has a high refractive index (approximately 2.5-2.9) and is also known as a photocatalyst. When used in the form of nanoparticles, it can function as a scattering layer while maintaining a high refractive index. Zirconium Oxide (ZrO2): Zirconium oxide also has a high refractive index (approximately 2.2). Its high thermal stability makes it suitable for optical applications. 2. Surface Smoothing Technologies Sol-Gel Method: The sol-gel method is a method of forming a uniform coating film on the nanoscale using the hydrolysis and condensation reaction of metal alkoxides. Using this technique, high-refractive index materials can be formed as a uniform, smooth layer on glass substrates. Spin Coating: Spin coating is a technique in which a drop of solution is placed on a substrate and rotated at high speed to form a thin, uniform film. This method makes it possible to coat a high-refractive index solution as a smooth layer on a substrate. 3. Light-scattering layer design Optimizing the refractive index contrast between layers: Optimizing the refractive index contrast between the high-refractive-index layer and the glass substrate can enhance the light-scattering effect. Stacking multiple layers with different refractive indices can improve light extraction efficiency.Introducing microstructures: By introducing microstructures to the surface (for example, patterns formed by nanoimprint lithography), it is possible to further enhance the light scattering effect. This allows light that tends to be trapped inside to be extracted more effectively.

[0088] Thus, when the large-scale language model 1 is queried for ideas without providing a technology classification to the large-scale language model 1 in advance and without specifying an international patent classification in the input task, the response obtained from the large-scale language model 1 includes a proposal for using titanium oxide or zirconium oxide, which are common high-refractive index materials, and for the planarization process, the well-known sol-gel method. Furthermore, the response from the large-scale language model 1 only suggests scattering using a common nanoimprint structure. These proposals merely propose general technologies for the necessary technical requirements, and do not offer any suggestive or insightful glass scattering layer. Therefore, it is difficult for the large-scale language model 1 to output an idea such as the combination of a glass film formation technology using a lens glass material used in optical lenses and a glass frit paste for forming a front dielectric used in plasma televisions. Therefore, by providing the large-scale language model 1 with a technology classification list in advance, the information processing device 10 can extract technically suggestive, insightful, and inspirational ideas from the large-scale language model 1.

[0089] For example, if the input task input by the user does not include a technology category, the technology category preparation unit 102 may extract a technology category to be given to the large-scale language model 1 from all of the technology categories in the technology category list 103. For example, if the input task input by the user does not include a technology category, the technology category preparation unit 102 may cause the large-scale language model to extract a technology category based on the input task input by the user. The large-scale language model for extracting a technology category may be a model different from the large-scale language model 1 for proposing ideas, or may be the same model.

[0090] For example, the following is a case where the input unit 101 receives the following input task: "With organic EL lighting, only 20% of the emitted light can be extracted to the outside. One possible solution is to form a highly refractive light-scattering layer on a glass substrate and then create an organic EL on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, because with organic EL devices, if the base is uneven, the two electrodes will easily short-circuit. Please propose a scattering layer that meets these conditions. When doing so, please state which IPC number you used and also show how to create the scattering layer."

[0091] In this case, the technology classification preparation unit 102 prompts the large-scale language model for extracting technology classifications to extract an appropriate technology classification (here, the International Patent Classification) from the content of the input task, and uses the response from the large-scale language model to extract a technology classification for idea generation.

[0092] When outputting ideas for solving a problem from all technology categories to the large-scale language model 1, which technology category to start with is important for efficiently obtaining inspiration. Therefore, the large-scale language model for extracting technology categories may be specialized for extracting technology categories. Specifically, the large-scale language model for extracting technology categories may be trained with data related to technologies related to the idea to be output by fine tuning or the like, and the technology categories may be output to the large-scale language model using RAG (Retrieval Augmented Generation).

[0093] Fine-tuning the large-scale language model for extracting technology classifications makes it easier for the large-scale language model to output technology classifications that include technologies for solving the input task entered by the user, compared to a case where the model is not fine-tuned. For example, an international patent classification that includes the technology of the solution idea for the glass scattering layer for light extraction in the above-mentioned organic EL lighting was proposed using both a non-fine-tuned large-scale language model and a fine-tuned large-scale language model. The fine-tuned language model used GPT-3.5 turbo-1106, and hyperparameters were set to Auto except for Epochs, which was set to 7. The training data, including the characteristics of frit paste and films fired from the frit paste, were prepared in the form of questions and answers. Tables 1 to 3 show the training data used for fine-tuning.

[0094]

[0095]

[0096]

[0097]

[0098] As a result, as shown in Tables 5 and 6, the fine-tuned large-scale language model output C03C, which includes the technology of glass scattering layers, at a higher rank. In Tables 5 and 6, "Frequency" refers to the frequency of occurrence of an international patent classification, and "Probably" refers to the probability (%) of occurrence of the international patent classification. Fine-tuning a large-scale language model that extracts a technology classification using the desired technical information makes it possible to extract a technology classification that takes that technical information into account. However, the large-scale language model may also be a model trained using short-shot learning on the desired technical information or a model trained using prompts in which the desired technical information is input. Furthermore, other methods are also applicable, such as storing the desired information in a RAG, extracting related information from the RAG and providing it to the large-scale language model as an input task for extracting a technology classification, or adding the desired information directly to the input task. As shown here, there are no particular restrictions on the large-scale language model for extracting technology classifications, but it is desirable for the number of parameters of the model to be 100 billion or more, such as GPT-3.5, and it is particularly desirable for the number of parameters to be 500 billion or more, such as GPT-4.

[0099]

[0100]

[0101] The International Patent Classification (IPC) is a classification system based on the technical content of patent documents that is used internationally. However, the granularity of the classification varies depending on the technical field. On the other hand, even if the granularity of a technical field is coarse in the IPC, other classification systems may have finer granularity. For example, the FI and F-term systems compiled by the Japan Patent Office (JPO) classify technical fields based on multiple technical perspectives that differ from the IPC. Furthermore, even if the granularity of a technical field is coarse in the IPC, companies working in that field may categorize the technologies in that field more precisely. Therefore, the information processing device 10 may extract a technology classification using FI, F-terms, or a technology classification list created independently by a company, etc., instead of or in addition to the IPC, as the technology classification list 103, add the extracted technology classification information to the prompt, and output ideas to the large-scale language model 1. For example, C03C 21 / 00 has the same definition in both the International Patent Classification and FI as "surface treatment of glass not in the form of fibers or filaments by diffusing ions or metals onto the surface," but while the International Patent Classification has no subclasses below it, FI has subclasses below it, allocating 101 to "chemical strengthening." In this way, the information processing device 10 can output more appropriate ideas to the large-scale language model 1 by providing the large-scale language model 1 with FI, F-terms, or information independently classified by companies, etc., in addition to the International Patent Classification as a technology classification list 103, or by allowing the large-scale language model 1 to refer to such technology classification list 103.

[0102] The technology classification list 103 may also be a list of technology classifications that include technical terminology of the inputted technical subject.

[0103] Technical classifications such as the International Patent Classification (IPC) describe one technology per code, and often specify substances in subordinate codes. For example, the International Patent Classification (IPC) B01J 21 / 00 describes "a catalyst consisting of magnesium, boron, aluminum, carbon, silicon, titanium, zirconium, or hafnium, or their oxides or hydroxides," with each substance specified in subordinate codes. However, substances may have multiple characteristics, and sometimes a problem can only be solved by utilizing multiple characteristics possessed by a substance. In other words, simply using a technology classification and its corresponding description may not always provide the inspiration needed to solve the problem.

[0104] Therefore, the idea generating unit 104 may cause the large-scale language model 1 to propose ideas, with the physical or chemical characteristics of substances added to the technology classification list 103 to be given to the large-scale language model 1.

[0105] 6 is a diagram showing an example of a technical classification list 103 to which the physical or chemical characteristics of substances are assigned. The IPC code column, hierarchy column, and title column are columns in which information prepared in advance in the International Patent Classification is stored. The feature column is a column in which information on the physical or chemical characteristics of the substance described in the title column is stored. For example, the feature column may describe information output by a large-scale language model in response to a query about the physical or chemical characteristics of the substance.

[0106] Below, a comparative example is a case where the physical or chemical characteristics of a substance are not assigned to the technology classification list 103, and an example is a case where the physical or chemical characteristics of a substance are assigned, and the results of the idea generation unit 104 having the large-scale language model 1 propose ideas are compared.

[0107] When applying an InGaZnOx transistor (IGZO-TFT) to a DRAM (Dynamic Random Access Memory), the channel length is only a few tens of nanometers. Therefore, when IGZO is reduced with titanium, the reduction extends to the area around the electrodes, resulting in a low-resistance connection between the source and drain electrodes, preventing the transistor from functioning. Therefore, when applying IGZO-TFT to a DRAM, it is necessary to reduce the IGZO only in a localized area very close to the electrodes.

[0108] The source and drain electrodes of IGZO-TFTs for DRAM applications are made of materials that satisfy the following two requirements: 1. To promote the reaction in a low-temperature process, H ２ 1. A metal electrode with catalytic ability that dissociates molecules into active atomic hydrogen. 2. A hydrogen-permeable metal electrode with moderate solubility that is favorable for high-speed hydrogen diffusion and easy absorption and release of hydrogen, allowing for the rapid transport of active hydrogen to the interface. A material that simultaneously achieves the above 1. and 2. above is palladium, which combines catalytic performance, hydrogen storage capacity, and charge transport capacity. The International Patent Classification does not state that palladium has such performance and capabilities. However, by using a technical classification list 103 that adds the physical or chemical characteristics of substances to the International Patent Classification, the idea generation unit 104 can more easily obtain solution ideas from the large-scale language model 1.

[0109] (Example) An example will be given below in which the idea generating unit 104 obtains ideas from the large-scale language model 1 when the input unit 101 receives the following input task: "Transistors (IGZO-TFTs) made of amorphous oxide semiconductor InGaZnOx, developed in previous research, can be fabricated at room temperature and exhibit mobility several tens of times that of amorphous silicon, and are therefore widely used in flat panel displays (FPDs). In this case, the channel length of the transistor is 5 to 10 microns. The contact resistance between the IGZO and the source and drain electrodes is made ohmic by reducing (removing oxygen from) the IGZO in contact with the electrodes and increasing the electron concentration. Specifically, this is possible when annealing is performed using Ti for the electrodes, as the IGZO loses oxygen to the Ti. On the other hand, if this is to be used in a DRAM, the channel length is only several tens of nanometers. As mentioned above, If IGZO is reduced, it will also reduce the area around the electrode, resulting in a low-resistance connection between the source and drain electrodes, preventing it from functioning as a transistor. Therefore, to use IGZO in DRAM, a reduction that acts only on a very localized area near the electrode, rather than reducing IGZO with Ti, must be used. Please solve this problem using the IPC code technology in the attached file. Please observe the following: Use the information in the attached list for the IPC code, not your own information. Please list 10 ideas based on these conditions. When doing so, please include the IPC number, the name of the specific substance in the attached list that you applied, and why that technology is effective for localized IGZO reduction.

[0110] When this input task was received, the idea generation unit 104 obtained a response including the following answer from the large-scale language model 1 given the technology classification list 103 shown in FIG. 6: "IPC code: B01J 23 / 44 Technology summary: Catalyst using palladium Physicochemical characteristics of the material: Excellent hydrogen absorption ability Reason for application in reduction: Palladium has the ability to utilize hydrogen to promote local reduction reactions, so it is effective for the local reduction of IGZO."

[0111] (Comparative Example) On the other hand, when the input unit 101 receives the above input task while providing the large-scale language model 1 with a technology classification list 103 that does not include the physical or chemical characteristics of the substances, the idea generation unit 104 obtains the idea of ​​using platinum group metals from the large-scale language model 1, but is unable to obtain an answer regarding the material palladium or the key technology for producing IGZO-TFTs, namely reduction with hydrogen.

[0112] Therefore, by providing the large-scale language model 1 with the technical classification list 103 to which the physical or chemical characteristics of substances have been assigned, the idea generation unit 104 can obtain effective ideas for technical problems from the large-scale language model 1 compared to when the physical or chemical characteristics of substances have not been assigned to the technical classification list 103.

[0113] The information processing device 10 may evaluate ideas proposed for technical problems by the large-scale language model 1. Fig. 7 is a block diagram showing an example of the functional configuration of the information processing device 10. The information processing device 10 shown in Fig. 7 has a configuration in which an idea evaluation unit 106 is added to the configuration shown in Fig. 3.

[0114] The idea evaluation unit 106 evaluates the ideas output by the idea generation unit 104 to the large-scale language model 1. Specifically, the idea evaluation unit 106 evaluates the ideas output by the large-scale language model 1 in terms of novelty, inventive step, and feasibility. The idea evaluation unit 106 generates a prompt for the large-scale language model for evaluation prepared for idea evaluation to evaluate the idea output by the large-scale language model 1, and inputs the generated prompt to the large-scale language model for evaluation. The large-scale language model for evaluation may be previously trained on technical content in the technical classification trained by the large-scale language model that output the idea. Furthermore, the large-scale language model for evaluation may be the same model as the large-scale language model 1 that proposes the idea, or may be a different model.

[0115] The idea evaluation unit 106 may use the large-scale language model for evaluation when evaluating the ideas output by the large-scale language model 1 in terms of novelty, inventive step, and feasibility. For example, the idea evaluation unit 106 may have the large-scale language model for evaluation evaluate the ideas output by the large-scale language model 1 on a scale of 10 points. Specifically, the idea evaluation unit 106 provides the following prompt to the large-scale language model for evaluation: "Please rate the proposed idea on a scale of 10 points for novelty, inventive step, and feasibility. Use all of your skills when evaluating."

[0116] Furthermore, the idea evaluation unit 106 may cause the large-scale language model for evaluation to evaluate the similarity of the idea output by the large-scale language model 1 to examples of inspiration prepared in advance.

[0117] The output unit 105 outputs to the display unit 16 the ideas that the idea generation unit 104 has caused the large-scale language model 1 to propose, and the results of the idea evaluation unit 106 evaluating the ideas that the large-scale language model 1 has caused to propose, to the display unit 16. Table 7 shows an example of the evaluation results of ideas by the idea evaluation unit 106.

[0118]

[0119] The output unit 105 also stores the ideas that the idea generation unit 104 has the large-scale language model 1 propose, and the results of the idea evaluation unit 106 evaluating the ideas that the large-scale language model 1 has proposed, in the storage 14 in a predetermined file format, for example, a CSV (Comma-Separated Values) format in which the ideas and scores are written separated by commas.

[0120] A user who has seen the ideas generated by the idea generation unit 104 in the large-scale language model 1 and output by the output unit 105 may want to specify a different technology classification and have the large-scale language model 1 generate ideas again. For this reason, the information processing device 10 may have the user manually select a technology classification and have the large-scale language model 1 generate ideas again using the technology classification selected by the user.

[0121] The performance of large-scale language models is improving daily as the number of parameters increases. In particular, GPT-4, a large-scale language model created by OpenAI, is estimated to have more than one trillion parameters (Reference: Multibody Models Generated from Natural Language, Multibody System Dynamics (2023) Internet <URL: https: / / doi.org / 10.1007 / s11044-023-09962-0>). The idea of ​​the glass scattering layer of organic EL lighting described above is output using various large-scale language models with different numbers of parameters, thereby demonstrating the number of parameters of large-scale language model 1 suitable for this embodiment. In the following description, the number of parameters of large-scale language model 1 suitable for this embodiment is verified using OpenAI API, an API (Application Programming Interface) related to natural language processing technology provided by OpenAI. Below, we will show the results of verification using four large-scale language models created by OpenAI: GPT-2, GPT-3, GPT-3.5, and GPT-4.

[0122] The similarity of the ideas generated by each model was evaluated by using a large-scale language model for evaluation to see whether they were similar to ideas that could lead to inspiration that had been prepared in advance.An example of an idea that could lead to inspiration is, "If you prepare a frit paste by mixing high-refractive-index glass frit with ceramic particles of a different refractive index, a resin such as ethyl cellulose, and a solvent such as terpineol, and then print this on a glass substrate by screen printing, and then heat it, remove the binder, and bake it near its softening point, a high-refractive-index scattering layer with a smooth surface can be formed on the glass substrate."

[0123] (GPT-4) The prompts used during verification with GPT-4 are as follows: "(You are an expert in {use_IPCs[f'IPC_subclass_{IPC_subclass[i]}']} of the International Patent Classification (IPC). Please output in Japanese. Please also describe in detail how and why you created the idea.) Here is your task: "{Request}" To solve your task, use the techniques of the following IPC numbers. "{IPC_subclass[i]}" An overview of these IPC numbers is as follows. Please also refer to this explanation. "{completion0_list}" Create {n} ideas to solve your task. Before you start writing down your ideas, number them, such as "1.", "2.", etc. When doing so, please tell us all the IPC numbers you used."

[0124] In the above prompt, {Request} is a description of the issue or request entered by the user of the information processing device 10, {IPC_subclass[i]} is the International Patent Classification number (e.g., C03B) of the entered subclass, and {completion0_list} is a summary that is explained by the large-scale language model by entering the IPC number up to the subgroup and the title of the IPC number.

[0125] The parameters used to verify idea generation using GPT-4 are as follows: model_name = "gpt-4" temperature = 0.7 The data to be input into GPT-4 is as follows: The following data indicates that the content specified in the Request will be output to GPT-4 in two parts, with International Patent Classifications C03B and C03C. Request = "With OLED lighting, only 20% of the emitted light can be extracted. One possible solution is to form a highly refractive, light-scattering layer on a glass substrate and then create the OLED on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, as OLED devices easily short-circuit the electrodes if the underlying surface is uneven. Please show us how to create a scattering layer (made of glass). " IPC_subclass = ["C03B", "C03C"]

[0126] Figures 8A and 8B show the ideas output by GPT-4 and the scoring results for each idea. While some of the ideas shown in Figures 8A and 8B are not feasible, they offer rich technical suggestions and deep insights compared to the results obtained when applying the large-scale language model described below. This tendency is particularly evident when using C03C as the International Patent Classification. Ideas Nos. 14, 16, and 18 in Figure 8B were related to the idea of ​​a glass scattering layer for organic electroluminescence (EL) lighting.

[0127] (GPT-3.5) The prompts used when verifying with GPT-3.5 are as follows: "(You are an expert in {use_IPCs} of International Patent Classification (IPC). Please output in Japanese. Please also describe in detail how and why you created the idea.) Your task is this: "{Request}" Use the following IPC number techniques to solve your task. "{use_IPCs}" An overview of these IPC numbers is as follows. Please also refer to this explanation. "{completion0_list}" Create {n} ideas to solve your task. Before you start writing down your ideas, number them, such as "1.", "2.", etc. When doing so, please tell us all the IPC numbers you used."

[0128] In the above prompt, {Request} is a description of the problem or request entered by the user of the information processing device 10, {use_IPCs} is the International Patent Classification number (e.g., C03B 1 / 00) up to the main group with the entered subtype, and {completion0_list} is a summary that is explained by the large-scale language model by entering the IPC number up to the subgroup and the title of the IPC number.

[0129] The parameters used to verify idea generation using GPT-3.5 are as follows: model_name = "gpt-3.5-turbo" temperature = 0.7 n (number of ideas generated) = 20 The data input to GPT-3.5 is as follows: The following data indicates that GPT-3.5 will output two international patent classifications, C03B and C03C, for the content specified in the Request. Request = "With OLED lighting, only 20% of the emitted light can be extracted. One possible solution is to form a highly refractive, light-scattering layer on a glass substrate and then create the OLED on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, as OLED devices easily short-circuit between the two electrodes if the underlying surface is uneven. Please show us how to create a scattering layer (made of glass). IPC_subclass = ["C03B", "C03C"]

[0130] Figures 9A and 9B show the ideas output by GPT-3.5 and the scoring results for each idea. Compared to the results of applying GPT-4, the 20 ideas shown in Figures 9A and 9B appear to be more general in terms of technology and less insightful. Furthermore, the 20 ideas shown in Figures 9A and 9B did not include any ideas related to the aforementioned inspiration for the glass scattering layer in OLED lighting.

[0131] (GPT-3) The prompts used in the GPT-3 verification are the same as those used in the GPT-3.5 verification. The parameters used in the GPT-3 idea generation verification are as follows: model_name = "davinci-002" temperature = 0.7 max_tokens = 500 n (number of ideas generated) = 20

[0132] The data to be input into GPT-3 is as follows. The following data indicates that the content specified in prompt will be output to GPT-3 for two international patent classifications, C03B and C03C. prompt = " With OLED lighting, only 20% of the emitted light can be extracted to the outside. One solution is to form a highly refractive, light-scattering layer on a glass substrate and then create the OLED on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, as an OLED device's electrodes can easily short-circuit if the underlying surface is uneven. As an expert in {IPC_subclass} for International Patent Classification (IPC), I will explain an example of a solution using {IPC_subclass} technology. The method for creating a scattering layer (made of glass) is as follows:" IPC_subclass = ["C03B", "C03C"]

[0133] Figures 10A and 10B show excerpts of ideas output by GPT-3. None of the ideas output by GPT-3 captured the intent of the input, and as a result, they were deemed unevaluable.

[0134] (GPT-2) The prompts used in the GPT-2 verification are the same as those used in the GPT-3.5 verification. The parameters used in the verification of idea generation using GPT-2 are as follows. Note that in GPT-2, the process of creating one idea per response is repeated five times. model_name = "gpt2-xl" max_new_tokens = 100 n (number of ideas generated) = 20

[0135] Furthermore, the data input to GPT-2 is the same as the data input to GPT-3.

[0136] The ideas output by GPT-2 were not at all sufficient in Japanese to begin with, and this was a problem that went beyond the evaluation of the ideas.

[0137] The evaluation results of each model are summarized in Table 8 below.

[0138]

[0139] As shown in Table 8, the ideas output by GPT-4 included multiple ideas that could be considered inspirations, and the technical descriptions were detailed. On the other hand, the ideas output by GPT-3.5, while not significantly different in evaluation score (total) from GPT-4, seemed technically shallow and did not include any ideas that could be considered inspirations. Furthermore, GPT-2 and GPT-3 were unable to output ideas that reflected the input intent, and therefore could not be evaluated.

[0140] Based on the above verification results, it is particularly preferable that the large-scale language model 1 that proposes ideas for solving technical problems has a parameter count of preferably 500 billion or more, more preferably 1 trillion or more, and for example, that GPT-4 be used.

[0141] There are often restrictions on the number of characters or tokens in a prompt input to the large-scale language model 1, and in order to have such a large-scale language model 1 generate ideas, the information processing device 10 shown in Fig. 3 extracts a technology classification that the technology classification preparation unit 102 provides to the large-scale language model 1. However, if there were no restrictions on the number of characters or tokens in a prompt input to the large-scale language model 1, or if the restrictions were significantly relaxed, it would be possible to include all information in the International Patent Classification in the input task and have the large-scale language model 1 generate ideas.

[0142] FIG. 11 is a block diagram showing another example of the functional configuration of the information processing device 10.

[0143] 11, the information processing device 10 has, as its functional configuration, an input unit 101, an idea generation unit 104, and an output unit 105. Each functional configuration is realized by the CPU 11 reading and executing an information processing program stored in the ROM 12 or the storage 14. The functional configuration of the information processing device 10 shown in FIG. 11 is a configuration in which the technology classification preparation unit 102 and the technology classification list 103 are removed from the functional configuration of the information processing device 10 shown in FIG. 3.

[0144] 11 is configured in a case where there is no limit or the limit is significantly relaxed on the number of characters or tokens of a prompt to be input to the large-scale language model 1. In this case, the input unit 101 receives an input task for making the large-scale language model 1 propose an idea. Then, the idea generation unit 104 generates a prompt including information on all technology categories and inputs it to the large-scale language model 1 in order to make the large-scale language model 1 generate an idea.

[0145] For example, when the input unit 101 receives an input task such as, "With organic EL lighting, only 20% of the emitted light can be extracted to the outside. One possible solution is to form a highly refractive, light-scattering layer on a glass substrate and then create an organic EL element on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, because an uneven surface on an organic EL device can easily cause a short circuit between the two electrodes. Please use all of your knowledge of glass-related technology to propose a scattering layer that meets these conditions," the idea generation unit 104 generates a prompt that specifies the input task as an assignment and all international patent classifications as the technology classifications to be used. In this prompt, the idea generation unit 104 causes the large-scale language model 1 to select an international patent classification to be used for idea generation from among all international patent classifications.

[0146] In this way, by using a large-scale language model 1 in which there is no limit on the number of characters or tokens in the prompt to be input, or in which the limit is significantly relaxed, the information processing device 10 can generate a prompt containing information on all technology classifications and input it into the large-scale language model 1 in order to have the large-scale language model 1 generate ideas.

[0147] So far, we have described the information processing device 10 that causes the large-scale language model 1 to propose ideas for solving technical problems, but the information processing device 10 according to this embodiment may input, as an input task, a task that causes the large-scale language model 1 to propose new ideas. Even in this case, by using a technology classification list 103 such as the International Patent Classification or FI number, the information processing device 10 according to this embodiment can cause the large-scale language model 1 to propose ideas that are rich in technical suggestions and have deep insights compared to when a technology classification list is not used.

[0148] Using the information processing device 10 according to this embodiment, a task was considered in which a new technical idea is output by combining a technology related to chemically strengthened glass and a technology related to fluororesin. In this embodiment, an example is shown in which a new idea is output to the large-scale language model 1 by the following five methods: (1) Simply providing a prompt in the form of a sentence to the large-scale language model 1 and outputting the idea to the large-scale language model 1; (2) Selecting an International Patent Classification number corresponding to a technology related to chemically strengthened glass and fluororesin, and outputting the idea to the large-scale language model 1 using this combination; (3) Providing a prompt in the form of a sentence and selecting an International Patent Classification number corresponding to a technology related to chemically strengthened glass and fluororesin, and outputting the idea to the large-scale language model 1 using this combination; (4) Selecting an FI number corresponding to a technology related to chemically strengthened glass and fluororesin, and outputting the idea to the large-scale language model 1 using this combination; (5) Providing a prompt in the form of a sentence and selecting an FI number corresponding to a technology related to chemically strengthened glass and fluororesin, and outputting the idea to the large-scale language model 1 using this combination.

[0149] In the case of (1), the input task acquired by the input unit 101 is, for example, as follows: "Please give us five novel ideas that have not yet been seen in the world using chemically strengthened glass and fluororesin technology." In this case, the ideas output by the large-scale language model 1 are as follows: "1. Super water-repellent smart windows: Using chemically strengthened glass coated with fluororesin, we have developed smart windows for automobiles that maintain clear visibility even when it rains. 2. Self-healing smartphone screens: We have developed smartphone screens that combine fluororesin and chemically strengthened glass, which have the ability to self-repair even when they get small scratches. 3. Fire- and heat-resistant transparent walls: By alternating layers of chemically strengthened glass and fluororesin, we have developed transparent wall materials that are highly fire- and heat-resistant, bringing innovation to the architecture and aerospace industries. 4. Stain-repellent kitchenware: By utilizing the non-stick properties of fluororesin and the durability of chemically strengthened glass, we have developed a series of kitchenware that is stain-resistant and easy to clean. 5. Energy-efficient solar panels: By combining an anti-reflective coating using fluororesin with chemically strengthened glass, we have developed more efficient solar panels." As you can see, the ideas output by large-scale language model 1 were largely mundane.

[0150] In the case of (2), the input task acquired by the input unit 101 is, for example, as follows: "You are an expert in the IPC codes shown below. Your areas of expertise are divided into Group A and Group B. Group A is C03C 21 / 00, and Group B is B32B 27 / 30, B32B 27 / 08, B32B 27 / 32, B32B 1 / 08, B32B 15 / 08. Please provide five new ideas that combine the technologies of these two groups and are not yet available in the world. Please attach a list of related IPC codes." In this case, the ideas output by the large-scale language model 1 were as follows: 1. Self-healing glass composite: Special glass that is scratch-resistant and self-healing has been developed using the self-healing polymer of B32B 27 / 30. 2. Radiation protection glass: Radiation-protective window glass has been developed by combining glass improved with C03C 21 / 00 technology with the radiation shielding layer of B32B 27 / 08. 3. High-durability, high-transparency building materials: High-transparency glass improved with C03C 21 / 00 and the strengthening bonding technology of B32B 27 / 32 have been used to develop building materials that are both durable and beautiful. 4. Multi-functional smart windows: Smart windows incorporating environmental sensors have been developed using the technologies of B32B 1 / 08 and B32B 27 / 30 with glass that has the properties of C03C 21 / 00. 5. Fire-resistant composite glass panels: Glass with improved fire resistance using C03C 21 / 00 technology and B32B "A highly safe architectural panel made by combining 15 / 08 metal layers." In this case, large-scale language model 1 output many more useful ideas than in case (1), but it did not output any ideas containing keywords such as chemical strengthening or fluororesin.

[0151] In the case of (3), the input task acquired by the input unit 101 is, for example, as follows: "Please give five new ideas that have not yet been seen in the world by combining the technologies of chemically strengthened glass and fluororesin. When doing so, please refer to the technologies that correspond to the related IPC codes shown below. You are an expert on the IPC codes shown below. The related IPC codes are C03C 21 / 00, B32B 27 / 30, B32B 27 / 08, B32B 27 / 32, B32B 1 / 08, and B32B 15 / 08. The list of related IPC codes is attached." In this case, the ideas output by the large-scale language model 1 were as follows: "1. Flexible displays: By utilizing the flexibility of chemically strengthened glass and fluororesin, we have developed bendable displays for smartphones and tablets. 2. Impact- and chemical-resistant windows: By combining chemically strengthened glass that has been ion-diffused with a fluororesin coating, we have developed windows that are highly impact-resistant and chemical-resistant. 3. Self-repairing smartphone screens: By combining fluororesin and chemically strengthened glass, we have produced smartphone screens that can self-repair small scratches. 4. Super water-repellent glass panels: By using chemically strengthened glass coated with fluororesin, we have created architectural glass panels that repel water and dirt. 5. Heat- and UV-resistant car glass: We have developed automotive glass that is both UV-blocking and heat-resistant, and can withstand prolonged exposure to direct sunlight." In this case, large-scale language model 1 output ideas that included keywords such as chemically strengthened and fluororesin, compared to the case of (2).

[0152] In the case of (4), the input task acquired by the input unit 101 is, for example, as follows: "You are an expert in the technologies corresponding to the FI codes (Japanese Patent Classification) C03C 21 / 00 101, F28F 19 / 04 A, A41D 31 / 00 504 G, C25D 11 / 18 306 A, and B32B 27 / 30 D. Please give us 10 new ideas that have not yet been developed by combining C03C 21 / 00 101 with technologies corresponding to other FI codes (F28F 19 / 04 A, A41D 31 / 00 504 G, C25D 11 / 18 306 A, and B32B 27 / 30 D)." In this case, the ideas output by the large-scale language model 1 were as follows: "1. Chemically strengthened glass is coated with a fluororesin to improve its heat and chemical resistance. Particularly applicable to industrial machinery and laboratory equipment. 2. A thin film of fluororesin is applied to a heat exchanger made from chemically strengthened glass to improve corrosion resistance and thermal efficiency. 3. Chemically strengthened glass fiber and fluororesin are used to develop protective clothing that is both abrasion-resistant and chemical-resistant. 4. Chemically strengthened glass powder is combined with fluororesin and applied to metal surfaces to create a decorative and durable coating. 5. A new type of composite material with high physical strength is developed by alternating layers of chemically strengthened glass and fluororesin." In this case, large-scale language model 1 output ideas that were most notable for technological innovation, including the development of new types of materials and manufacturing methods.

[0153] In the case of (5), the input task acquired by the input unit 101 is, for example, as follows: "Please give us 10 new ideas that have not yet been seen in the world by combining the technologies of chemically strengthened glass and fluororesin. Please refer to the following FI codes (Japanese Patent Classification) when doing so: C03C 21 / 00 101, F28F 19 / 04 A, A41D 31 / 00 504 G, C25D 11 / 18 306 A, B32B 27 / 30 D. A list of related FI codes is attached. Please ignore spaces and backslashes. Please explain the technology of the FI code, taking into account subclasses. FI is based on IPC, but a hierarchical structure is created by adding numbers or letters to the IPC code. It is sometimes written as fluororesin or fluoropolymer." In this case, the ideas output by the large-scale language model 1 were as follows. "1. Coating chemically strengthened glass with a fluororesin improves its heat and chemical resistance, particularly for industrial machinery and laboratory equipment. 2. Applying a thin film of fluororesin to a heat exchanger made from chemically strengthened glass to improve corrosion resistance and thermal efficiency. 3. Developing protective clothing that is both abrasion-resistant and chemical-resistant using chemically strengthened glass fiber and fluororesin. 4. Applying chemically strengthened glass powder in combination with fluororesin to metal surfaces to create a decorative and durable coating. 5. Developing a new type of composite material with high physical strength by alternating layers of chemically strengthened glass and fluororesin." In this case, large-scale language model 1 output ideas that focused on durability and developing materials to meet specific industrial requirements.

[0154] Of the outputs for the input tasks (1) to (5), (4) was the most innovative and unique. On the other hand, the output for (2) did not include keywords such as chemical strengthening or fluororesin. This is because the International Patent Classification (IPC) does not have technical classifications corresponding to chemical strengthening or fluororesin, while the FI is a more detailed classification based on the IPC, including a classification corresponding to chemical strengthening (C03C 21 / 00) and a classification corresponding to fluororesin (A41D 31 / 00 504). Thus, it is desirable for technical classifications to include keywords. In the above example, FI included technical classifications containing keywords. However, technical classifications may be added to the existing technical classification list to include keywords. In this embodiment, if the specific technical field is defined as a technology related to chemically strengthened glass, then a technology related to fluororesin can be considered a technical field other than the specific technical field.

[0155] When the idea generation unit 104 causes the large-scale language model 1 to propose new ideas, the output unit 105 may arrange the ideas proposed by the large-scale language model 1 in a technology classification matrix and display the arranged ideas on the display unit 16. Fig. 12 is a diagram showing an example in which the ideas proposed by the large-scale language model 1 are arranged in an International Patent Classification matrix. By the output unit 105 displaying a matrix such as that shown in Fig. 12 on the display unit 16, the user of the information processing device 10 can understand which ideas correspond to which combinations of technology classifications.

[0156] When providing an input task to the large-scale language model 1, the idea generation unit 104 may assign priorities to the order of access to technology classifications and then provide the input task to the large-scale language model 1. When outputting ideas for composite materials to the large-scale language model 1, the number of material combinations can become enormous depending on the number of materials to be combined. By assigning priorities to the order of access to technology classifications, the idea generation unit 104 can output ideas to the large-scale language model 1 more effectively than when no priorities are assigned.

[0157] For example, when selecting international patent classification numbers corresponding to technologies related to chemically strengthened glass and fluororesin and providing them to the large-scale language model 1, the idea generation unit 104 sets priorities for the international patent classification for chemically strengthened glass and the international patent classification related to fluororesin and provides them to the large-scale language model 1. Then, by setting an upper limit on the number of ideas output from the large-scale language model 1, the idea generation unit 104 can cause the large-scale language model 1 to preferentially output ideas related to international patent classifications with higher priorities.

[0158] According to the information processing device 10 of this embodiment, by having such a configuration, it is possible to effectively output technical ideas to the large-scale language model 1 by providing the large-scale language model 1 with a technology classification list, which is information on technology classifications that have been listed in advance.

[0159] Next, the operation of the information processing device 10 will be described.

[0160] 13 is a flowchart showing the flow of information processing by the information processing device 10. The CPU 11 reads out an information processing program from the ROM 12 or the storage 14, loads it into the RAM 13, and executes it, thereby performing information processing.

[0161] In step S101, the CPU 11 acquires an input task for outputting a technical idea to the large-scale language model 1. The input task is input by the user of the information processing device 10 using the input unit 15, and may contain, for example, the following content: "With organic EL lighting, only 20% of the emitted light can be extracted to the outside. One possible solution is to form a highly refractive light-scattering layer on a glass substrate and then create an organic EL element on top of that. The refractive index must be 1.9 or higher. The surface must also be very smooth, because organic EL devices easily short-circuit between the two electrodes if the underlying surface is uneven. Please propose a scattering layer that meets these conditions using technology related to the IPC numbers shown in the attached C03B and C03C. When doing so, please indicate which numbers you used and how to create the scattering layer."

[0162] Following step S101, in step S102, the CPU 11 extracts, from the technology classification list 103, a technology classification to be given to the large-scale language model 1, based on the input task acquired in step S101. When the above input task is acquired, the CPU 11 extracts C03B and C03C of the International Patent Classification from the technology classification list 103 as the technology classification.

[0163] Following step S102, in step S103, the CPU 11 generates a prompt using the input task acquired in step S101 and the technology classification extracted in step S102, and provides the generated prompt to the large-scale language model 1. As described above, if the input task originally included an International Patent Classification, the CPU 11 provides the content of the input task as is as a prompt to the large-scale language model 1. If the input task did not include an International Patent Classification, the CPU 11 generates a prompt that proposes an idea using the technology related to the International Patent Classification extracted in step S102 in the content of the input task, and provides the generated prompt to the large-scale language model 1.

[0164] Following step S103, in step S104, the CPU 11 obtains an output from the large-scale language model 1. A technical idea is output from the large-scale language model 1 in response to the prompt given to the large-scale language model 1. That is, in step S104, the CPU 11 obtains the technical idea output by the large-scale language model 1.

[0165] Following step S104, in step S105, the CPU 11 displays on the display unit 16 or records in the storage 14 the technical idea output by the large-scale language model 1 acquired in step S104.

[0166] According to the information processing device 10 of this embodiment, by executing such processing and providing the large-scale language model 1 with a technology classification list, which is information on technology classifications that have been listed in advance, it is possible to make the large-scale language model 1 output technical ideas more effectively than in the case where the technology classification list is not provided.

[0167] It should be noted that the ideas output from the large-scale language model 1 are not necessarily correct. This is known as a phenomenon called hallucination, and the same applies to the present disclosure.

[0168] Although the embodiments of the present disclosure have been described in detail above with reference to the accompanying drawings, the technical scope of the present disclosure is not limited to such examples. It is clear that a person skilled in the art of the present disclosure can conceive of various modifications or alterations within the scope of the technical idea described in the claims, and it is understood that these modifications or alterations also naturally fall within the technical scope of the present disclosure.

[0169] Furthermore, the effects described in the above embodiments are explanatory or exemplary and are not limited to those described in the above embodiments. In other words, the technology according to the present disclosure may achieve other effects that are obvious to a person skilled in the art of the present disclosure from the description in the above embodiments, in addition to or instead of the effects described in the above embodiments.

[0170] In each of the above embodiments, the information processing performed by the CPU after reading the software (program) may be performed by various processors other than the CPU. Examples of such processors include a PLD (Programmable Logic Device) whose circuit configuration can be changed after manufacture, such as an FPGA (Field-Programmable Gate Array), and a dedicated electrical circuit, such as an ASIC (Application Specific Integrated Circuit), which is a processor having a circuit configuration designed specifically to perform specific processing. Furthermore, the information processing may be performed by one of these various processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). Furthermore, the hardware structure of these various processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor elements.

[0171] The present disclosure can also be applied to programs and program products. In the above embodiments, the information processing program is pre-stored (installed) in a ROM or storage device, but this is not limiting. The program may be provided in a form recorded on a non-transitory recording medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The program may also be downloaded from an external device via a network.

[0172] The disclosure of Japanese Patent Application No. 2024-114323, filed on July 17, 2024, is incorporated herein by reference in its entirety. In addition, all documents, patent applications, and technical standards described herein are incorporated herein by reference to the same extent as if each individual document, patent application, and technical standard were specifically and individually indicated to be incorporated by reference.

Claims

1. An information processing method in which a processor, in response to an input task of having a large-scale language model propose ideas in a specific technical field, provides the large-scale language model with at least one technical field from a technology classification list, which is information on listed technology classifications, and causes the large-scale language model to generate ideas related to the specific technical field.

2. The information processing method according to claim 1, wherein the processor provides the large-scale language model with at least one technical field different from the specific technical field.

3. The information processing method of claim 1, wherein the processor extracts a technology classification to be given to the large-scale language model from the technology classification list based on the input task, gives the input task to the large-scale language model so that the extracted technology classification is included, and causes the large-scale language model to generate ideas related to the specific technology field.

4. The information processing method according to claim 3, wherein the processor provides the input task to a large-scale language model that extracts technology classifications, and extracts technology classifications based on a response from the large-scale language model.

5. The information processing method according to claim 4, wherein the processor provides the input task to a large-scale language model specialized in extracting technology classifications, and extracts technology classifications based on a response from the large-scale language model.

6. The information processing method according to claim 1, wherein the processor causes the large-scale language model to output an evaluation of the idea obtained from the large-scale language model.

7. The information processing method of claim 1, wherein the processor uses the input task and the technology classification to instruct the large-scale language model to generate ideas specific to a particular technology classification.

8. The information processing method according to claim 1, wherein the large-scale language model is a model that has been trained in advance on at least technical information related to the technology classification list.

9. The information processing method of claim 8, wherein the processor incorporates technical information from a Retrieval Augmented Generation (RAG) that accumulates technical information related to at least the technical classification list and provides the technical information to the large-scale language model.

10. The information processing method according to claim 8, wherein the large-scale language model is a model that has been trained in advance by fine tuning at least technical information related to the technical classification list.

11. The information processing method according to claim 8, wherein the large-scale language model is a short-shot trained model or a model trained by prompting input of technical information related to the technical classification list.

12. The information processing method according to claim 1, wherein the large-scale language model is a model that can refer to the technology classification list.

13. The information processing method according to claim 1, wherein the technology classification list is the International Patent Classification.

14. An information processing method as described in claim 1, wherein the technology classification list is a list in which existing technology classifications are subdivided using a uniquely prepared technology classification, or a list in which information is added to existing technology classifications.

15. The information processing method according to claim 1, wherein the processor, when providing the input task to the large-scale language model, provides the large-scale language model with a priority given to the order of access to the technical classification.

16. The information processing method according to claim 1, wherein the input task is a task of having the large-scale language model propose ideas for solving a technical problem.

17. The information processing method according to claim 1, wherein the input task is a task of having the large-scale language model propose a new idea.

18. The information processing method according to claim 1, wherein the number of parameters of the large-scale language model for which the processor proposes ideas is 500 billion or more.

19. An information processing device comprising: an input unit that acquires an input task for causing a large-scale language model to propose ideas in a specific technical field; and an idea generation unit that provides the large-scale language model with at least one technical field from a technology classification list, which is information on listed technology classifications, for the input task, and causes the large-scale language model to generate ideas related to the specific technical field.

20. An information processing program that causes a computer to execute a process in which, in response to an input task of having a large-scale language model propose ideas in a specific technology field, at least one technology field is provided to the large-scale language model from a technology classification list, which is information on listed technology classifications, and the large-scale language model generates ideas related to the specific technology field.

Images