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

Abstract

A user database (151) stores a model and user information corresponding to each of a plurality of users. A prompt editing unit (104) creates a prompt for generating a response to an utterance entered by a first user, on the basis of the utterance and information about the first user acquired from the user database (151). A training data creation unit (106) creates training data for training a model corresponding to the first user from a history of combinations of the first user's utterances and responses generated using the prompt in response to the first user's utterances. When the model corresponding to the first user is stored in the user database (151), a user information management unit (102) makes settings so that a response is generated using the prompt and the model corresponding to the first user, whereas when the model corresponding to the first user is not stored in the user database (151), the user information management unit (102) makes settings so that a response is generated using the prompt and a model common to the plurality of users.

Description

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

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

[0002] The recent rise of LLMs (Large Language Models) has made it easier to create dialogue systems capable of natural dialogue based on past conversations (see, for example, Non-Patent Documents 1-3).

[0003] Also, a method is known in which an LLM is specialized for a user by using a dialogue history with the user (see, for example, Non-Patent Document 4).

[0004] BlenderBot 3: a deployed conversational agent that continually learns to responsibly engage, [online], [Retrieved April 10, 2024], Internet (https: / / blenderbot.ai / ). Turn your enterprise data into production-ready LLM applications, [online], [Retrieved April 10, 2024], Internet (https: / / www.llamaindex.ai / ). MemGPT, [online], [Retrieved April 10, 2024], Internet (https: / / memgpt.ai / ). Fine-tuning Llama 2 with DPOs, [online], [Retrieved April 10, 2024], Internet (https: / / note.com / npaka / n / nfe7391a1d28d). LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS, [online], [Retrieved April 10, 2024], Internet (https: / / arxiv.org / pdf / 2106.09685.pdf). Development of personalized dialogue technology that efficiently reproduces individual speech in large-scale language models. [online], [Retrieved April 10, 2024], Internet (https: / / group.ntt / jp / newsrelease / 2024 / 01 / 17 / 240117a.html). Direct Preference Optimization: Your Language Model is Secretly a Reward Model, [online], [Retrieved April 10, 2024], Internet (https: / / arxiv.org / pdf / 2305.18290.pdf).

[0005] However, conventional dialogue systems have a problem in that it is sometimes difficult to have a dialogue in a manner specific to the user.

[0006] For example, the dialogue system described in Non-Patent Document 1 generates a response to a user by referring to the dialogue system's own past dialogue history and external knowledge. Therefore, even if a dialogue system has been conversing with a single user for a long period of time, it does not necessarily understand the user's personality and values ​​like a human being and generate a response based on the content of that understanding.

[0007] Furthermore, in the method described in Non-Patent Document 4, it is difficult to generate utterances appropriate for the user until sufficient data is collected.

[0008] In order to solve the above-mentioned problems and achieve the object, an information processing device is characterized by having: a user DB that stores models and user information corresponding to each of a plurality of users; a prompt editing unit that creates a prompt for generating a response to an utterance based on an utterance input by a first user and information about the first user obtained from the user DB; a user information management unit that, if a model corresponding to the first user is stored in the user DB, sets the response to be generated using the prompt and the model corresponding to the first user, and, if a model corresponding to the first user is not stored in the user DB, sets the response to be generated using the prompt and a model common to the plurality of users; and a learning data creation unit that creates learning data for training the model corresponding to the first user from a history of combinations of utterances of the first user and responses generated using the prompt in response to the utterances of the first user.

[0009] According to the present invention, it is possible to have a dialogue in a manner that is specific to the user.

[0010] FIG. 1 is a diagram illustrating an example of a configuration of a dialogue device according to a first embodiment. FIG. 2 is a flowchart illustrating a processing flow of the dialogue device according to the first embodiment. FIG. 3 is a flowchart illustrating a flow of a start process. FIG. 4 is a diagram illustrating an example of a prompt for reproducing an agent. FIG. 5 is a diagram illustrating an example of a prompt for reproducing an agent. FIG. 6 is a diagram illustrating an example of a prompt for estimating emotions. FIG. 7 is a diagram illustrating an example of a prompt including a dialogue history. FIG. 8 is a diagram illustrating an example of a prompt for updating a profile. FIG. 9 is a diagram illustrating an example of a prompt including a dialogue history. FIG. 10 is a diagram illustrating an example of a prompt for generating a response. FIG. 11 is a diagram illustrating an example of a prompt for referring to a past dialogue history. FIG. 12 is a flowchart illustrating a flow of a termination process. FIG. 13 is a diagram illustrating an example of learning data. FIG. 14 is a diagram illustrating an example of a prompt for generating a user's utterance. FIG. 15 is a diagram illustrating an example of a computer that executes a dialogue program.

[0011] Hereinafter, embodiments of an information processing device, an information processing method, and an information processing program according to the present application will be described in detail with reference to the accompanying drawings. Note that the present invention is not limited to the embodiments described below. Note that the interactive device is an example of an information processing device.

[0012] [Configuration of First Embodiment] First, the configuration of the dialogue device according to the first embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the configuration of the dialogue device according to the first embodiment.

[0013] The dialogue device 10 uses LLM to engage in dialogue with a user. The dialogue device 10 receives an utterance from the user as input. The dialogue device 10 then generates an utterance in response to the user's utterance and outputs the generated utterance. The utterance output by the dialogue device 10 may be called a response. Both the utterance and the response are data that can be expressed in text. The utterance and the response may also be audio data.

[0014] The dialogue device 10 generates a response tailored to the user by selectively using two techniques: prompt engineering and model learning. Prompt engineering is a technique for creating a prompt that corresponds to a desired response, which is a prompt to be input into an LLM. Model learning is a technique for learning (including re-learning or additional learning) a model that includes an LLM. The model to be learned in model learning is not limited to the LLM itself, but may also be a model that works with the LLM. Note that the model that includes the LLM in this embodiment may also be referred to as a machine learning model.

[0015] The terminal 20 is a terminal used by a user. The terminal 20 is a personal computer, a smartphone, or the like. The terminal 20 has a client application 21 for receiving a dialogue service provided by the dialogue device 10. The client application 21 may be a web browser. The dialogue device 10 may be a server connected to the terminal 20 via a network.

[0016] The dialogue device 10 includes an agent information management unit 101, a user information management unit 102, an emotion estimation unit 103, a prompt editing unit 104, an utterance generation unit 105, a learning data creation unit 106, a model learning unit 107, a model usability determination unit 108, a user DB 151, and an agent DB 152.

[0017] The flow of processing will be described along with details of each unit of the dialogue device 10 using the flowchart of Fig. 2. Fig. 2 is a flowchart showing the flow of processing of the dialogue device according to the first embodiment.

[0018] As shown in Fig. 2, first, the interactive device 10 performs a start process (step S11). The start process will be described in detail with reference to Fig. 3. Fig. 3 is a flowchart showing the flow of the start process.

[0019] The start process begins when the user inputs a user ID and password via the client application 21. If the correct combination of user ID and password is input, the client application 21 displays a list of agents and accepts the selection of an agent. The client application 21 transmits the user ID and the agent ID of the selected agent to the interactive device 10.

[0020] 3, the interactive device 10 receives input of a user ID and an agent ID (step S1101). Next, the agent information management unit 101 acquires a prompt for reproducing an agent based on the agent ID (step S1102). Note that if the user does not specify an agent, step S1102 may be skipped. The prompt for reproducing an agent is associated with the agent ID and stored in the agent DB 152.

[0021] 4 and 5 are diagrams showing examples of prompts for recreating an agent. As shown in Fig. 4, the prompt for recreating an agent may be a combination of a command statement and a profile. Also, as shown in Fig. 5, the prompt for recreating an agent may instruct the user to become a fictional character appearing in content such as an anime.

[0022] The user information management unit 102 also acquires the dialogue history between the user and the agent from the DB based on the user ID and the agent ID (step S1103). The dialogue history between the user and the agent is associated with the user ID and the agent ID and stored in the user DB 151. The user information management unit 102 may acquire the entire content of past dialogues between the user and the agent (dialogue device 10) as the dialogue history, or may acquire a sentence summarizing the content of the dialogue.

[0023] The user information management unit 102 also acquires a user profile based on the user ID (step S1104). The user profile is associated with the user ID and stored in the user DB 151. The profile is information indicating the user's attributes and preferences, such as age, sex, blood type, highest level of education, hometown, occupation, hobbies, and favorite foods.

[0024] Here, a model may be created for each user. If a model for the user exists (step S1105, Yes), the user information management unit 102 acquires the model for the user from a DB (user DB 151) based on the user ID (step S1106). The model for each user is created in the termination process described below and stored in the user DB 151. If a model for the user does not exist (step S1105, No), the user information management unit 102 proceeds to step S1106. The user DB 151 stores user models, i.e., models corresponding to each of multiple users.

[0025] Here, the model for each user is not the LLM itself, but the LoRA (Low-Rank Adaptation) used in combination with the LLM (see, for example, Non-Patent Documents 5 and 6). In other words, the parameters of the LLM are fixed, and the LoRA part is the subject of user-specific learning. This can be said to be a model for adapting the LLM to the user.

[0026] If a model corresponding to the first user is stored in the user DB 151, i.e., if a LoRA corresponding to the first user exists in the user DB 151, the user information management unit 102 sets the unit 102 to generate a response using a prompt and a model corresponding to the first user (e.g., a combination of LLM and LoRA). On the other hand, if a model corresponding to the first user is not stored in the user information management unit 102, the user information management unit 102 sets the unit 102 to generate a response using a prompt and a model common to multiple users (e.g., LLM). The utterance generation unit 105 generates a response based on the settings of the user information management unit 102.

[0027] 2, the feeling estimation unit 103 receives an input of a user's utterance (step S12). The feeling estimation unit 103 estimates the user's emotion based on the utterance (step S13).

[0028] (Estimating Emotions Using an Emotion Estimation Model) The emotion estimation unit 103 can estimate emotions using an emotion estimation model by the method described in Reference 1. The emotion estimation model has been trained to classify a pair of utterances from two speakers into one of the emotions contained in one of the utterances (five types of emotions: joy, anger, sadness, happiness, and neutral).

[0029] Reference 1: BERT base Japanese (unidic-lite with whole word masking, CC-100 and jawiki-20230102) (URL: https: / / huggingface.co / tohoku-nlp / bert-base-japanese-v3)

[0030] The input to the emotion estimation model is expressed as "[CLS] A's utterance [SEP] B's utterance [SEP]". In response to this, the emotion estimation model outputs one of the above five types as a classification result. In this embodiment, the emotion estimation unit 103 inputs a pair of an utterance from the dialogue device 10 and an utterance from the user to the trained emotion estimation model, and estimates the emotion.

[0031] (Estimating Emotions Using LLM) The emotion estimation unit 103 can estimate emotions using LLM. In this case, the emotion estimation unit 103 inputs a prompt of a command statement as shown in Fig. 6 into the LLM together with a prompt of a dialogue history as shown in Fig. 7 .

[0032] FIG. 6 is a diagram showing an example of a prompt for estimating emotions. The prompt shown in FIG. 6 allows the emotion estimation unit 103 to instruct the LLM to evaluate five types of emotions, namely, "joy," "anger," "sadness," "pleasure," and "neutral," on a five-point scale and to describe the reason for the evaluation. FIG. 7 is a diagram showing an example of a prompt including a dialogue history. When estimating emotions, the emotion estimation unit 103 causes the LLM to refer to the prompt in the dialogue history shown in FIG. 7.

[0033] The user information management unit 102 adds the utterance and the estimated emotion to the dialogue history (step S14). Then, the user information management unit 102 determines whether the user's profile needs to be updated based on the utterance (step S15). If the profile needs to be updated (step S16, Yes), the user information management unit 102 updates the user's profile (step S17). If the profile does not need to be updated (step S16, No), the user information management unit 102 proceeds to step S18.

[0034] The user information management unit 102 may input the user's current profile and interaction history to the LLM, determine whether the profile needs to be updated, and update the profile if necessary. The user information management unit 102 may also update the profile stored in the user DB 151 based on the output of the LLM.

[0035] The user information management unit 102 inputs a command statement prompt such as that shown in Fig. 8 into the LLM along with a dialogue history prompt such as that shown in Fig. 9. For example, in this case, the LLM updates the "Place of Residence" item in the user's profile from "Tokyo" to "Kyoto Prefecture." Fig. 8 is a diagram showing an example of a prompt for updating a profile. Fig. 9 is a diagram showing an example of a prompt including a dialogue history.

[0036] 2, the prompt editing unit 104 creates a prompt based on the dialogue history (step S18). At this time, the prompt editing unit 104 creates (or edits) a prompt that instructs the user to reply based on the user's inner thoughts and past dialogues, based on the dialogue history to which the user's emotions have been added, received from the user information management unit 102.

[0037] For example, the prompt editing unit 104 creates a prompt that instructs the user to generate an utterance that will make the user happy or enjoyable. In this case, the prompt editing unit 104 adds a part or all of the past dialogue history to the prompt and creates a prompt that instructs the user to refer to the added dialogue history.

[0038] An example of a prompt created by the prompt editing unit 104 is shown in Fig. 10. Fig. 10 shows an example of a prompt for generating a reply. As shown in Fig. 10, the dialogue history includes an estimated emotion of the user (corresponding to A).

[0039] The prompt editing unit 104 may also use Retrieval Augmented Generation (RAG) (see Non-Patent Document 2 and Reference 2) to have the LLM refer to past dialogue history. Fig. 11 is a diagram showing an example of a prompt for referencing past dialogue history. RAG is a technique for having a machine learning model refer to external information.

[0040] Reference 2: Introduction to LLAMA-INDEX (URL: https: / / qiita.com / LyW / items / ff5c222a3a7e0a1610d6)

[0041] 2 , the utterance generation unit 105 inputs the prompt into the model to generate a response (step S19). The user information management unit 102 adds the response to the dialogue history (step S20). The utterance generation unit 105 then outputs the response (step S21). The response is output via the client application 21 of the terminal 20. The user information management unit 102 then updates the user DB 151 with the dialogue history to which the response has been added.

[0042] If the dialogue has ended (step S22, Yes), the dialogue device 10 performs an end process (step S23). If the dialogue has not ended (step S22, No), the dialogue device 10 returns to step S12. The dialogue device 10 may determine whether the dialogue has ended based on whether the user has explicitly performed an operation to end the dialogue (for example, pressing an end button).

[0043] 12 is a flowchart showing the flow of the termination process. As shown in FIG. 12, the learning data creation unit 106 acquires past learning data based on the user ID and agent ID (step S2301). The past learning data was created based on the dialogue history in a previous termination process and is stored in the user DB 151.

[0044] Next, the learning data creation unit 106 creates learning data based on the dialogue history and adds it to the acquired learning data (step S2302). In other words, the learning data creation unit 106 updates the past learning data to the latest one using the latest dialogue history.

[0045] 13 is a diagram showing an example of training data. The training data is a combination of a prompt for each utterance, an utterance for which the user's emotion is estimated to be "joy" or "happiness," and an utterance for which the user's emotion is estimated to be "anger" or "sadness." In other words, the training data is data in which utterances, emotions, and prompts are associated with each other.

[0046] For example, suppose the utterance generated by the utterance generation unit 105 based on a prompt including the imperative sentence shown in FIG. 13 is "Wow, that sounds interesting. Shall we go see it together sometime?" Furthermore, suppose the user's utterance in response to this utterance is "That's great! It's been a long time since we last enjoyed a movie together." In this case, the emotion "relaxed" is estimated from the user's utterance. Therefore, the actual utterance "That's great! It's been a long time since we last enjoyed a movie together." is stored as the utterance estimated as "joy" or "relaxed" in the learning data.

[0047] On the other hand, there is no utterance in which the user's emotion is estimated to be "anger" or "sadness." Therefore, as shown in FIG. 14 , the training data creation unit 106 inputs a prompt for generating a user's utterance into the LLM, and generates an utterance (fictitious utterance) in which the user's emotion is estimated to be "anger" or "sadness." FIG. 14 is a diagram showing an example of a prompt for generating a user's utterance. Then, the utterance generated by the LLM is stored in the training data as an utterance inferred to be "anger" or "sadness."

[0048] If a model for the user (e.g., LoRA) exists (step S2303, Yes), the model learning unit 107 updates the model for the user using the training data (step S2304). On the other hand, if a model for the user does not exist (step S2303, No), the model learning unit 107 creates a model for the user using the training data (step S2305).

[0049] For example, the model learning unit 107 creates a LoRA for the user using DPO (Direct Preference Optimization) described in Non-Patent Document 7.

[0050] The model updated in step S2304 or the model created in step S2305 is called a new model. The model usability determination unit 108 determines whether the model for the user, i.e., the new model, is usable (step S2306). If the new model is usable (step S2307, Yes), the model usability determination unit 108 updates the model in the user DB 151 with the new model (step S2308). On the other hand, if the new model is not usable (step S2307, No), the model usability determination unit 108 does not update the model in the user DB 151. In this case, the new model may be discarded.

[0051] In this way, the model usability determination unit 108 determines whether or not the model is usable when training of the model corresponding to the first user managed by the user information management unit 102 is performed using the training data created by the training data creation unit 106. This makes it possible to prevent a decrease in accuracy due to model updates.

[0052] For example, the model usability determination unit 108 determines that the new model is usable if the rewards / margins (the average difference between the selected rewards and the corresponding rejected rewards) output in the DPO is greater than a threshold value (e.g., 0.6).

[0053] [Effects of the First Embodiment] As described above, when a model corresponding to the first user is stored in the user DB 151, the user information management unit 102 is configured to generate a response using a prompt and the model corresponding to the first user, and when a model corresponding to the first user is not stored in the user information management unit 102, the user information management unit 102 is configured to generate a response using a prompt and a model common to multiple users. Furthermore, the utterance generation unit 105 generates a response based on the settings of the user information management unit 102.

[0054] This enables dynamic model creation and updating using learning data created during a dialogue. As a result, the dialogue device 10 can gradually understand the dialogue partner (acquire a model adapted to the user) while collecting data about the dialogue partner by dynamically switching between prompt engineering and model learning. Furthermore, the dialogue device 10 generates a response tailored to the user by selectively using the two techniques of prompt engineering and model learning.

[0055] [System Configuration, etc.] The components of each device shown in the figure are conceptual functional units and do not necessarily have to be physically configured as shown. In other words, the specific form of distribution and integration of each device is not limited to that shown in the figure, and all or part of the devices can be functionally or physically distributed or integrated in any unit depending on various loads, usage conditions, etc. Furthermore, all or any part of the processing functions performed by each device can be realized by a CPU (Central Processing Unit) and a program analyzed and executed by the CPU, or can be realized as hardware using wired logic. The program may be executed not only by the CPU but also by other processors such as a GPU.

[0056] Furthermore, among the processes described in this embodiment, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically using a known method.In addition, the information including the processing procedures, control procedures, specific names, various data and parameters shown in the above documents and drawings can be changed as desired unless otherwise specified.

[0057] [Program] In one embodiment, the dialogue device 10 can be implemented by installing a dialogue program that executes the above-described learning process as package software or online software on a desired computer. For example, by executing the dialogue program on an information processing device, the information processing device can function as the dialogue device 10. The information processing device referred to here includes desktop and notebook personal computers. Other information processing devices also include mobile communication terminals such as smartphones, mobile phones, and PHS (Personal Handyphone Systems), as well as slate terminals such as PDAs (Personal Digital Assistants).

[0058] The dialogue device 10 may also be implemented as a server device that provides services related to the dialogue processing to a client terminal device used by the user. For example, the server device may be implemented as a server device that receives a user's utterance as input and outputs a response. In this case, the server device may be implemented as a web server or as a cloud that provides services related to the dialogue processing by outsourcing.

[0059] 15 is a diagram showing an example of a computer that executes an interactive program. The computer 1000 includes, for example, a memory 1010 and a CPU 1020. The computer 1000 also includes a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. These components are connected by a bus 1080.

[0060] The memory 1010 includes a read-only memory (ROM) 1011 and a random access memory (RAM) 1012. The ROM 1011 stores a boot program such as a basic input / output system (BIOS). The hard disk drive interface 1030 is connected to a hard disk drive 1090. The disk drive interface 1040 is connected to a disk drive 1100. A removable storage medium such as a magnetic disk or optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to a mouse 1110 and a keyboard 1120, for example. The video adapter 1060 is connected to a display 1130, for example.

[0061] The hard disk drive 1090 stores, for example, an OS 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process of the interaction device 10 is implemented as a program module 1093 in which computer-executable code is written. The program module 1093 is stored, for example, in the hard disk drive 1090. For example, the program module 1093 for executing processes similar to those of the functional configuration of the interaction device 10 is stored in the hard disk drive 1090. Note that the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).

[0062] Furthermore, setting data used in the processing of the above-described embodiment is stored as program data 1094, for example, in the memory 1010 or the hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 or the program data 1094 stored in the memory 1010 or the hard disk drive 1090 into the RAM 1012 as necessary, and executes the processing of the above-described embodiment.

[0063] The program module 1093 and program data 1094 may not necessarily be stored in the hard disk drive 1090, but may also be stored in a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and program data 1094 may be stored in another computer connected via a network (such as a local area network (LAN) or a wide area network (WAN)). The program module 1093 and program data 1094 may then be read by the CPU 1020 from the other computer via the network interface 1070.

[0064] REFERENCE SIGNS LIST 10 Dialogue device 101 Agent information management unit 102 User information management unit 103 Emotion estimation unit 104 Prompt editing unit 105 Utterance generation unit 106 Learning data creation unit 107 Model learning unit 108 Model usability determination unit 151 User DB 152 Agent DB

Claims

1. An information processing device comprising: a user DB that stores models and user information corresponding to each of a plurality of users; a prompt editing unit that creates a prompt for generating a response to an utterance based on an utterance input by a first user and information about the first user obtained from the user DB; a user information management unit that, if a model corresponding to the first user is stored in the user DB, sets the system to generate a response using the prompt and the model corresponding to the first user, and, if a model corresponding to the first user is not stored in the user DB, sets the system to generate a response using the prompt and a model common to the plurality of users; and a learning data creation unit that creates learning data for training the model corresponding to the first user from a history of combinations of utterances of the first user and responses generated using the prompt in response to the utterances of the first user.

2. The information processing device according to claim 1, further comprising: an utterance generation unit that generates a response based on the settings of said user information management unit.

3. The information processing device according to claim 1, further comprising an emotion estimation unit that estimates the emotion of the first user based on an utterance of the first user, and wherein the prompt editing unit creates a prompt for generating the response based on the utterance of the first user associated with the emotion estimated by the emotion estimation unit.

4. The information processing device according to claim 3, characterized in that the learning data creation unit creates the learning data including utterances of the first user associated with emotions estimated by the emotion estimation unit.

5. The information processing device described in claim 4, further comprising a model usability determination unit that determines whether the model is usable when learning of the model corresponding to the first user managed by the user information management unit is performed using learning data created by the learning data creation unit.

6. An information processing method executed by an information processing device, comprising: a prompt editing process for creating a prompt for generating a response to an utterance input by a first user based on the utterance and information of the first user obtained from a user DB that stores models and user information corresponding to each of a plurality of users; a user information management process for, if a model corresponding to the first user is stored in the user DB, setting to generate a response using the prompt and the model corresponding to the first user, and, if a model corresponding to the first user is not stored in the user DB, setting to generate a response using the prompt and a model common to the plurality of users; and a learning data creation process for creating learning data for training a model corresponding to the first user from a history of combinations of the utterances of the first user and responses generated using the prompt in response to the utterances of the first user.

7. An information processing program that causes a computer to execute the following steps: a prompt editing step that creates a prompt for generating a response to an utterance based on an utterance input by a first user and information about the first user obtained from a user DB that stores models and user information corresponding to each of multiple users; a user information management step that, if a model corresponding to the first user is stored in the user DB, sets the system to generate a response using the prompt and the model corresponding to the first user, and, if a model corresponding to the first user is not stored in the user DB, sets the system to generate a response using the prompt and a model common to the multiple users; and a learning data creation step that creates learning data for training a model corresponding to the first user from a history of combinations of the first user's utterances and responses generated using the prompt in response to the first user's utterances.

Images