Solution generation device and solution generation method

Abstract

This solution generation device finds a solution to a problem by using a language model. The solution generation device comprises: an input unit (3) that inputs, from a problem database (1) and a technical element database (2), a plurality of combinations of a problem and a technical element for solving the problem; a generation unit (4) that generates instruction information for a language model for each combination of a problem and a technical element input by the input unit, inputs the instruction information into the language model, sets, for the language model, diversity parameters (8) for specifying the diversity of solutions to be found by the language model, and instructs the language model to generate solution text; and an output unit (5) that outputs solutions found by the language model when using, as items, the problems and technical elements input by the input unit.

Description

Solution generation device and solution generation method

[0001] The present invention relates to a solution generation device and a solution generation method for generating solutions.

[0002] There is a technique for generating candidate answer sentences that serve as answers to new questions submitted by users, based on a question and answer database.

[0003] For example, Patent Document 1 describes a system that includes a processor and a storage device, and the storage device stores a question and answer database that stores past question sentences in association with past answer sentences to the past question sentences. When a user inputs a group of item words selected by a user from a group of item candidate words generated based on the past question sentences and new question sentences stored in the question and answer database and a new question sentence, the processor generates question information including the group of item words based on the group of item words and the new question sentence, calculates a similarity between each of the past question sentences stored in the question and answer database and the question information, extracts past question sentences similar to the question information from the question and answer database based on the similarity, and extracts past answer sentences associated with the past question sentences similar to the extracted question information from the question and answer database to set them as first candidate answer sentences.

[0004] Japanese Patent Application Laid-Open No. 2023-091791

[0005] In use cases (such as generating solution ideas) in which multiple questions about new issues or new technological elements are input, such as when generating solution ideas that respond to changes in technological elements in response to changes in social trends, it is difficult to apply the above-mentioned prior art that extracts similar questions from an answer database.

[0006] An object of the present invention is to provide a solution generation device and a solution generation method that can generate documents describing solutions even for new problems and new technical elements.

[0007] In order to solve the above-mentioned problems, the solution generation device of the present invention is a solution generation device that finds solutions to problems using a language model, and comprises: an input unit that inputs a problem and a plurality of combinations of technical elements that solve the problem from a problem database and a technical element database; a generation unit that generates instruction information for a language model for each combination of problem and technical element input by the input unit, inputs the instruction information into the language model, sets a diversity parameter instructing the diversity of solutions found by the language model in the language model, and instructs the language model to generate sentences of solutions; and an output unit that outputs the solutions found by the language model using the problem and technical elements input by the input unit as items.

[0008] According to the present invention, it is possible to generate text describing solutions even for new problems and new technical elements.

[0009] 1 is a functional block diagram showing the overall configuration of a solution generation device of an embodiment;

[0023] FIG. 1 is a diagram explaining an example of a specific hardware configuration of a solution generation device;

[0024] FIG. 2 is a diagram explaining the data structure of problems stored in a problem database;

[0025] FIG. 3 is a diagram explaining the data structure of technical elements stored in a technical element database;

[0026] FIG. 4 is a diagram explaining the problem selection process in the problem / technology selection unit;

[0027] FIG. 5 is a diagram explaining the technical element selection process in the problem / technology selection unit;

[0028] FIG. 6 is a diagram explaining additional viewpoint information;

[0029] FIG. 7 is a diagram explaining the diversity parameter setting process in the setting unit;

[0030] FIG. 8 is a diagram showing an example of the setting values ​​of the diversity parameters Temperature and Top_p;

[0031] FIG. 9 is a flowchart explaining the solution document generation process executed in the generation unit;

[0032] FIG. 10 is a diagram showing an example of the display of a solution notified to the solution display unit;

[0033] FIG. 11 is a diagram showing an example of the display of an additional viewpoint processing area for inputting additional viewpoints;

[0034] FIG. 11 is a flowchart explaining the solution document generation process executed in the generation unit for adding additional viewpoints and digging deeper.

[0010] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[0011] 1 is a functional block diagram showing the overall configuration of a solution generation system according to an embodiment of the present invention. The solution generation system is composed of a problem database 1 (problem DB), a technology element database 2 (technology element DB), an input unit 3, a generation unit 4, an output unit 5, a solution display unit 6, a setting unit 7 for setting diversity parameters 8, and a large-scale language model unit 9.

[0012] The issue database 1 stores a list of issues to be solved, as will be described in detail later. Examples of issues include long-term social issues such as global warming, management issues of the organization to which the user belongs, development issues for the services and products provided by the organization, and mid-term management plans for companies. It also includes megatrends, which are changes that have a major impact on people's lives, such as the economic impact of climate change and resource shortages.

[0013] The technical element database 2 (technical element DB) stores a list of technical elements that contribute to solving problems. Here, technical elements refer to technologies owned by users involved in the solution study, new technologies that are expected to be put to practical use in the near future, etc.

[0014] The input unit 3 is composed of a problem / technology selection unit 31, a solution selection unit 32, and an additional viewpoint input unit 33. Then, problem / technology elements that will be the basis for generating prompts for the generation AI that constitutes the large-scale language model unit 9 (described later) are selected, and these are used as input information for the generation unit 4 together with additional viewpoints for in-depth consideration.

[0015] The problem / technology selection unit 31 selects, based on user instructions, problems and technical elements to be considered in the study of solutions from the problems and technical elements stored in the problem database 1 and the technical element database 2 .

[0016] Based on a user instruction, the solution selection unit 32 selects a solution to be explored from the solutions output by the output unit 5 to the solution display unit 6. Based on a user instruction, the additional viewpoint input unit 33 inputs additional viewpoints to be considered in the solution to be explored selected by the solution selection unit 32. The solution selection unit 32 and the additional viewpoint input unit 33 can be collectively called a deep exploration processing unit.

[0017] The generation unit 4 generates prompts (answer instruction information) that are input information to be given to the large-scale language model unit 9 based on the information input by the input unit 3. At this time, for the multiple tasks and technical elements selected by the task / technology selection unit 31, one-to-one combinations of tasks and technical elements are found, and a prompt is generated for each combination. Note that a "prompt" is used as an example of "instruction information."

[0018] Furthermore, the generation unit 4 generates a prompt from the additional perspective input by the additional perspective input unit 33 and the problem and technology corresponding to the solution for which the additional perspective has been input. The generation unit 4 then inputs the generated prompt to the large-scale language model unit 9, instructs it to generate a sentence describing the solution based on the diversity parameter 8, and outputs the obtained solution to the output unit 5.

[0019] The output unit 5 notifies the solution display unit 6 of the solution text obtained from the generation unit 4 in order to present it to the user. At this time, as will be described in detail later, the output unit 5 outputs solutions in a table format for 1:1 combinations of problems and technical elements, with the multiple problems and technical elements selected by the problem / technology selection unit 31 as items.

[0020] The solution display unit 6 displays solutions for each combination of problem and technology in a table format, and also displays the displayed solutions so that the user can select them. The solution selection unit 32 selects the solution selected by the user. The solution display unit 6 also displays a text box window for inputting an additional viewpoint using the additional viewpoint input unit 33, superimposed on the solutions displayed in the table format.

[0021] The setting unit 7 has a diversity setting selection unit 71 and sets a diversity parameter 8, which is setting information for the large-scale language model unit 9. The diversity parameter 8 is a parameter indicating the diversity of the output of the large-scale language model unit 9. High diversity in the output of the large-scale language model unit 9 means that the large-scale language model unit 9 is in a state where it outputs noise and errors, but that it is possible to expect the seeds of new ideas and unexpected output. For example, the diversity parameter 8 is represented by Temperature and Top_p in GPT (registered trademark, Generative Pre-trained Transformer), which is a type of large-scale language model (LLM). GPT is a registered trademark of OpenAI Opco LLC.

[0022] Although details will be described later, the diversity setting selection unit 71 of the setting unit 7 allows the user to select from a plurality of mode settings, namely, "divergence mode," "discussion mode," and "convergence mode," for the diversity (creativity) of the large-scale language model unit 9. Then, the setting unit 7 sets Temperature and Top_p according to the mode setting selected by the user, and notifies the generation unit 4 of these as diversity parameters 8.

[0023] The mode settings are not limited to the "divergence mode," "discussion mode," and "convergence mode." The diversity setting selection unit 71 selects at least two of the "divergence mode," "discussion mode," and "convergence mode."

[0024] The large-scale language model unit 9 provides the generation unit 4 with a natural-text response as a solution in response to the input of natural-text generated by the generation unit 4 from a combination of a problem and a technology or a combination of a problem, a technology, and an additional perspective. At this time, the large-scale language model unit 9 generates a sentence describing a solution with diversity corresponding to the diversity parameter 8 input from the generation unit 4.

[0025] As described above, the solution generation device of the embodiment is a solution generation device that finds solutions to problems using a large-scale language model, and includes: an input unit 3 that inputs problems and multiple combinations of technical elements that solve the problems from a problem database 1 and a technical element database 2; a generation unit 4 that generates prompts for each combination of problem and technical element input by the input unit 3, inputs the prompts to a language model, sets a diversity parameter 8 in the language model that indicates the diversity of solutions found by the language model, and instructs the language model to generate sentences of solutions; and an output unit 5 that outputs the solutions found by the language model using the problem and technical elements input by the input unit 3 as items. The language model is a language model generated by deep learning. Examples of language models include a large-scale language model, a small-scale language model, and a multimodal language model.

[0026] Next, an example of a specific hardware configuration of the solution generation device will be described with reference to Fig. 2. Specifically, the solution generation device of the embodiment is realized by a computer 200 shown in the hardware configuration diagram of Fig. 2. The computer 200 has a CPU (Central Processing Unit) 201, a ROM (Read Only Memory) 202, a RAM 203, an HDD (Hard Disk Drive) 204, an input / output I / F (Interface) 205, a communication I / F 206, and a media I / F 207. The HDD 204 may be an SSD (Solid State Drive).

[0027] The CPU 201 operates based on programs stored in the ROM 202 or the HDD 204, and controls each unit of the computer 200. The ROM 202 stores a boot program executed by the CPU 201 when the computer 200 is started up, programs related to the hardware of the computer 200, and the like.

[0028] The CPU 201 controls an input device 210 such as a mouse or keyboard, and an output device 211 such as a display or printer, via an input / output I / F 205. The CPU 201 acquires data from the input device 210 via the input / output I / F 205, and outputs generated data to the output device 211. Note that a GPU (Graphics Processing Unit) or the like may be used as a processor together with the CPU 201.

[0029] The HDD 204 stores programs executed by the CPU 201 and data used by the programs. The communication I / F 206 receives data from other devices via a communication network (e.g., NW (Network) 220) and outputs the data to the CPU 201, and also transmits data generated by the CPU 201 to other devices via the communication network.

[0030] The media I / F 207 reads a program or data stored in the recording medium 212 and outputs it to the CPU 201 via the RAM 203. The CPU 201 loads a program related to a target process from the recording medium 212 onto the RAM 203 via the media I / F 207, and executes the loaded program. The recording medium 212 is an optical recording medium such as a DVD (Digital Versatile Disc) or a PD (Phase Change Rewritable Disc), a magneto-optical recording medium such as an MO (Magneto Optical Disc), a magnetic recording medium, a semiconductor memory, or the like.

[0031] For example, when the computer 200 or the like functions as the solution generation device of the embodiment, the CPU 201 of the computer 200 executes a program loaded on the RAM 203 to realize the functions of each processing unit of the solution generation device.

[0032] The CPU 201 reads and executes a program related to a target process from the recording medium 212. Alternatively, the CPU 201 may read a program related to a target process from another device via a communication network (NW 220).

[0033] In detail, the problem database 1 and the technical element database 2 are stored in the HDD 204. The solution display unit 6 is realized by the output device 211. The large-scale language model unit 9, the generation unit 4, the output unit 5, and the setting unit 7 realize their functions by the CPU 201 executing a program. Furthermore, the input unit 3 realizes its function by controlling the input device 210 and the output device 211 by the CPU 201 executing a program.

[0034] The solution generation device may not only realize its functions by the computer 200 (information processing device) shown in FIG. 2, but also provide the functions of the solution generation device as a cloud computing service.

[0035] In more detail, a problem database 1 and a technical element database 2 are set up on the cloud, and a cloud virtualization server realizes the functions of a large-scale language model unit 9, a generation unit 4, an output unit 5, a setting unit 7, and an input unit 3. The output unit 5 on the cloud displays solutions on the display unit of a user terminal connected to the cloud as a solution display unit 6, and the input unit 3 and setting unit 7 on the cloud control the user terminal to make a selection.

[0036] The configuration and operation of the solution generator will be explained in more detail below.

[0037] 3 is a diagram showing the data structure of the problems 11 to be solved recorded in the problem database 1 (FIG. 1). The problems 11 are composed of the problem ID, the name of the problem to be solved, and a problem summary (description of the problem content and trends) for each problem to be solved.

[0038] Problem 11 is registered in advance in problem database 1 before a solution to the problem is found by large-scale language model unit 9. At this time, the outline of the problem may be found by inputting the name of the problem to be solved into large-scale language model unit 9. This makes it easy to register problem 11.

[0039] 4 is a diagram showing the data structure of the technical elements 21 for solving problems recorded in the technical element database 2 (FIG. 1). The technical elements 21 are composed of the ID, name, and summary of the technical element (description of the technical content and progress) for each technical element.

[0040] The technical elements 21 are registered in advance in the technical element database 2 before a solution to the problem is found by the large-scale language model unit 9. At this time, the outline of the technical element may be found by inputting the name of the technical element into the large-scale language model unit 9. This makes it easy to register the technical elements 21.

[0041] Fig. 5 is a diagram explaining the problem selection process in the problem / technology selection unit 31 (Fig. 1) of the solution generation device, and shows the problem selection information 311. The problem / technology selection unit 31 reads out the problems 11 explained in Fig. 3 from the problem database 1, and for each problem to be solved registered in the problem 11, provides a selection item for setting whether or not to select the problem as a problem for which a solution is sought by the large-scale language model unit 9, and generates the problem selection information 311.

[0042] 5 and selects the problem for which the user has marked an "X" in the selection field as the problem to be solved. The generation unit 4 generates a prompt for the large-scale language model unit 9 based on the selected problem. In detail, the problem / technology selection unit 31 regards the problem to be solved corresponding to the position tapped by the user on the screen of the displayed problem selection information 311 as the selected problem, and displays an "X" in the selection field to indicate that it has been selected as the problem to be solved.

[0043] Fig. 6 is a diagram for explaining the selection process of technological elements in the problem / technology selection unit 31 (Fig. 1) of the solution generation system, and shows the selection information of technological elements 312. The problem / technology selection unit 31 reads out the technological elements 21 explained in Fig. 4 from the technological element database 2, and provides a selection item for each technological element registered in the technological elements 21 to set whether or not to select it as a technological element when the large-scale language model unit 9 finds a solution, and generates the selection information of technological elements 312.

[0044] 6 and selects the technological element for which the user has marked an "X" in the selection field as the technological element that solves the problem. The generation unit 4 generates a prompt for the large-scale language model unit 9 based on the selected technological element. In detail, the problem / technology selection unit 31 selects the technological element corresponding to the position tapped by the user on the screen of the displayed technological element selection information 312 as the selected technological element, and displays an "X" in the selection field to indicate that the technological element has been selected as the technological element that solves the problem.

[0045] 7 shows the additional viewpoint information 331 of the additional viewpoint input unit 33 (FIG. 1) of the solution generation device. As will be described in detail later, the additional viewpoint input unit 33 records, as an additional viewpoint, an explanatory text entered by the user in a text box displayed superimposed on the solution in the solution display unit 6 in the additional viewpoint information 331. The generation unit 4 generates a prompt for the large-scale language model unit 9 based on the input additional information.

[0046] Fig. 8 is a diagram showing the process of setting the diversity parameters 8 in the setting unit 7 of the solution generation device. The diversity parameters should include at least a first parameter indicating the distribution of word appearance probabilities and a second parameter indicating the threshold of the top cumulative probabilities when selecting words to appear. In the following, we will explain the case where the diversity parameters 8 are represented by Temperature and Top_p in the GPT. In the following, we will explain the case where the first parameter is Temperature, the second parameter is Top_p, and words are tokens.

[0047] As shown in FIG. 8, the diversity setting selection unit 71 displays mode setting buttons for "divergence mode," "discussion mode," and "convergence mode," which are diversity selection buttons with preset values ​​for the diversity parameters Temperature and Top_p, as well as a mode setting button for "custom mode," which allows the user to arbitrarily set diversity parameters. It also displays the set values ​​for the diversity parameters Temperature and Top_p. In FIG. 8, the "divergence mode" has been selected by the user, and the set values ​​of the diversity parameters at that time, "Temperature: 1.0" and "Top_p: 1.0," are displayed.

[0048] FIG. 9 is a diagram showing an example of the set values ​​of Temperature and Top_p of the diversity parameter 8 in the mode settings of "divergence mode," "discussion mode," and "convergence mode."

[0049] Temperature is a parameter used to calculate the probability distribution when selecting a token in sentence generation in the large-scale language model unit 9 (GPT). When Temperature is set high, the probability of occurrence of a token with a low probability of being selected as the next token is calculated to be high, and when Temperature is set low, the probability of occurrence of a token with a high probability of being selected as the next token is calculated to be higher. Therefore, setting Temperature high has the advantage of increasing the diversity of the sentences generated. On the other hand, setting Temperature low has the advantage of increasing the consistency of the sentences generated, and setting Temperature to 0 results in the token with the highest probability of being selected as the next token always being selected.

[0050] Top_p is the threshold value of the top cumulative probability when selecting tokens from the token occurrence probability distribution during sentence generation in the large-scale language model unit 9 (GPT). For example, if Top_p is set to 0.9, only tokens that make up the top 90% probability will be considered as selection candidates. Therefore, setting Top_p higher will generate a solution sentence from a wider and more diverse set of token candidates. Setting Top_p lower will generate a solution sentence from a more consistent and more confident set of token candidates.

[0051] Next, the solution document generation process executed by the generator 4 will be explained with reference to the flowchart in Fig. 10. The flowchart in Fig. 10 shows the process of finding solutions for each combination of the problem to be solved selected in Fig. 4 and the technical elements selected in Fig. 5.

[0052] In step S101, the generation unit 4 reads the diversity parameters 8 set by the setting unit 7. In step S102, the generation unit 4 reads the first problem in the list of problems to be solved, which is selected by the problem / technology selection unit 31. In step S103, the generation unit 4 reads the first technical element in the list of technical elements that solve the problem, which is selected by the problem / technology selection unit 31.

[0053] In step S104, the generation unit 4 generates a prompt to be input to the large-scale language model unit 9 from the read problem and technical elements, and notifies the large-scale language model unit 9 of the generated prompt and the read diversity parameters 8, instructing it to create a sentence describing a solution.The generation unit 4 then obtains the sentence describing the solution created by the large-scale language model unit 9 and notifies the output unit 5.

[0054] For example, the generator 4 generates the following prompt (instruction sentence): You are an expert in forced idea generation in design thinking workshops. Under the following conditions, generate a solution using "technical elements that contribute to the solution" for the specified "problem to be solved."

[0055] Issue to be solved: Aging society Overview of the issue: The population of people aged 65 and over accounted for 7.1% of Japan's total population in 1970, but will reach 37.7% in 2050. Technological element that contributes to the solution: AI technology Overview of the technology: AI technology has improved significantly with the emergence of large-scale language models that can communicate with users in natural language.

[0056] The large-scale language model unit 9 generates, for example, the following sentence in response to the above instruction sentence, i.e., the prompt starting with "You have improved...": "An autonomous vehicle exclusively for the elderly. This vehicle is equipped with AI that takes into account the specific problems of the elderly (decreased reaction speed, decreased eyesight and hearing, etc.) and supports safe driving."

[0057] In step S105, the generation unit 4 checks whether the loaded technical element is the last item in the list of technical elements that solve the problem selected by the problem / technology selection unit 31. If it is the last item (Yes in S105), the process proceeds to step S107, and if it is not the last item (No in S105), the process proceeds to step S106.

[0058] In step S106, the generation unit 4 reads the next technological element of the item selected by the problem / technology selection unit 31 in the list of technological elements that solve the problem, and returns to step S104.

[0059] In step S107, the generation unit 4 checks whether the loaded problem is the last item in the list of problems to be solved selected by the problem / technology selection unit 31. If it is the last item (Yes in S107), the process ends, and if it is not the last item (No in S105), the process proceeds to step S108.

[0060] In step S108, the generation unit 4 reads the next problem in the list of problems to be solved after the item selected by the problem / technology selection unit 31, and returns to step S103.

[0061] 11 is a diagram showing an example of the display of the solution 51 that the generation unit 4 finds according to the flowchart of FIG. 10 and notifies the solution display unit 6 by the output unit 5. The solution 51 is output in a table format that can be viewed at a glance, with the horizontal axis displaying the selected problem to be solved and the vertical axis displaying the selected technical elements. Then, the solution text (solution 51) generated by the large-scale language model unit 9 by combining the selected problem to be solved and the technical elements is displayed. １１ ~Solution ３３ ) is displayed. Solution １１ The message displayed will be, "An autonomous vehicle exclusively for the elderly. This vehicle is equipped with AI that takes into account the specific problems faced by the elderly (decreased reaction speed, reduced eyesight and hearing, etc.) and provides safe driving support." as described above.

[0062] In FIG. 11, a case is described in which a solution is proposed to solve the problem to be solved using one technical element, but it is also possible to propose solutions to solve the problem to be solved using each combination of selected technical elements.

[0063] In other words, when the technological elements of "AI technology," "EV popularization," and "IoT" are selected, a solution may be proposed that combines each of the technological elements of "AI technology," "EV popularization," "IoT," "AI technology + EV popularization," "EV popularization + IoT," "IoT + AI technology," and "AI technology + EV popularization + IoT" with the problem to be solved.

[0064] The solution generation device described above makes it possible to automatically generate text describing a solution that combines the problem to be solved with technical elements that have the potential to solve that problem, and depending on the progress phase of the discussion meeting where multiple users are discussing solutions, it is possible to generate more diverse solutions, for example, when multiple ideas are being widely considered in the early stages of the discussion meeting, and to generate more consistent solutions in the later stages of the discussion meeting, thereby making it possible to propose solutions that are appropriate to the progress of the discussion.

[0065] Next, a process of digging deeper into a solution by combining a predetermined problem to be solved and a technical element will be described with reference to Fig. 12. Fig. 12 shows the solution indicated by the user as the solution to be digged deeper in the display of the solution 51 notified to the solution display unit 6 described in Fig. 11. １１ 10 is a diagram showing a display example of an additional viewpoint processing area 332 for inputting an additional viewpoint regarding the object.

[0066] The additional viewpoint processing area 332 allows the user to select solutions to be explored further, such as solutions １１ When the user taps on this area, it is displayed as a pop-up window. The additional viewpoint processing area 332 is configured from a text box for inputting additional viewpoints for digging deeper into the solution, an area for displaying the additional viewpoints registered in the additional viewpoint information 331, and the digged-out solution.

[0067] As will be described in detail later, the solution selection unit 32 of the input unit 3 selects a solution to be explored further from the solutions output by the output unit 5 to the solution display unit 6 based on a user instruction. Then, the issues and technical elements corresponding to the selected solution are determined. The additional perspective input unit 33 of the input unit 3 displays an additional perspective processing area 332 and inputs additional perspectives to be considered in the solution to be explored further that is selected by the solution selection unit 32. The additional perspective input unit 33 additionally registers the input additional perspective in the additional perspective information 331.

[0068] The generation unit 4 then generates a prompt to be input to the large-scale language model unit 9 from the problem and technical elements corresponding to the solution selected by the solution selection unit 32 and the additional perspective input by the additional perspective input unit 33, and notifies the large-scale language model unit 9 of the generated prompt to instruct it to create a sentence describing the solution. The generation unit 4 then acquires the sentence describing the in-depth solution created by the large-scale language model unit 9 and notifies the output unit 5. The output unit 5 displays the notified sentence as an "in-depth solution" in the additional perspective processing area 332.

[0069] Next, the solution document generation process executed by the generator 4 for digging deeper by adding additional viewpoints will be described with reference to the flowchart in Fig. 13. In step S131, the generator 4 reads the diversity parameters 8 set by the setting unit 7.

[0070] In step S132, the generation unit 4 selects, via the solution selection unit 32, the solution indicated by the user as the solution to be explored in depth from the solutions 51 in the solution display unit 6. In step S133, the generation unit 4 refers to the additional viewpoint information 331 (FIG. 7) and reads from the additional viewpoint information 331 the additional viewpoint input by the additional viewpoint input unit 33 to the additional viewpoint processing area 332 described in FIG.

[0071] In step S134, the generation unit 4 reads the problem corresponding to the solution to be explored selected by the solution selection unit 32. In step S135, the generation unit 4 reads the technical element corresponding to the solution to be explored selected by the solution selection unit 32.

[0072] In step S136, the generation unit 4 generates a prompt to be input to the large-scale language model unit 9 from the additional perspective read in step S133, the problem read in step S134, and the technical elements read in step S135, and notifies the large-scale language model unit 9 of the generated prompt and the read diversity parameters 8, instructing it to create a sentence describing a solution.The generation unit 4 then acquires the sentence describing the solution created by the large-scale language model unit 9, and displays it as an "in-depth solution" in the additional perspective processing area 332.

[0073] For example, the generator 4 generates the following prompt (instruction sentence): You are an expert in forced idea generation in design thinking workshops. You have generated a "solution" for the specified "problem to be solved" as a solution using "technical elements that contribute to the solution." Please propose a new, more advanced "solution" using the "additional perspectives" added by the user.

[0074] Issue to be solved: Aging society Overview of the issue: The population of people aged 65 and over accounted for 7.1% of Japan's total population in 1970, but will reach 37.7% by 2050. Technological element that contributes to the solution: AI technology Overview of the technology: The emergence of large-scale language models that can communicate with users in natural language has led to significant improvements in AI technology. Solution: Self-driving cars designed specifically for the elderly. These cars are equipped with AI that takes into account issues specific to the elderly (slower reaction speed, reduced eyesight and hearing, etc.) and provide safe driving support. Additional perspective 1: Targets the automotive industry. Additional perspective 2: Support functions are configured by obtaining individual user issues (reduced eyesight, hearing, muscle strength, etc.) from a hospital database. Security technology is applied to connect to external databases.

[0075] The large-scale language model unit 9 generates sentences such as the following in response to the above instruction: Development of a personalized self-driving car utilizing AI technology. This car obtains individual user issues, such as reduced eyesight, hearing, or muscle strength, from a hospital database and provides driving support based on that information. For example, for a user with reduced eyesight, the vehicle recognizes the surrounding situation and provides driving information via voice. For a user with reduced muscle strength, the vehicle provides support tailored to the user's condition, such as automatically performing driving operations. This system will provide new business opportunities for the automotive industry and solve mobility problems in an aging society. Furthermore, the latest security technology will be applied to the handling of personal information, ensuring thorough privacy protection.

[0076] In this way, the user can select a solution that they want to develop from the text describing the solutions generated by the solution generation device, add perspectives to consider, and regenerate the solution, allowing the user to improve the solution through confirmation and addition of perspectives, thereby realizing the brushing up of solutions through collaboration between large-scale language models (generative AI) and humans.

[0077] Furthermore, the present invention is not limited to the above-described embodiments, and includes various modifications. The above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment.

[0078] REFERENCE SIGNS LIST 1 Problem database 2 Technology element database 3 Input section 31 Problem / technology selection section 32 Solution selection section 33 Additional viewpoint input section 4 Generation section 5 Output section 6 Solution display section 7 Setting section 71 Diversity setting selection section 8 Diversity parameters 9 Large-scale language model section

Claims

1. A solution generation device that finds solutions to problems using a language model, comprising: an input unit that inputs problems and multiple combinations of technical elements that solve the problems from a problem database and a technical element database; a generation unit that generates instruction information for a language model for each combination of problem and technical element input by said input unit, inputs said instruction information into said language model, sets a diversity parameter in said language model that indicates the diversity of solutions found by said language model, and instructs said language model to generate sentences of solutions; and an output unit that outputs the solutions found by said language model using the problem and technical elements input by said input unit as items.

2. A solution generation system according to claim 1, wherein said diversity parameters include a first parameter indicating a distribution of word occurrence probabilities, and a second parameter indicating a threshold of the top cumulative probabilities when selecting words to appear.

3. The solution generation system according to claim 2, wherein the first parameter is Temperature, which indicates the distribution of the probability of token appearance, and the second parameter is Top_p, which indicates the threshold of the top cumulative probability when selecting tokens to appear.

4. The solution generating system according to claim 1, further comprising a setting unit for setting said diversity parameters in accordance with a plurality of preset modes of diversity.

5. A solution generation device according to claim 4, wherein the modes include mode settings of a divergence mode, a discussion mode, and a convergence mode, and when the divergence mode is input, the setting unit sets the first parameter value and the second parameter value higher than the first parameter value and the second parameter value of the discussion mode, and when the discussion mode is input, the setting unit sets the first parameter value and the second parameter value higher than the first parameter value and the second parameter value of the convergence mode.

6. A solution generation device as claimed in claim 1, comprising: a solution selection unit which selects a solution to be explored in depth; and an additional viewpoint input unit which inputs additional viewpoints when exploring in depth, wherein the generation unit generates instruction information for the language model from the problem and technical elements corresponding to the solution selected by the solution selection unit and the additional viewpoints input by the additional viewpoint input unit, inputs this to the language model, and sets a diversity parameter in the language model to instruct the language model to generate a sentence of the solution, and the output unit outputs the solution found by the language model from the problem, technical elements and additional viewpoints.

7. The solution generation system according to claim 1, wherein said output unit outputs said solutions in a table format.

8. The solution generator according to claim 1, wherein said instruction information is a prompt.

9. A solution generation method for finding a solution to a problem using a language model, comprising the steps of: inputting a problem and a plurality of combinations of technical elements that solve the problem from a problem database and a technical element database; generating instruction information for the language model for each input combination of problem and technical element; setting the instruction information and a diversity parameter that indicates the diversity of solutions found by the language model in the language model, and instructing the language model to generate a sentence of the solution; and outputting the solution found by the language model using the problem and the technical element as items.

10. A solution generation method according to claim 9, further comprising the steps of: selecting a solution to be explored in depth; inputting additional perspectives to be used when exploring in depth; reading in the issues and technical elements corresponding to the selected solution to be explored in depth; generating instruction information for the language model from the read in issues and technical elements and the input additional perspectives; setting in the language model the instruction information and a diversity parameter indicating the diversity of solutions sought by the language model, and instructing the language model to generate a sentence of the solution; and outputting the solution sought by the language model as an explored solution.

Images