Information processing method, information processing device, and computer program

Abstract

This information processing device 16 acquires a first answer of a generative AI based on prescribed input information. The information processing device 16 causes the generative AI to execute an evaluation based on a plurality of evaluation items for the acquired first answer. The information processing device 16 acquires a second answer obtained by correcting the first answer on the basis of the result of the evaluation. The information processing device 16 outputs the acquired second answer.

Description

Information Processing Method, Information Processing Apparatus, and Computer Program

[0001] The present disclosure relates to data processing technology, and particularly to an information processing method, an information processing apparatus, and a computer program.

[0002] Patent Document 1 below describes a technique for automatically evaluating learning materials.

[0003] Japanese Patent Application Laid-Open No. 2025-130431

[0004] In recent years, the use of generative AI has become widespread. Generative AI can generate answers regarding various fields. The correct answers of learning materials such as those in Patent Document 1 are based on general facts in the world and are basically determined to be one. On the other hand, among the answers of generative AI, there are some with a wide range of correct answers, or in other words, those where the correct answer is not necessarily determined to be one. One object of the present disclosure is to provide a technique for suitably ensuring the quality of answers utilizing generative AI.

[0005] To solve the above problems, an information processing method according to an aspect of the present disclosure is an information processing method in a computer, including obtaining a first answer of a generative AI based on predetermined input information, causing the generative AI to perform an evaluation based on a plurality of evaluation items for the first answer, obtaining a second answer obtained by correcting the first answer based on the result of the evaluation, and outputting the second answer. <​​​​​​​​​

[0009] This is a diagram showing the configuration of the cooking support system of the embodiment. This is a block diagram showing the functional blocks provided by the information processing device of the embodiment. This is a flowchart showing the operation of the information processing device of the embodiment. This is a diagram showing an example of similarity evaluation between multiple reference general recipe information and target general recipe information. This is a diagram showing an example of fine-tuning of the embedding model. This is a diagram showing an example of similarity calculation rules. This is a diagram illustrating a modified version of the fine-tuning of the embedding model. This is a diagram showing an example of quality evaluation of automatic cooking recipe information. This is a diagram showing an example of an evaluation prompt. This is a diagram showing an example of an evaluation prompt. This is a diagram showing an example of an automatic cooking recipe information before modification. This is a diagram showing an example of a modification instruction prompt. This is a diagram showing an example of an automatic cooking recipe information after modification.

[0010] The subject of the apparatus or method in this disclosure comprises a computer. The functions of the subject of the apparatus or method in this disclosure are realized by the computer executing a program. The computer comprises a processor as its main hardware component, which operates according to the program. The processor is of any type as long as it can realize its functions by executing the program. The processor consists of one or more electronic circuits, including an Integrated Circuit (IC) or a Large Scale Integration (LSI). Multiple electronic circuits may be integrated on one chip or provided on multiple chips. Multiple chips may be aggregated in one device or provided on multiple devices. The program may be recorded on a non-temporary recording medium such as a Read Only Memory (ROM), optical disc, or hard disk drive that is readable by the computer, or on a temporary storage medium such as a Random Access Memory (RAM) that is readable by the computer. The program may be pre-stored on the recording medium, or it may be supplied to the recording medium or storage medium via a wide-area communication network, including the Internet.

[0011] Generative AI, also known as generative artificial intelligence, is an artificial intelligence system that generates text, images, and other content in response to prompts. A prompt is input information for the generative AI. A prompt includes natural language text describing the task that the generative AI should perform; in other words, it includes input text that indicates instructions to the generative AI. A prompt may also include non-text content such as images. The generative AI in this embodiment uses a large-scale language model (hereinafter also called "LLM" (Large Language Model)) that imitates human conversation through text-based dialogue; in other words, it functions as an artificial intelligence chatbot. The LLM, for example, generates and outputs response information (response text) in response to the input prompt.

[0012] The embodiments are outlined below. The user obtains the desired deliverables using LLM. The deliverables in these embodiments are recipe information for creating dishes using an automatic cooker (hereinafter also referred to as "automatic cooking recipe information"). The automatic cooker is also called a cooking appliance, electric automatic cooker, or automatic cooking pot. The recipe information includes, for example, the name of the dish, the ingredients for creating the dish, and information about the cooking process.

[0013] The cooking support system of this embodiment has, as its first feature, a configuration that utilizes LLM to create appropriate automatic cooking recipe information.

[0014] Specifically, the first feature includes the following (1)-(4): (1) The LLM generates a general recipe for the dish desired by the user and evaluates its similarity to existing recipes designated for the automatic cooking device. (2) If the similarity is low, the desired dish is deemed unsuitable for creation by the automatic cooking device, and automatic cooking recipe information is not output. (3) If the similarity is high, the LLM generates automatic cooking recipe information corresponding to the desired dish, and the quality of the generated automatic cooking recipe information is evaluated (e.g., error detection). (4) If the quality is low (e.g., an error is detected), the LLM corrects the problematic part (e.g., the error), and the corrected automatic cooking recipe information is output.

[0015] According to the first feature, if the dish desired by the user is unsuitable for cooking with an automatic cooker, the output of automatic cooking recipe information can be suppressed, thereby reducing the amount of computation and computational cost. Furthermore, if the quality of the automatic cooking recipe information is low, the quality of the automatic cooking recipe information can be improved by modifying it.

[0016] A second feature of the cooking support system of this embodiment is that it has a configuration for evaluating the similarity of recipes with high accuracy.

[0017] Specifically, the second feature includes the following (1)-(3): (1) Evaluating the similarity between a general recipe that does not use an automatic cooker and an existing recipe for an automatic cooker for a dish desired by the user. (2) Evaluating the similarity between recipes using vector search with an embedded model. (3) Fine-tuning the embedded model based on the category classification of the dish and the cooking process.

[0018] According to the second feature, it can evaluate the similarity between recipes with higher accuracy than conventional recipe searches based on dish names or ingredients.

[0019] As a third feature of the cooking support system of this embodiment, when creating automatic cooking recipe information using LLM, it is configured to create high-quality automatic cooking recipe information.

[0020] Specifically, the third feature includes the following (1)-(3): (1) The LLM generates automatic cooking recipe information to create the dish desired by the user, and uses the LLM to perform a quality evaluation based on multiple items (e.g., error detection). (2) If the quality of the automatic cooking recipe information generated in (1) is evaluated as low (e.g., errors are detected), the automatic cooking recipe information is regenerated or adjusted. (3) If the quality of the automatic cooking recipe information generated in (1) is evaluated as high (e.g., no errors are detected), the automatic cooking recipe information is output.

[0021] According to the third feature, the quality of automated cooking recipe information can be evaluated based on multiple perspectives corresponding to the required quality, and automated cooking recipe information that meets the required quality can be output. Furthermore, it can output high-quality recipes that do not contain errors.

[0022] The details of the embodiment will now be described. Figure 1 shows the configuration of the cooking support system 10 of the embodiment. The cooking support system 10 can also be called an information processing system and comprises an automatic cooker 12, an LLM server 14, and an information processing device 16. Each device in Figure 1 has a communication function and sends and receives data via a communication network 18. The communication network 18 includes, for example, a LAN, WAN, or the Internet.

[0023] The automatic cooker 12 is a cooking appliance that automatically performs heating and pressure cooking when ingredients and seasonings are added and a start operation is input. The automatic cooker 12 has multiple functions such as heating, pressurizing, stirring, standby, and notification. The automatic cooker 12 has multiple APIs (Application Programming Interfaces) which are interfaces for instructing the execution of these multiple functions from an external device (for example, an information processing device 16). As an example, the automatic cooker 12 may be a microwave oven, a microwave oven, or a rice cooker. Furthermore, the automatic cooker 12 does not necessarily have to have all of the multiple functions such as heating, pressurizing, stirring, standby, and notification; it is sufficient to have one or more functions.

[0024] The LLM server 14 is a computer that stores data of the generated AI, including the LLM, and provides a dialogue service using the generated AI. In this embodiment, the first LLM and the second LLM may be the same LLM or different LLMs. Different LLMs may differ in at least one of the following: session, server, type, performance, or price (price for API use). The type of LLM may be classified by at least one of the following: model structure, learning method, application / optimization. Furthermore, either the first LLM or the second LLM, or both, may be AI agents.

[0025] In this embodiment, an LLM server 14, separate from the information processing device 16, is configured to hold the LLM. However, as a variation, the automatic cooker 12 or the information processing device 16 may store the LLM and provide an interactive service using the LLM.

[0026] The information processing device 16 is a computer operated by a user. The information processing device 16 may be a PC, a tablet terminal, a smartphone, or a smart speaker. Alternatively, the information processing device 16 may be a terminal that can be attached to a person or device. Furthermore, the functions of the information processing device 16 may be included in the automatic cooker 12; in other words, the automatic cooker 12 and the information processing device 16 may be configured as an integrated unit.

[0027] Figure 2 is a block diagram showing the functional blocks of the information processing device 16 in this embodiment. The various functions shown in Figure 2 are realized by the computer mounted on the information processing device 16 executing a program. The computer includes hardware such as memory for loading programs, one or more processors (CPUs, etc.) for executing the loaded programs, auxiliary storage devices, and other LSIs. The processor is composed of multiple electronic circuits, including semiconductor integrated circuits and LSIs, and these multiple electronic circuits may be mounted on a single chip or on multiple chips. The functional blocks shown in Figure 2 are realized through the cooperation of hardware and software, and therefore, it will be understood by those skilled in the art that these functional blocks can be realized in various ways using hardware alone, software alone, or a combination thereof.

[0028] The information processing device 16 comprises an input unit 20, a display unit 22, a processing unit 24, a storage unit 26, and a communication unit 28. The input unit 20 includes operating elements such as buttons, touch panels, and keyboards, and receives user operations and inputs user operation information (input information) to the processing unit 24. The input unit 20 may also receive operation instructions based on voice input to a microphone (not shown). The display unit 22 displays various information. The display unit 22 may include a liquid crystal display or an organic EL display.

[0029] The processing unit 24 performs various information processing. The processing unit 24 may be implemented by a processor provided in the information processing device 16. The storage unit 26 stores various data that are referenced or updated by the processing unit 24. The storage unit 26 may be implemented by a memory or storage provided in the information processing device 16. The communication unit 28 communicates with external devices according to a predetermined communication protocol. The processing unit 24 sends and receives data with the automatic cooker 12 and the LLM server 14 via the communication unit 28.

[0030] The memory unit 26 includes a cooking appliance information memory unit 30, a recipe information memory unit 32, and an embedding model memory unit 34. The cooking appliance information memory unit 30 stores information about the automatic cooking appliance 12 used by the user (hereinafter also referred to as "cooking appliance information"). The cooking appliance information includes, for example, information about cooking processes that the automatic cooking appliance 12 can perform, information about ingredients that should not be put into the automatic cooking appliance 12, information about the correct and incorrect use of the automatic cooking appliance 12, and information about tips for cooking using the automatic cooking appliance 12.

[0031] The recipe information storage unit 32 stores predetermined recipe information (automatic cooking recipe information) relating to dishes that can be prepared using the automatic cooker 12. The automatic cooking recipe information includes, for example, the name of the dish, the ingredients for making the dish, and information about the cooking process (including the functions of the automatic cooker 12 used).

[0032] The embedding model storage unit 34 stores the data of the embedding model. The embedding model is a machine learning model (also called a mathematical model or transformation model) that converts arbitrary data such as text and images into numerical vectors. In this embodiment, the embedding model converts strings related to ingredients and cooking processes contained in recipe information into numerical vector representations.

[0033] In this embodiment, the information processing device 16 is configured to hold the embedding model, but as a modification, an external device to the information processing device 16 may store the embedding model and convert the recipe information into numerical values.

[0034] The processing unit 24 includes an input information acquisition unit 40, a prompt generation unit 42, an answer acquisition unit 44, a similarity evaluation unit 46, a quality evaluation unit 48, an answer output unit 50, an answer correction unit 52, and an embedding model update unit 54.

[0035] The functions of the multiple functional blocks included in the processing unit 24 may be implemented in a computer program. This computer program may be recorded on a non-temporary recording medium, or installed from that recording medium to the storage of the information processing device 16. Alternatively, this computer program may be downloaded from a predetermined server via a network and installed to the storage of the information processing device 16. The processor of the information processing device 16 may perform the functions of these multiple functional blocks by reading this computer program into main memory and executing it.

[0036] The input information acquisition unit 40 acquires user input information for the LLM server 14 (LLM). The input information in this embodiment includes information about the dish the user wants to make using the automatic cooker 12 (hereinafter also referred to as the "desired dish").

[0037] The prompt generation unit 42 generates a prompt that includes instructions for the LLM server 14. The prompt includes, for example, input information acquired by the input information acquisition unit 40.

[0038] The response acquisition unit 44 inputs the prompt generated by the prompt generation unit 42 to the LLM (first LLM) of the LLM server 14, thereby acquiring the first response from the first LLM regarding the output that the user expects (which can be described as the user's desired dish, or as the user's desired automatic cooking recipe information). In this embodiment, the first response is general recipe information for making the desired dish without using the automatic cooking device 12 (hereinafter also referred to as "target general recipe information").

[0039] The similarity evaluation unit 46 uses the embedding model stored in the embedding model storage unit 34 to evaluate the similarity between the reference information relating to the automatic cooking recipe information and the target general recipe information as the first answer, as a first evaluation. The reference information relating to the automatic cooking recipe information is information about recipes for making dishes predetermined as dishes that can be made using the automatic cooker 12. In this embodiment, the above reference information is general recipe information (hereinafter also referred to as "reference general recipe information") for making dishes that can be made with the automatic cooker 12 without using the automatic cooker 12. The reference information includes multiple reference general recipe information corresponding to multiple dishes that can be made with the automatic cooker 12.

[0040] The first evaluation includes an evaluation of the similarity between the textual information (including strings indicating ingredients and processes) contained in each of the multiple reference general recipe information and the textual information (including strings indicating ingredients and processes) contained in the target general recipe information. Specifically, the similarity evaluation unit 46 evaluates the similarity between the numerical vector information corresponding to the text of each of the multiple reference general recipe information, which has been transformed using an embedding model, and the numerical vector information corresponding to the text of the target general recipe information.

[0041] The response acquisition unit 44 acquires a second response from the LLM (first LLM) of the LLM server 14 regarding automatic cooking recipe information, in accordance with the result of the first evaluation by the similarity evaluation unit 46. In this embodiment, the second response includes automatic cooking recipe information, which is recipe information for making a desired dish using the automatic cooker 12. In this embodiment, "first response" and "second response" are names used to distinguish and explain the responses, and do not necessarily have to be different in content. For example, the second response may be the same as the first response.

[0042] The quality evaluation unit 48 performs a quality evaluation (second evaluation) of the automatic cooking recipe information indicated by the second response, in accordance with the results of the first evaluation by the similarity evaluation unit 46. The quality evaluation unit 48 causes the LLM (second LLM) of the LLM server 14 to perform a second evaluation based on one or more evaluation items (multiple evaluation items in the embodiment) for the automatic cooking recipe information indicated by the second response.

[0043] Based on the result of the second evaluation by the quality evaluation unit 48, the response output unit 50 outputs a second response (in the embodiment, automatic cooking recipe information). When it is evaluated that the second response satisfies one or more evaluation items (in the embodiment, a plurality of evaluation items) in the second evaluation, the response output unit 50 outputs the second response to the outside. When the second response does not satisfy one or more evaluation items (in the embodiment, a plurality of evaluation items), the response output unit 50 suppresses outputting the second response to the outside.

[0044] Based on the result of the second evaluation by the quality evaluation unit 48, the response correction unit 52 obtains a new second response (new automatic cooking recipe information) obtained by correcting the second response of the evaluation target. The response correction unit 52 may correct the second response of the evaluation target based on the rule based on the result of the second evaluation, or may input the result of the second evaluation to the LLM (first LLM) of the LLM server 14 to cause the first LLM to generate a new second response. Based on the result of the second evaluation by the quality evaluation unit 48 for the new second response, the response output unit 50 outputs the new second response to the outside.

[0045] The embedding model update unit 54 fine-tunes the embedding model stored in the embedding model storage unit 34. The embedding model update unit 54 updates the data (parameters, weights, etc.) of the embedding model by fine-tuning, and stores the updated data of the embedding model in the embedding model storage unit 34.

[0046] The operation of the information processing apparatus 16 having the above configuration will be described. FIG. 3 is a flowchart showing the operation of the information processing apparatus 16 according to the embodiment. The input information acquisition unit 40 acquires the input information of the user (hereinafter also referred to as "desired cooking information") input via the input unit 20 (S10). The desired cooking information includes at least the name of the desired cooking of the user. The desired cooking information may include information such as the ingredients the user wants to use, food preferences, and expected tastes in addition to the name of the desired cooking.

[0047] In addition, the user's input information may be an image or photograph of a desired dish, or information that does not directly indicate the name of the desired dish, such as an explanation of the desired dish or information about the information source. The information about the information source may include the URL of a recipe site, information for specifying a cookbook and page, etc. In this case, the information processing device 16 may estimate the user's desired dish based on the user's input information and use the name of the estimated dish as the desired dish information.

[0048] The prompt generation unit 42 generates a prompt including content instructing the creation of general recipe information (target general recipe information) for making the dish indicated by the desired dish information without using the automatic cooking device 12. This prompt includes the desired dish information (such as the dish name) and content instructing to answer the ingredients and cooking steps for making the dish. Also, information about the automatic cooking device 12 may be excluded in this prompt. The answer acquisition unit 44 transmits the prompt generated by the prompt generation unit 42 to the LLM server 14 and acquires the target general recipe information as the first answer generated by the first LLM of the LLM server 14 from the LLM server 14 (S12).

[0049] The prompt generation unit 42 generates a prompt including content instructing the creation of general recipe information (multiple reference general recipe information) for making each of the multiple dishes that can be made by the automatic cooking device 12 without using the automatic cooking device 12. This prompt includes information about the multiple dishes that can be made by the automatic cooking device 12 (such as the name of each dish) and content instructing to answer the ingredients and cooking steps for making the dish. Also, information about the automatic cooking device 12 may be excluded in this prompt. The answer acquisition unit 44 transmits the prompt generated by the prompt generation unit 42 to the LLM server 14 and acquires the multiple reference general recipe information generated by the first LLM of the LLM server 14 from the LLM server 14 (S14).

[0050] The similarity evaluation unit 46 uses the embedding model stored in the embedding model storage unit 34 to evaluate the similarity between multiple reference general recipe information and the target general recipe information (S16).

[0051] Figure 4 shows an example of similarity evaluation between multiple reference general recipe information and target general recipe information. Each of the multiple reference general recipe information and the target general recipe information contains a string indicating the ingredients and cooking steps for making the dish. The similarity evaluation unit 46 inputs the strings of each of the multiple reference general recipe information into the embedding model and obtains multiple numerical vectors corresponding to the multiple reference general recipe information output by the embedding model. The similarity evaluation unit 46 also inputs the string of the target general recipe information into the embedding model and obtains a numerical vector corresponding to the target general recipe information output by the embedding model.

[0052] The similarity evaluation unit 46 calculates the cosine similarity (a value between 0 and 1) for each pair of numerical vectors of the target general recipe information and the respective reference general recipe information using a known method. The cosine similarity may also be the dot product of the numerical vectors, with a value closer to 1 indicating higher similarity and a value closer to 0 indicating lower similarity. The similarity evaluation unit 46 extracts the reference general recipe information (existing menu of the automatic cooker 12) with the highest cosine similarity to the target general recipe information (desired dish).

[0053] The embedding model update unit 54 fine-tunes the embedding model stored in the embedding model storage unit 34 when predetermined conditions are met. The predetermined conditions may be that a predetermined period of time has elapsed since the last fine-tuning, or that an instruction has been input from an administrator or operator to instruct fine-tuning.

[0054] Figure 5 shows an example of fine-tuning the embedding model. The embedding model update unit 54 performs the processing shown in Figure 5 for each pair of existing menus extracted from a plurality of existing menus predetermined for the automatic cooker 12 and stored in the recipe information storage unit 32. Hereinafter, the two existing menus to be processed will be referred to as existing menu A and existing menu B.

[0055] The recipe information storage unit 32 stores information for each of several existing menus. The information for each existing menu includes the name of the dish and classification information. The classification information includes multiple items, such as the genre of the dish (stew, soup, curry, seafood, etc.), the main ingredients before preparation (chicken thigh, pork loin, minced meat, etc.), the preparation process (human work, etc.), the ingredients added during preparation (seasonings, ingredients, etc.), whether or not mixing is done, and the cooking method (pressure cooking, waterless cooking, stew, etc.).

[0056] The embedding model update unit 54 calculates the correct similarity score (a value between 0 and 1) between existing menu A and existing menu B based on predefined rules (i.e., rule-based). Figure 6 shows an example of the similarity calculation rules. The embedding model update unit 54 compares the classification information of existing menu A and existing menu B item by item, and calculates the correct similarity score by adding predetermined scores to the matching items. The higher the similarity, the closer the score approaches 1, and the lower the similarity, the closer the score approaches 0.

[0057] Note that a single item in the classification information may have multiple values ​​set. For example, "stew, soup" may be set as the first dish category, or "mussels, shrimp, squid" may be set as the main ingredients before preparation. In this case, the embedding model update unit 54 may calculate the number of matching values ​​relative to the total number of values ​​as the matching rate, and add the result of multiplying the score of the item by the matching rate to the correct score.

[0058] Returning to Figure 5, the embedding model update unit 54 inputs the dish names for existing menu A and existing menu B into the LLM of the LLM server 14, thereby obtaining the general recipe information for each of existing menu A and existing menu B generated by the LLM (recipe information that does not use the automatic cooker 12, and in Figure 5, general recipe A and general recipe B). The method for obtaining the general recipe information here may be the same as the method for obtaining the target general recipe information and reference general recipe information described above.

[0059] The embedding model update unit 54 inputs the string for general recipe A into the embedding model and obtains the numerical vector corresponding to general recipe A output by the embedding model. The embedding model update unit 54 also inputs the string for general recipe B into the embedding model and obtains the numerical vector corresponding to general recipe B output by the embedding model.

[0060] The embedding model update unit 54 calculates the cosine similarity (a value between 0 and 1) between the numerical vector of general recipe A and the numerical vector of general recipe B using a known method. The embedding model update unit 54 updates the data (parameters, weights, etc.) of the embedding model so that the cosine similarity of each pair of existing menus extracted from a predetermined set of existing menus by the automatic cooker 12 approaches the correct score.

[0061] Figure 7 illustrates a modified example of fine-tuning the embedding model. Figure 7 shows the names and categories (genres) of several pre-defined existing menus in the automatic cooker 12. In this modified example as well, the embedding model update unit 54 obtains a general recipe for each dish, including ingredients and steps, based on the dish name.

[0062] In this modified version, the embedding model update unit 54 fine-tunes the embedding model using the Triplet loss method. Specifically, the embedding model update unit 54 creates all possible combinations of assigning three dishes to anchor, positive, and negative categories. The anchor is a reference sentence (in this case, a dish), the positive is a sentence close to the anchor (in this case, a dish in the same category as the anchor), and the negative is a sentence far from the anchor (in this case, a dish in a different category than the anchor). The embedding model update unit 54 keeps the sets where the anchor and positive categories match, and the anchor and negative categories are different, and uses them as training data.

[0063] In this modified example, the embedding model update unit 54 performs training on the embedding model such that the anchor dish (general recipe) and the positive dish (general recipe) are closer in the vector space, and the anchor dish (general recipe) and the negative dish (general recipe) are further apart in the vector space, that is, it determines the parameters and weights of the embedding model.

[0064] Returning to Figure 3, if the similarity between the target general recipe information and the reference general recipe information is less than a predetermined threshold (for example, a cosine similarity of 0.7), i.e., if the similarity is low (Y in S18), the response output unit 50 outputs information indicating that it cannot create automatic cooking recipe information to make the dish desired by the user (S20). The reference general recipe information here is the reference general recipe information with the highest cosine similarity to the target general recipe information among multiple reference general recipe information.

[0065] In S20, the response output unit 50 may output information indicating that the dish desired by the user cannot be prepared using the automatic cooker 12. The response output unit 50 may display this information on the display unit 22, or it may transmit this information to a user's terminal (not shown) for display. That is, if the similarity between the target general recipe information and the reference general recipe information is less than a predetermined threshold, the information processing device 16 will not perform the second evaluation described later, nor will it output the automatic cooking recipe information as the second response.

[0066] If the similarity between the target general recipe information and the reference general recipe information is greater than or equal to a predetermined threshold (for example, a cosine similarity of 0.7), that is, if the similarity is high (N in S18), the prompt generation unit 42 generates a prompt that includes content instructing the creation of recipe information (automatic cooking recipe information) for making the dish indicated by the desired dish information entered by the user using the automatic cooker 12. This prompt includes the desired dish information (such as the name of the dish) and includes content instructing the user to respond with information such as the ingredients for making the dish, the cooking process, and the functions of the automatic cooker 12 to be used.

[0067] The response acquisition unit 44 transmits the prompt generated by the prompt generation unit 42 to the LLM server 14 and acquires the automatic cooking recipe information as the second response, which has been generated by the first LLM of the LLM server 14 (S22). Note that "acquiring the second response" is not limited to the method of generating and acquiring a new response by the first LLM, but also includes the method of adopting (identifying) the already acquired first response as the second response. In this case, the second response may be the same as the first response.

[0068] The quality evaluation unit 48, as a second evaluation, uses the LLM (second LLM) of the LLM server 14 to evaluate the quality of the automatic cooking recipe information acquired in S22 based on multiple evaluation items (S24). The second evaluation includes determining whether or not the automatic cooking recipe information satisfies multiple evaluation items.

[0069] Figure 8 shows an example of quality evaluation of automated cooking recipe information. The prompt generation unit 42 generates multiple prompts (hereinafter also referred to as "evaluation prompts") corresponding to multiple evaluation items for evaluating the quality of automated cooking recipe information. Evaluation prompts can also be called error detection prompts.

[0070] The multiple evaluation items include items relating to the requirements that the automatic cooking recipe information must satisfy, and items relating to whether or not the automatic cooking recipe information contains errors. Furthermore, the multiple evaluation items include at least one of the following: evaluation items relating to the ingredients for making the dish, and evaluation items relating to the cooking process for making the dish; in embodiments, both are included. Furthermore, the multiple evaluation items include at least one of the following: evaluation items relating to the manner in which the automatic cooker 12 is used, and evaluation items relating to the cooking process using the automatic cooker 12; in embodiments, both are included.

[0071] Figure 8 shows the following (1) to (4) as examples of multiple evaluation items: (1) Evaluation items related to errors in the cooking process. (2) Evaluation items related to errors in post-processing (the process of adding ingredients after the first automatic cooking process and before the second automatic cooking process). (3) Evaluation items related to ingredients prohibited from use in the automatic cooker 12, and errors in the use of the automatic cooker 12. (4) Evaluation items related to tips for cooking with the automatic cooker 12.

[0072] The quality evaluation unit 48 transmits the automatic cooking recipe information acquired in S22 and each of the multiple evaluation prompts corresponding to the multiple evaluation items to the LLM server 14 and inputs them to the multiple second LLMs, thereby causing the multiple second LLMs to perform evaluations of each of the multiple evaluation items for the automatic cooking recipe information. In other words, the quality evaluation unit 48 causes the multiple second LLMs to perform evaluations of the automatic cooking recipe information for each evaluation item. Note that the multiple second LLMs may be provided by one LLM server 14, or by multiple LLM servers 14. As a modification, the quality evaluation unit 48 may cause one second LLM to perform evaluations of each of the multiple evaluation items for the automatic cooking recipe information multiple times.

[0073] Furthermore, each of the multiple evaluation prompts includes instructions to include information regarding the basis for the evaluation results based on evaluation criteria for the automated cooking recipe information in the response. This causes multiple second LLMs to output information regarding the basis for the evaluation results based on multiple evaluation criteria for the automated cooking recipe information.

[0074] Multiple second LLMs of the LLM server 14 calculate a similarity score (for example, a value between 0 and 1) according to the input evaluation prompt, and generate a response that includes that score and information indicating the basis for the evaluation (evaluation basis).

[0075] The quality evaluation unit 48 aggregates the results of evaluations performed by multiple second LLMs. As an example, if evaluations for each of multiple evaluation items are performed multiple times by one second LLM, the quality evaluation unit 48 aggregates the results of the evaluations performed multiple times by the one second LLM.

[0076] In this embodiment, the quality evaluation unit 48 obtains the mode of the scores from multiple second LLMs and the basis for the evaluation for each evaluation item. The quality evaluation unit 48 also aggregates the evaluation results from multiple LLMs for each evaluation item. For example, the quality evaluation unit 48 calculates a value that aggregates the scores (e.g., the mode) for each evaluation item. The aggregated value may be the average of the scores. The quality evaluation unit 48 also generates a document that aggregates the basis for the evaluation for each evaluation item (e.g., a document that excludes duplicate basis for evaluation).

[0077] Figure 9 shows an example of an evaluation prompt. The evaluation prompt 100 includes scoring rules, scoring procedures, scoring criteria, an example of an OK (1 point in Figure 9) automated cooking recipe information, and an example of an NG (0 points in Figure 9) automated cooking recipe information. Of these, at least the scoring rules differ for each evaluation item. The scoring rules of the evaluation prompt 100 in Figure 9 correspond to the evaluation items related to errors in the cooking process shown in Figure 8, or in other words, they are rules for detecting errors in the cooking process.

[0078] Although not shown in Figure 9, the evaluation prompt 100 may further include instructions to provide the basis for the evaluation (basis for scoring). The basis for the evaluation may include the reason for the evaluation (e.g., the location of the error in the automated cooking recipe information).

[0079] Figure 10 also shows an example of an evaluation prompt. The evaluation prompt 102 in Figure 10 has different scoring rules than the evaluation prompt 100 in Figure 9, but the other items are the same as the evaluation prompt 100 in Figure 9. The scoring rules for evaluation prompt 102 in Figure 10 correspond to the evaluation items related to late-input errors shown in Figure 8, or in other words, they are rules for detecting late-input errors.

[0080] Figure 11 also shows an example of an evaluation prompt. The scoring rules for evaluation prompt 104 in Figure 11 differ from those for evaluation prompt 100 in Figure 9, but the other items are the same as those for evaluation prompt 100 in Figure 9. The scoring rules for evaluation prompt 104 in Figure 11 correspond to the evaluation items related to ingredients prohibited from use in the automatic cooker 12 and errors in the use of the automatic cooker 12, as shown in Figure 8. In other words, these are rules for detecting the use of prohibited ingredients and errors in the use of the automatic cooker 12.

[0081] The quality of the second response of the first LLM (automatic cooking recipe information in the embodiment) for a given deliverable includes, in order of importance, basic quality, and attractive quality. Basic quality includes the minimum elements that must be met and the elements for ensuring safety. Basic quality includes elements that determine the essential value of the deliverable (dish), such as taste and texture (e.g., the quantity of ingredients and seasonings). Basic quality can also be called essential quality or basic quality. Attractive quality includes elements that, when met, create added value (e.g., appearance and presentation).

[0082] The four evaluation items shown in Figure 8 all assess basic quality. Therefore, the scoring criteria in the evaluation prompts from Figures 9 to 11 stipulate that if any of the prohibited items apply, the score will be 0. On the other hand, the scoring criteria in the evaluation prompts corresponding to the evaluation items related to unified quality or attractive quality do not have to be binary values ​​of OK (e.g., 1 point) and NG (e.g., 0 points), but may be set as a value between 0 and 1 or a continuous score.

[0083] Figure 12 shows an example of an evaluation prompt. The scoring rules in evaluation prompt 106 in Figure 12 are based on the expertise of cooking professionals and correspond to the evaluation items of unified quality. The scoring criteria for evaluation prompt 106 have a range of scores depending on the number of rules that are followed. Thus, the scoring rules included in evaluation prompt 100 are not limited to one rule, but may include multiple rules. The scoring procedure may define a procedure to check the description of the automated cooking recipe information being evaluated step by step for each of these multiple rules. The scoring criteria may have a range of scores depending on the number of rules that are met or not met among the multiple rules. For example, a first score may be given if all rules are met, a second score if a predetermined number or more rules are met, and a third score if none of the rules are met. The scores may be discrete values ​​such as 1 point, 0.5 points, and 0 points, or they may be values ​​between 0 and 1 point.

[0084] Returning to Figure 3, if the result of the quality evaluation in S24 (aggregate result) is below a predetermined threshold, in other words, if the quality of the automated cooking recipe information to be evaluated is low (Y in S26), the response correction unit 52 generates new automated cooking recipe information by correcting the automated cooking recipe information that was evaluated (S28). The automated cooking recipe information to be evaluated includes both the automated cooking recipe information generated in S22 and the automated cooking recipe information newly generated in S28.

[0085] As shown in Figure 8, if all of the evaluation items are related to basic quality, the threshold may be 1. The quality evaluation unit 48 may determine the aggregated score to be 0 if at least one of the multiple evaluation results (scores) corresponding to the multiple evaluation items is 0. In this way, if the automated cooking recipe information being evaluated is evaluated as not meeting at least one of the multiple evaluation items (basic quality), the aggregated score will be 0, which is below the threshold.

[0086] In this embodiment, if the result of the quality evaluation (aggregate result) in S24 is less than a predetermined threshold, that is, if the automated cooking recipe information to be evaluated is evaluated as not satisfying at least one of the multiple evaluation items, the response correction unit 52 inputs information regarding the evaluation result (evaluation result information) to the first LLM of the LLM server 14, causing the first LLM to generate new automated cooking recipe information with the automated cooking recipe information to be evaluated corrected. The evaluation result information may also include information regarding the basis for the evaluation generated by the second LLM in S24.

[0087] After S28, the process returns to S24, and the quality evaluation unit 48 instructs the second LLM of the LLM server 14 to perform an evaluation of the newly modified automatic cooking recipe information based on multiple evaluation items. Thereafter, if the automatic cooking recipe information to be evaluated does not satisfy at least one of the multiple evaluation items, the information processing device 16 repeatedly performs the acquisition (modification) and evaluation of new automatic cooking recipe information. In other words, the information processing device 16 repeatedly performs the acquisition (modification) and evaluation of new automatic cooking recipe information until the automatic cooking recipe information to be evaluated satisfies all of the multiple evaluation items.

[0088] If the result of the quality evaluation in S24 (aggregate result) is above a predetermined threshold, in other words, if the quality of the automated cooking recipe information to be evaluated (automated cooking recipe information generated in S22, or automated cooking recipe information newly generated in S28) is high (N in S26), the response output unit 50 outputs the automated cooking recipe information to be evaluated (S30). For example, the response output unit 50 displays the automated cooking recipe information to be evaluated (ingredients, cooking process, how to use the automated cooker 12, etc.) on the display unit 22.

[0089] Although not shown in Figure 3, the information processing device 16 may terminate the acquisition and evaluation of automatic cooking recipe information and suppress the output of the automatic cooking recipe information if the repeated acquisition and evaluation of automatic cooking recipe information as a second answer satisfies predetermined termination conditions. For example, the answer correction unit 52 (or quality evaluation unit 48) may count the number of times the automatic cooking recipe information is acquired (or evaluated), and may terminate the acquisition (or evaluation) of the automatic cooking recipe information if the number of times the automatic cooking recipe information is acquired (or evaluated) reaches a predetermined threshold (for example, 10 times) or more. The answer output unit 50 may suppress the output of the automatic cooking recipe information if the number of times the automatic cooking recipe information is acquired (or evaluated) reaches a predetermined threshold or more.

[0090] Alternatively, the quality evaluation unit 48 may evaluate the quality using a different LLM than the previous LLM if the repeated acquisition and evaluation of the automated cooking recipe information as the second response satisfies a predetermined termination condition (for example, if the number of repetitions reaches a predetermined threshold as described above). The different LLM may have a higher benchmark evaluation (for example, evaluation of language comprehension ability, reasoning ability, accuracy, dialogue ability, response performance, etc.) than the previous LLM. Or, the different LLM may have a higher API usage fee (token unit price or request unit price) than the previous LLM.

[0091] This section explains an example of correcting automated cooking recipe information. Figure 13 shows an example of automated cooking recipe information before correction. The automated cooking recipe information 120 in Figure 13 is a recipe for beef stew. In this case, the fact that the lid was opened and ingredients were added in "Cooking process 1 by the automated cooker" was detected as an error (prohibited item) in the quality evaluation by the 2nd LLM.

[0092] Figure 14 shows an example of a correction instruction prompt. The correction instruction prompt 122 includes instructions to correct the automatic cooking recipe information, and further includes the results of the quality evaluation by the second LLM (in this case, information on detected errors). The response correction unit 52 causes the prompt generation unit 42 to generate the correction instruction prompt 122, and by inputting the correction instruction prompt 122 to the first LLM of the LLM server 14, it obtains the new automatic cooking recipe information (corrected automatic cooking recipe information) generated by the first LLM.

[0093] Figure 15 shows an example of corrected automatic cooking recipe information. In the automatic cooking recipe information 124 of Figure 15, the lid is no longer opened in "Cooking process 1 by automatic cooker," meaning that the error in cooking process 1 before the correction has been eliminated.

[0094] According to the cooking support system 10 (information processing device 16) of this embodiment, the user can be provided with automatic cooking recipe information for creating any dish desired by the user using the automatic cooker 12. Furthermore, by determining in advance whether or not the desired dish can be cooked using the automatic cooker 12, the quality of the automatic cooking recipe information for creating the user's desired dish can be ensured.

[0095] Furthermore, according to the cooking support system 10 (information processing device 16) of this embodiment, when the similarity between the target general recipe information relating to the user's desired dish and the reference general recipe information relating to dishes that can be created by the automatic cooker 12 is high, the LLM is made to generate automatic cooking recipe information relating to the user's desired dish, and the quality of that automatic cooking recipe information is evaluated. This reduces the frequency of generating and evaluating the quality of automatic cooking recipe information, thereby reducing the computational load on the computer.

[0096] Furthermore, according to the cooking support system 10 (information processing device 16) of this embodiment, it is possible to obtain highly accurate evaluation results based on multiple evaluation items for automatic cooking recipe information for creating a dish desired by the user. In addition, based on the highly accurate evaluation results, it is possible to obtain new automatic cooking recipe information by appropriately modifying the automatic cooking recipe information that was evaluated.

[0097] The present disclosure has been described above based on embodiments. The embodiments are illustrative, and it will be understood by those skilled in the art that various modifications are possible for each component or combination of processing processes of the embodiments, and that such modifications are also within the scope of the present disclosure.

[0098] A first modified example will be described. The answer acquisition unit 44 of the information processing device 16 may, as the target of similarity evaluation, have the first LLM generate automatic cooking recipe information for creating the user's desired dish using the automatic cooker 12, rather than general recipe information. The similarity evaluation unit 46 of the information processing device 16 may determine the similarity between the automatic cooking recipe information generated by the first LLM and a plurality of pre-registered automatic cooking recipe information (a plurality of reference information) relating to a plurality of dishes that the automatic cooker 12 can create. For example, the answer acquisition unit 44 may acquire recipe information (automatic cooking recipe information) for creating the user's desired dish using the automatic cooker 12 as the first answer. In this case, the information processing device 16 may treat the acquired first answer as the second answer, evaluate the quality of the second answer (which may be the same as the first answer) using the second generation AI, and output the second answer according to the result of the quality evaluation.

[0099] The information processing device 16 may evaluate the quality of the automatic cooking recipe information if the similarity (highest similarity) between the automatic cooking recipe information and the multiple reference pieces is greater than or equal to a predetermined threshold. The information processing device 16 may suppress the quality evaluation and output of the automatic cooking recipe information if the similarity (highest similarity) between the automatic cooking recipe information and the multiple reference pieces is less than a predetermined threshold.

[0100] A second modification will be described. If the similarity score between the target general recipe information and the reference general recipe information relating to the user's desired dish is above a predetermined threshold, the information processing device 16 may, instead of having the LLM generate automatic cooking recipe information, output existing automatic cooking recipe information pre-installed in the automatic cooker 12 that corresponds to the reference general recipe information with a high similarity to the target general recipe information. The existing automatic cooking recipe information corresponding to the reference general recipe information can also be said to be existing automatic cooking recipe information for creating the dish indicated by the reference general recipe information.

[0101] In the second modified example, the quality evaluation unit 48 may or may not perform a quality evaluation on existing automated cooking recipe information corresponding to reference general recipe information that has a high similarity to the target general recipe information. This is because existing automated cooking recipe information is typically guaranteed to be of high quality. Furthermore, if the similarity score between the target general recipe information and the reference general recipe information relating to the user's desired dish is below a predetermined threshold, the information processing device 16 may, as in the embodiment, have the LLM generate automated cooking recipe information and evaluate the quality of that automated cooking recipe information.

[0102] A third modified example will be described. When a predetermined operation is input by the user to the screen of the display unit 22 which displays the automatic cooking recipe information, the response output unit 50 of the information processing device 16 may transmit the automatic cooking recipe information to the automatic cooking device 12 by calling one or more APIs provided on the automatic cooking device 12, thereby starting the operation of the automatic cooking device 12.

[0103] A fourth modification will be described. The information processing device 16 may further include a feedback receiving unit and a feedback reflection unit. The feedback receiving unit may receive user feedback information input via the input unit 20 or from an external device, and may receive feedback information regarding dishes cooked using the automatic cooker 12 based on the output automatic cooking recipe information. The feedback information may include at least information indicating whether the dish is delicious (success) or not (failure).

[0104] The feedback reflection unit may reflect feedback information in at least one of the similarity score calculation process and the content of the evaluation items (evaluation prompts) in the quality evaluation (error detection). According to this modification, the accuracy of the similarity evaluation or quality evaluation can be improved based on feedback regarding whether it tastes good (success) or bad (failure).

[0105] A fifth modified example will be described. The quality evaluation unit 48 may write to a separate evaluation prompt for each evaluation item (evaluation perspective) and query the second LLM (one or more second LLMs) for each evaluation item. Alternatively, the quality evaluation unit 48 may write information about multiple evaluation items to a single prompt and query the second LLM (one or more second LLMs) only once.

[0106] A sixth modified example will be described. The quality evaluation unit 48 may obtain a single evaluation result from the second LLM based on a single evaluation prompt and determine the quality of the automatic cooking recipe information based on that single evaluation result. Alternatively, the quality evaluation unit 48 may obtain multiple evaluation results from the second LLM based on a single evaluation prompt and determine the quality of the automatic cooking recipe information by calculating the average score of the multiple evaluation results, or by determining the quality of the automatic cooking recipe information by majority vote.

[0107] A seventh modification will be described. In the embodiment, the information processing device 16 evaluates whether the similarity between the target general recipe and the reference general recipe (existing recipe) is above a predetermined threshold, in other words, it evaluates the similarity in two stages: high and low. As a modification, the information processing device 16 may evaluate the similarity between the target general recipe and the reference general recipe in three stages: high, medium, and low. If the similarity is medium, the information processing device 16 may not immediately output the automatic cooking recipe information, but instead may request verification by a human (for example, a judgment of whether the similarity is high or low). Then, depending on the result of the human verification, it may perform either the low or high similarity processing. The human here may be, for example, a business operator using the cooking support system. Note that, as in the embodiment, the information processing device 16 evaluates the similarity between the target general recipe and the reference general recipe in two stages: high and low, and if the similarity is low, it may request verification by a human (for example, a business operator) as described above.

[0108] An eighth modification will now be described. The technology for generating suitable automatic cooking recipe information described in the embodiment is applicable to the creation of various deliverables. The technology of the embodiment is not limited to the automatic cooker 12, but is applicable to determining the control modes of various home appliances, for example, it is applicable to creating operation sequences for washing machines, dryers, or dishwashers. Furthermore, the technology of the embodiment is applicable to creating various text information, for example, it is applicable to creating learning materials, documents, or manuals.

[0109] An example of applying the technology of the embodiment to the creation of a washing machine operation sequence (a series of steps performed until washing is completed, hereinafter also referred to as the "washing sequence") will be described. The input information here includes information indicating the laundry specified by the user, or information indicating the laundry automatically detected by a sensor or the like. In the first evaluation, the similarity between the washing sequence generated by the first LLM based on the laundry indicated by the input information and several existing washing courses pre-registered in the washing machine (for example, delicate wash) is calculated using vector search.

[0110] In this case, a high degree of similarity means that there is a washing course similar to the washing sequence based on the laundry indicated by the input information, and a certain level of quality can be expected from the washing sequence. A moderate degree of similarity means that there is a washing course that is somewhat similar to the washing sequence based on the laundry indicated by the input information, but the quality of the washing is unclear and verification is required. A low degree of similarity means that there is no washing course similar to the washing sequence based on the laundry indicated by the input information, and it is highly likely that it will be difficult to wash the laundry indicated by the input information in the washing machine.

[0111] In the second evaluation, the second LLM is used to evaluate the quality of the washing sequence generated by the first LLM based on the laundry indicated by the input information, and, for example, detects errors in the washing sequence. Multiple evaluation items for the washing sequence may include, for example, the amount of water, the amount of detergent, constraints based on the type of laundry, and the quality of the finished wash. If the quality of the washing sequence to be evaluated is below a predetermined threshold, the information processing device 16 generates a new washing sequence with the washing sequence corrected. On the other hand, if the quality of the washing sequence to be evaluated is above the predetermined threshold, the information processing device 16 outputs the washing sequence.

[0112] An example of applying the technology of the embodiment to the creation of learning materials is described. Learning materials include teaching materials and documents prepared for learning and education. The input information here includes information entered by the user specifying the content they wish to learn (for example, high school third-year physics for the Common Test). In the first evaluation, the similarity between the learning material generated by the first LLM based on the learning content indicated by the input information and pre-registered existing learning materials is calculated using vector search.

[0113] In this case, a high degree of similarity means that there are existing learning materials similar to the learning materials generated based on the learning content indicated by the input information, and that the generated learning materials can be expected to be of a certain quality level. A moderate degree of similarity means that there are existing learning materials that are somewhat similar to the learning materials generated based on the learning content indicated by the input information, but the quality of the generated learning materials is unclear and verification is necessary. A low degree of similarity means that there are no existing learning materials similar to the learning materials generated based on the learning content indicated by the input information, and therefore it is highly likely that it will be difficult to generate learning materials of a certain quality level.

[0114] In the second evaluation, the second LLM is used to evaluate the quality of the learning materials generated by the first LLM based on the learning content indicated by the input information, and for example, errors contained in the learning materials are detected. Multiple evaluation items for the generated learning materials may include, for example, evaluation items related to consistency in question format (e.g., inconsistency if the common test is in a multiple-choice format but the learning materials are in a written format) and evaluation items related to consistency with the curriculum (educational plan). If the quality of the learning materials to be evaluated is below a predetermined threshold, the information processing device 16 generates new learning materials with the learning materials corrected. On the other hand, if the quality of the learning materials to be evaluated is above the predetermined threshold, the information processing device 16 outputs the learning materials.

[0115] Any combination of the embodiments and modifications described above is also useful as an embodiment of the present disclosure. The new embodiments resulting from such combinations will possess the combined effects of the respective embodiments and modifications. Furthermore, it will be understood by those skilled in the art that the functions to be performed by each component described in the claims can be achieved by each component shown in the embodiments and modifications individually or in combination thereof.

[0116] Furthermore, unless otherwise specified, terms such as "first," "second," etc., used in this disclosure do not indicate any order or importance, but are used to distinguish one configuration from another. In addition, unless otherwise specified, appropriate values ​​for "threshold," "predetermined value," etc., used in this disclosure may be determined by the developer's knowledge or by experiments using the cooking support system 10, automatic cooker 12, LLM server 14, information processing device 16, etc., as described in the embodiments.

[0117] <Note> Based on the above description of embodiments and modifications, the following technologies are disclosed. [Technology 1-1] An information processing method in a computer, comprising: acquiring input information for a generating AI; acquiring a first response from the generating AI regarding a predetermined deliverable by inputting the input information into the generating AI; performing a first evaluation regarding the similarity between reference information regarding the deliverable and the first response; acquiring a second response from the generating AI regarding the deliverable according to the result of the first evaluation; performing a second evaluation regarding the quality of the second response using the generating AI; and outputting the second response according to the result of the second evaluation. According to this information processing method, since the second evaluation is performed according to the result of the first evaluation, the frequency of the second evaluation can be reduced, and the computational load on the computer can be reduced. In addition, it becomes easier to ensure the quality of the output second response. [Technology 1-2] The information processing method according to Technology 1-1, wherein the first evaluation includes an evaluation regarding the similarity between textual information contained in the reference information and textual information contained in the first response. According to this information processing method, the similarity between the reference information and the first answer can be evaluated with high accuracy. [Technical 1-3] The information processing method according to Technical 1-2, wherein the similarity is the similarity of vector information corresponding to the textual information contained in the reference information and the textual information contained in the first answer. According to this information processing method, the similarity between the textual information contained in the reference information and the textual information contained in the first answer can be evaluated with high accuracy. [Technical 1-4] The information processing method according to Technical 1-1, wherein if the similarity in the first evaluation is above a predetermined threshold, the second evaluation is performed, and if the similarity is below the threshold, the second evaluation is not performed and the second answer is not output. According to this information processing method, the frequency of the second evaluation can be reduced, and the computational load on the computer can be reduced. [Technical 1-5] The information processing method according to Technical 1-1, wherein the second evaluation includes an evaluation of whether or not the second answer contains errors. According to this information processing method, the quality of the second answer can be guaranteed from the perspective of whether or not the second answer contains errors.[Technical 1-6] The information processing method according to Technical 1-1, wherein the second evaluation includes determining whether the second answer satisfies one or more predetermined evaluation items. This information processing method ensures the quality of the second answer in terms of whether or not it contains errors. [Technical 1-7] The information processing method according to Technical 1-6, wherein in the second evaluation, the second answer is output if the second answer satisfies one or more of the evaluation items, and the second answer is not output if the second answer does not satisfy one or more of the evaluation items. This information processing method ensures the quality of the outputted second answer. [Technical 1-8] The information processing method according to any one of Technical 1-1 to 1-7, wherein the first answer is information about a recipe for making a dish desired by the user, and the second answer is information about a recipe for making the dish using predetermined cooking equipment. This information processing method can output information about a recipe for making any dish desired by the user using predetermined cooking equipment. [Technical 1-9] The information processing method according to Technical 1-8, wherein the reference information is information about a recipe for making a dish predetermined as a dish that can be made using the cooking equipment. According to this information processing method, information about a recipe for making any dish desired by the user using a predetermined cooking equipment can be output. By performing a first evaluation using information about a recipe for making a dish predetermined for the cooking equipment as reference information, those that are highly likely to be made with the cooking equipment can be moved to the second evaluation. [Technical 1-10] The information processing method according to Technical 1-8, wherein the first answer is information about a recipe for making a dish desired by the user without using the cooking equipment, the reference information is information about a recipe for making a dish predetermined as a dish that can be made using the cooking equipment without using the cooking equipment, and in the first evaluation, if the similarity between the first answer and the reference information is above a predetermined threshold, the generating AI is made to generate information about a recipe for making a dish desired by the user using the cooking equipment as the second answer.According to this information processing method, by comparing a first answer, which is general recipe information that does not use cooking equipment, with reference information, it is possible to accurately determine whether or not the dish desired by the user can be made using the cooking equipment, and it is possible to suppress the first generating AI from unnecessarily generating a second answer. [Technical 1-11] The information processing method described in Technical 1-1, wherein the second answer is the same as the first answer. [Technical 1-12] An information processing apparatus comprising: an input information acquisition unit that acquires input information for a generating AI; a first answer acquisition unit that acquires a first answer from the generating AI regarding a predetermined deliverable by inputting the input information to the generating AI; a first evaluation unit that performs a first evaluation regarding the similarity between reference information regarding the deliverable and the first answer; a second answer acquisition unit that acquires a second answer from the generating AI regarding the deliverable according to the result of the first evaluation; a second evaluation unit that performs a second evaluation regarding the quality of the second answer using the generating AI; and an output unit that outputs the second answer according to the result of the second evaluation. This information processing device produces the same effects as the information processing method of Technology 1-1. [Technology 1-13] A computer program for causing a computer to perform the following: acquire input information for a generating AI; input the input information into the generating AI to acquire a first response from the generating AI regarding a predetermined deliverable; perform a first evaluation regarding the similarity between reference information regarding the deliverable and the first response; acquire a second response from the generating AI regarding the deliverable according to the result of the first evaluation; perform a second evaluation regarding the quality of the second response using the generating AI; and output the second response according to the result of the second evaluation. This computer program produces the same effects as the information processing method of Technology 1-1.

[0118] [Technical 2-1] An information processing method in a computer, comprising: obtaining a first response from a generating AI based on predetermined input information; causing the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; obtaining a second response obtained by modifying the first response based on the results of the evaluation; and outputting the second response. The first response may correspond to the second response before modification in the embodiment (for example, the automatic cooking recipe initially obtained in S22 of Figure 3). The second response may also correspond to the modified second response in the embodiment (for example, the automatic cooking recipe newly generated in response to the modification instruction prompt in S28 of Figure 3). According to this information processing method, it is possible to obtain a highly accurate evaluation result based on a plurality of evaluation items for the first response from the generating AI, and to output a second response that appropriately modifies the first response based on the highly accurate evaluation result. [Technical 2-2] The information processing method according to Technical 2-1, wherein the plurality of evaluation items include requirements that the first response should satisfy. According to this information processing method, highly accurate evaluation results can be obtained from the perspective of whether or not the first answer includes the requirements that it should satisfy. [Technology 2-3] The information processing method according to Technology 2-1, wherein when it is determined that the first answer or the second answer satisfies the multiple evaluation items, the answer is output. According to this information processing method, an appropriate answer that satisfies multiple evaluation items can be output. [Technology 2-4] The information processing method according to Technology 2-1, wherein the generating AI is further instructed to output information regarding the basis for the evaluation result of the first answer based on the multiple evaluation items. According to this information processing method, the accuracy of the correction of the first answer can be improved, or in other words, it becomes easier to appropriately correct the first answer. [Technology 2-5] The information processing method according to Technology 2-1, wherein when it is determined that the first answer does not satisfy at least one of the multiple evaluation items, the generating AI is instructed to input information regarding the result of the evaluation to generate the second answer. According to this information processing method, the generating AI can be used to obtain a second answer that appropriately corrects the first answer based on the evaluation result.[Technology 2-6] An information processing method according to Technology 2-5, wherein the generating AI is further instructed to output information regarding the basis for the evaluation result of the first answer based on the multiple evaluation items, and the generating AI is further instructed to input the information regarding the basis to generate the second answer. According to this information processing method, a second answer can be obtained in which the first answer has been appropriately modified based on the basis for the evaluation result. [Technology 2-7] An information processing method according to Technology 2-1, wherein a score is calculated according to the number of evaluation items that the first answer satisfies among the multiple evaluation items, and the generating AI is instructed to generate the second answer if the score is less than a predetermined threshold. According to this information processing method, it is possible to flexibly set whether or not to generate a second answer using a score according to the number of evaluation items that the first answer satisfies. [Technology 2-8] An information processing method according to Technology 2-1, wherein the generating AI is further instructed to perform an evaluation of the second answer based on the multiple evaluation items. According to this information processing method, a highly accurate evaluation result based on multiple evaluation items can be obtained for the second answer. [Technology 2-9] An information processing method according to Technology 2-8, wherein if the second answer does not satisfy at least one of the multiple evaluation items, the acquisition and evaluation of the second answer is repeatedly performed. This information processing method makes it easier to ensure the quality of the output second answer. [Technology 2-10] An information processing method according to Technology 2-9, wherein if the repeated acquisition and evaluation of the second answer satisfies a predetermined termination condition, the acquisition and evaluation of the second answer is terminated and the second answer is not output. This information processing method makes it possible to prevent excessive repetition of the acquisition and evaluation of the second answer. [Technology 2-11] An information processing method according to Technology 2-1, wherein multiple generating AIs perform evaluations for each of the multiple evaluation items for the first answer, or one generating AI performs evaluations for each of the multiple evaluation items multiple times. This information processing method makes it possible to improve the accuracy of the evaluation for each evaluation item. [Technical 2-12] The information processing method according to Technical 2-11, further comprising aggregating the results of evaluations performed by the multiple generating AIs, or the results of evaluations performed multiple times by the generating AI of the first generation.According to this information processing method, the accuracy of the evaluation for each evaluation item can be improved. [Technical 2-13] The first answer and the second answer are information relating to a recipe for making a dish using a predetermined cooking appliance, according to the information processing method described in any of Technical 2-1 to 2-12. According to this information processing method, accurate information relating to a recipe for making any dish desired by the user using a predetermined cooking appliance can be output. [Technical 2-14] The information processing method described in Technical 2-13, wherein the plurality of evaluation items include at least one of an evaluation item relating to the ingredients for making the dish and an evaluation item relating to the cooking process for making the dish. According to this information processing method, the accuracy of the recipe information for making any dish desired by the user can be improved from the standpoint of the ingredients and cooking process for making the dish. [Technical 2-15] The information processing method described in Technical 2-13, wherein the plurality of evaluation items include at least one of an evaluation item relating to the manner of use of the cooking appliance and an evaluation item relating to the cooking process using the cooking appliance. According to this information processing method, the accuracy of the recipe information for making the dish can be improved from the standpoint of the manner of use of the cooking appliance and the cooking process using the cooking appliance. [Technical 2-16] An information processing device comprising: a first response acquisition unit that acquires a first response from a generating AI based on predetermined input information; an evaluation unit that causes the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; a second response acquisition unit that acquires a second response obtained by modifying the first response based on the results of the evaluation; and an output unit that outputs the second response. This information processing device produces the same effects as the information processing method of Technical 2-1. [Technical 2-17] A computer program that causes a computer to perform: acquire a first response from a generating AI based on predetermined input information; cause the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; acquire a second response obtained by modifying the first response based on the results of the evaluation; and output the second response. This computer program produces the same effects as the information processing method of Technical 2-1.

[0119] The technology disclosed herein can be applied to information processing devices and information processing systems.

[0120] 10 Cooking support system, 12 Automatic cooker, 14 LLM server, 16 Information processing device, 40 Input information acquisition unit, 42 Prompt generation unit, 44 Response acquisition unit, 46 Similarity evaluation unit, 48 Quality evaluation unit, 50 Response output unit, 52 Response correction unit, 54 Embedding model update unit.

Claims

1. An information processing method in a computer, comprising: obtaining a first response from a generating AI based on predetermined input information; causing the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; obtaining a second response obtained by modifying the first response based on the results of the evaluation; and outputting the second response.

2. The information processing method according to claim 1, wherein the plurality of evaluation items include requirements that the first response must satisfy.

3. The information processing method according to claim 1, which outputs the first or second response if it is determined that the first or second response satisfies the multiple evaluation items.

4. The information processing method according to claim 1, further comprising causing the generating AI to output information regarding the basis for the evaluation result based on the plurality of evaluation items for the first answer.

5. The information processing method according to claim 1, wherein if the first response is evaluated as not satisfying at least one of the plurality of evaluation items, the second response is generated by inputting information regarding the evaluation result into the generating AI.

6. The information processing method according to claim 5, further comprising: causing the generating AI to output information regarding the basis for the evaluation result based on the plurality of evaluation items for the first answer; and further inputting the information regarding the basis into the generating AI to generate the second answer.

7. The information processing method according to claim 1, comprising: calculating a score corresponding to the number of evaluation items among the plurality of evaluation items that the first response satisfies; and causing the generating AI to generate the second response if the score is less than a predetermined threshold.

8. The information processing method according to claim 1, further comprising causing the generating AI to perform an evaluation of the second answer based on the plurality of evaluation items.

9. The information processing method according to claim 8, wherein if the second response does not satisfy at least one of the multiple evaluation items, the acquisition and evaluation of the second response are repeatedly performed.

10. The information processing method according to claim 9, wherein if the repetition of obtaining and evaluating the second answer satisfies a predetermined termination condition, the acquisition and evaluation of the second answer is terminated and the second answer is not output.

11. The information processing method according to claim 1, wherein multiple generating AIs perform evaluations of each of the multiple evaluation items for the first answer, or one generating AI performs evaluations of each of the multiple evaluation items multiple times.

12. The information processing method according to claim 11, further comprising aggregating the results of evaluations performed by the plurality of generating AIs, or the results of evaluations performed multiple times by the generating AI of claim 1.

13. The information processing method according to any one of claims 1 to 12, wherein the first and second answers are information relating to a recipe for preparing a dish using a predetermined cooking appliance.

14. The information processing method according to claim 13, wherein the plurality of evaluation items include at least one of evaluation items relating to the ingredients for making the dish and evaluation items relating to the cooking process for making the dish.

15. The information processing method according to claim 13, wherein the plurality of evaluation items include at least one of evaluation items relating to the manner of use of the cooking equipment and evaluation items relating to the cooking process using the cooking equipment.

16. An information processing device comprising: a first response acquisition unit that acquires a first response from a generating AI based on predetermined input information; an evaluation unit that causes the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; a second response acquisition unit that acquires a second response obtained by modifying the first response based on the results of the evaluation; and an output unit that outputs the second response.

17. A computer program for causing a computer to perform the following actions: obtain a first response from a generating AI based on predetermined input information; cause the generating AI to perform an evaluation of the first response based on a plurality of evaluation items; obtain a second response by modifying the first response based on the results of the evaluation; and output the second response.

Images