Training device, training method, information provision device, information provision method, and storage medium

Abstract

Provided is an information providing device, in which a directive acquisition means acquires a directive from a user. This response creation means uses an AI model including a storage device and a trained generative AI trained on the basis of information of a target to be trained to create a response to the directive. The determination means determines whether the response includes information that is conceptually close to restricted disclosure information. When the response includes information that is conceptually close to the restricted disclosure information, a response correction means corrects the response by inputting into the AI model a directive for instructing to conceal the conceptually close information. In addition, in the training device, a storage means stores experience data which associates an experience of the subject, a concept vector of the experience, and metadata including a classification tag for classifying the experience. A training means estimates a result for the experience on the basis of the experience, and trains the generative AI, using a pair of the experience and the result for the experience, as training data. An output means outputs a set of the experience data and the trained generative AI.

Description

Learning Device, Learning Method, Information Providing Device, Information Providing Method, and Recording Medium 【0001】 This disclosure relates to the learning and provision of AI. 【0002】 In recent years, various AI (Artificial Intelligence) services have been utilized in business, such as question-and-answer and document creation. For example, Patent Document 1 discloses a method of providing a model that outputs an answer to a user's question. Also, as an AI service, a service that substitutes for a person's work is also known. 【0003】 Japanese Unexamined Patent Application Publication No. 2023-076413 【0004】 In the learning of AI, there may be information that is not desired to be known by a third party in the process of collecting learning data. If an AI learned with such data is provided to a user, there is a possibility that the information that is not desired to be known will be disclosed to the service user. 【0005】 One object of this disclosure is to provide an information providing device capable of restricting the disclosure of information to users of an AI service. 【0006】 From one aspect of this disclosure, the information providing device includes: an instruction obtaining means for obtaining an instruction sentence from a user; a learned generation AI learned based on information to be learned, a storage device, and an answer creating means for creating an answer to the instruction sentence using an AI model including them; a determination means for determining whether information conceptually close to disclosure restriction information is included in the answer; and an answer correction means for correcting the answer by inputting an instruction sentence instructing to conceal the information conceptually close to the disclosure restriction information into the AI model when the information conceptually close to the disclosure restriction information is included in the answer. 【0007】In other respects of this disclosure, the information provision method includes obtaining an instruction from a user, creating a response to the instruction using an AI model that includes a trained generative AI that has learned based on the information to be learned and a storage device, determining whether the response contains information that is conceptually similar to the information that is restricted from disclosure, and if the response contains information that is conceptually similar to the information that is restricted from disclosure, modifying the response by inputting an instruction into the AI ​​model that instructs it to conceal the information that is conceptually similar. 【0008】 In yet another aspect of this disclosure, the recording medium records a program that causes a computer to perform a process of modifying the response by inputting an instruction into the AI ​​model that instructs the AI ​​model to conceal the information that is similar in concept to the information that is restricted from disclosure, if the response contains information that is similar in concept to the information that is restricted from disclosure. 【0009】 In one aspect of this disclosure, the learning device includes: a storage means for storing experience data that associates a subject's experience, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; a learning means for estimating results for the experience from the experience and training a generative AI using pairs of the experience and the results for the experience as training data; and an output means for outputting a set of the experience data and a trained generative AI. 【0010】 In other aspects of this disclosure, the learning method stores experience data that associates a subject's experience with a conceptual vector of the experience and metadata including classification tags for classifying the experience; estimates the results for the experience from the experience; trains a generative AI using pairs of the experience and the results for the experience as training data; and outputs a set of the experience data and the trained generative AI. 【0011】In yet another aspect of this disclosure, the recording medium stores experience data that associates a subject's experience with a conceptual vector of the experience and metadata including classification tags for classifying the experience; estimates the results for the experience from the experience; trains a generative AI using pairs of the experience and the results for the experience as training data; and records a program that causes a computer to perform the process of outputting a set of the experience data and the trained generative AI. 【0012】 This disclosure makes it possible to restrict the disclosure of information to users of AI services. 【0013】 This shows the overall configuration of the clone AI provision system. This is a block diagram showing the hardware configuration of the learning device related to this disclosure. This is a block diagram showing the functional configuration of the learning device related to this disclosure. This is a flowchart of the processing performed by the learning device related to this disclosure. This is a block diagram showing the hardware configuration of the information provision device related to this disclosure. This is a block diagram showing the functional configuration of the information provision device related to this disclosure. This is a flowchart of the processing performed by the information provision device related to this disclosure. This is an example of fine tuning. This is a block diagram showing the functional configuration of another information provision device related to this disclosure. This is a flowchart of the processing performed by another information provision device related to this disclosure. This is a block diagram showing the functional configuration of another learning device related to this disclosure. This is a flowchart of the processing performed by another learning device related to this disclosure. 【0014】 Preferred embodiments of this disclosure will be described below with reference to the drawings. 【0015】 <First Embodiment> [Overall Structure] This disclosure involves lending a clone AI that shares the experiences and knowledge of a certain person to another person. 【0016】Figure 1 shows the overall configuration of a clone AI provision system to which the learning device and information provision device related to this disclosure are applied. The clone AI provision system 1 includes terminal devices 10 for multiple subjects, an information provision device 20, and terminal devices 30 for multiple users. Terminal device 10 is an example of a learning device. A clone AI is an example of an AI model configured to be able to think, judge, and answer in a manner similar to that person by learning based on the experience and knowledge of that person. 【0017】 Terminal device 10 is a terminal device such as a smartphone, smart glasses, or smart speaker used by the target person who is the learning target of the clone AI, and communicates with information provision device 20 via a network such as the internet. Terminal device 10 is assumed to be equipped with a generating AI. 【0018】 Terminal device 10 accumulates the subject's experience and creates a clone AI of the subject. Specifically, the subject activates the microphone and camera of terminal device 10 at any time. Terminal device 10 accumulates the audio data collected by the microphone and the image data captured by the camera as the subject's experience. Then, terminal device 10 generates training data from the accumulated experience and trains the generating AI. Terminal device 10 transmits the set of the trained generating AI (knowledge) and the accumulated experience (experience) as a clone AI to the information providing device 20. This clone AI is mainly used to perform the subject's work on their behalf. 【0019】 The aforementioned arbitrary timings include, for example, the occurrence of events such as school events, company events, community events, and family events. The subject may keep the microphone and camera of the terminal device 10 activated at all times to accumulate experience. In addition, the terminal device 10 stores audio data collected by the microphone and image data captured by the camera as the subject's experience. Furthermore, the terminal device 10 may also provide the subject with a generating AI as an AI assistant and store the content of conversations and question-and-answer sessions between the subject and the generating AI as the subject's experience. 【0020】The information providing device 20 is composed of, for example, a server device and communicates with terminal devices 10 and 30 via a network such as the Internet. The information providing device 20 stores clone AIs of multiple subjects and provides the corresponding clone AI in response to the user's request. Specifically, when the information providing device 20 receives an instruction from the user's terminal device 30, it uses the clone AI to create a response to the instruction. Then, the information providing device 20 transmits the created response to the terminal device 30. 【0021】 Terminal device 30 is a terminal device such as a personal computer used by the user of the clone AI, and communicates with information provider device 20 via a network such as the Internet. Terminal device 30 sends instruction messages to the clone AI to information provider device 20 and receives responses to the instruction messages from information provider device 20. The instruction messages are created by the user and include, for example, inquiries about work or instructions for creating documents. 【0022】 [Terminal device] First, let's explain the terminal device 10. The terminal device 10 creates a clone AI of the target person. 【0023】 (Hardware Configuration) Figure 2 is a block diagram showing the hardware configuration of the terminal device 10 according to the first embodiment. As shown in the figure, the terminal device 10 includes an interface (I / F) 11, a processor 12, a memory 13, a recording medium 14, a database (DB) 15, an input unit 16, a display unit 17, a camera 18, and a microphone 19. 【0024】 I / F11 performs data input and output with external devices. Specifically, I / F11 transmits clone AI to the information providing device 20. 【0025】The processor 12 is a computer such as a CPU (Central Processing Unit) and controls the entire terminal device 10 by executing a pre-prepared program. The processor 12 may be a GPU (Graphics Processing Unit), DSP (Digital Signal Processor), MPU (Micro Processing Unit), FPU (Floating Point Number Processing Unit), PPU (Physics Processing Unit), TPU (Tensor Processing Unit), quantum processor, microcontroller, or a combination thereof. The processor 12 performs the learning process described later. 【0026】 Memory 13 consists of ROM (Read Only Memory), RAM (Random Access Memory), and the like. Memory 13 is also used as working memory while the processor 12 is executing various processes. 【0027】 The recording medium 14 is a non-volatile, non-temporary recording medium such as a disk-shaped recording medium or semiconductor memory, and is configured to be detachable from the terminal device 10. The recording medium 14 stores various programs that the processor 12 will execute. When the terminal device 10 performs various processes, the programs stored on the recording medium 14 are loaded into the memory 13 and executed by the processor 12. 【0028】 DB15 records the experience data described later. The input unit 16 is a touch panel or the like for the user to operate. The display unit 17 is a display that shows information based on the control of the processor 12, or a speaker that outputs sound. The camera 18 takes pictures of the surroundings and the target object. The microphone 19 collects the voice of the user and surrounding sounds. The terminal device 10 may also be equipped with GPS (Global Positioning System) or the like. 【0029】(Functional Configuration) Figure 3 is a block diagram showing the functional configuration of the terminal device 10 of the first embodiment. Functionally, the terminal device 10 comprises a data acquisition unit 101, a data storage unit 102, a learning data generation unit 103, a learning unit 104, and an output unit 105. 【0030】 The data storage unit 102 is implemented by the DB 15 shown in Figure 2. The data acquisition unit 101, the learning data generation unit 103, the learning unit 104, and the output unit 105 are all composed of the processor 12 shown in Figure 2. 【0031】 The data acquisition unit 101 acquires audio data collected by the microphone and image data captured by the camera as the subject's experience. 【0032】 The data acquisition unit 101 adds metadata such as date and time, location, and related parties to the acquired experience. Related parties are people or things that were involved with the subject. The data acquisition unit 101 may use existing image analysis methods to identify related parties who are visible in the image from the image data, or it may use existing voice analysis methods to identify related parties from the voice data. 【0033】 Furthermore, the data acquisition unit 101 converts the acquired experience into an embedding vector. For example, the data acquisition unit 101 can convert image data into an embedding vector using an embedding model such as ResNet (Residual Networks) or VGG. Alternatively, the data acquisition unit 101 can convert audio data into text data using existing speech recognition methods, and then convert the text data into an embedding vector using an embedding model such as BERT (Bidirectional Encoder Representations from Transformers), Universal Sentence Encoder, or Doc2Vec. The data acquisition unit 101 may also convert audio data into an embedding vector using an embedding model such as GRU (Gated Recurrent Unit) or LSTM (Long Short-Term Memory). The experience converted into an embedding vector will also be referred to as a "conceptual vector" below. 【0034】 The data acquisition unit 101 associates experience, conceptual vectors, and metadata and outputs them to the data storage unit 102. The data that associates experience, conceptual vectors, and metadata will also be referred to as "experience data" below. 【0035】 The data storage unit 102 stores the subject's experience data. 【0036】 The learning data generation unit 103 generates learning data for training the generated AI from the experience data stored in the data storage unit 102. The learning data generation unit 103 then outputs the generated learning data to the learning unit 104. 【0037】 Specifically, the learning data generation unit 103 generates learning data by associating what the subject actually experienced with the reactions of those involved to that experience and the subject's own actions (hereinafter also referred to as "results of the experience"). For example, if the subject takes a certain action during the experience, the learning data generation unit 103 estimates the reactions of those involved to that action and generates learning data by associating the subject's action with the reactions of those involved. Also, if the subject sees or hears something during the experience, the learning data generation unit 103 estimates the action the subject took at that time (i.e., their own action in response to the experience) and generates learning data by associating what the subject saw or heard with the action the subject took. 【0038】 More specifically, the learning data generation unit 103 generates learning data by estimating the causal relationships between multiple actions included in an experience and associating the causal action with the resulting action. For example, if, regarding a certain action A and a certain action B included in an experience, action B occurs immediately after action A in terms of time, the learning data generation unit 103 may estimate action A as the cause and action B as the result of action A. Alternatively, if action A and action B fit a general causal relationship, the learning data generation unit 103 may estimate that there is a causal relationship between action A and action B. A general causal relationship is a predictable causal relationship, such as "getting angry → worsening relationships with others" or "getting injured → seeking treatment." The learning data generation unit 103 generates learning data by associating action A and action B. 【0039】 Note that the learning data generation unit 103 can estimate the actions of the target person or related persons by detecting the postures and movements of the target person or related persons from a plurality of temporally continuous image data (i.e., video) using, for example, existing image analysis methods. In addition, the learning data generation unit 103 can estimate the actions and reactions of the target person or related persons by detecting the speech content and voice tone of the target person or related persons from voice data using existing speech recognition and speech analysis methods. 【0040】 Note that the method for generating the learning data shown above is an example and is not limited thereto. 【0041】 The learning unit 104 uses the learning data input from the learning data generation unit 103 to train the generation AI. The learning unit 104 outputs the trained generation AI to the output unit 105. The output unit 105 transmits a set of the trained generation AI and the experience data stored in the data storage unit 102 to the information providing device 20 as a clone AI. 【0042】 (Processing Flow) Next, the learning process of the clone AI by the terminal device 10 will be described. FIG. 4 is a flowchart of the learning process by the terminal device 10. This process is realized by the processor 12 shown in FIG. 2 executing a program prepared in advance and operating as each element shown in FIG. 3. 【0043】 First, the data acquisition unit 101 acquires the experience of the target person (step S101). Next, the data acquisition unit 101 outputs experience data in which the experience, the concept vector, and the metadata are associated to the data storage unit 102 (step S102). 【0044】 Next, the learning data generation unit 103 generates learning data for training the generation AI. Specifically, the learning data generation unit 103 estimates the result for the experience based on the experience of the target person (step S103). Then, the learning data generation unit 103 generates learning data by associating the experience and the result for the experience (step S104). 【0045】Next, the learning unit 104 learns the generation AI using the learning data input from the learning data generation unit 103 (step S105). The learning unit 104 outputs the learned generation AI to the output unit 105. Next, the output unit 105 transmits a set of the learned generation AI and the experience data stored in the data storage unit 102 to the information providing device 20 as a clone AI (step S106). Then, the process ends. 【0046】 [Information Providing Device] Next, the information providing device 20 will be described. The information providing device 20 provides the clone AI to the user. 【0047】 (Hardware Configuration) FIG. 5 is a block diagram showing the hardware configuration of the information providing device 20 according to the first embodiment. As shown in the figure, the information providing device 20 includes an interface (I / F) 21, a processor 22, a memory 23, a recording medium 24, and a database (DB) 25. 【0048】 The I / F 21 performs data input / output with an external device. Specifically, the I / F 21 receives the clone AI from the terminal device 10. Also, the I / F 21 receives an instruction sentence from the user's terminal device 30. 【0049】 The processor 22 is a computer such as a CPU (Central Processing Unit), and controls the entire information providing device 20 by executing a program prepared in advance. Note that the processor 22 may be a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating Point number Processor), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination thereof. The processor 22 executes the information providing process described later. 【0050】Memory 23 is composed of ROM (Read Only Memory), RAM (Random Access Memory), and the like. Memory 23 is also used as working memory while the processor 22 is executing various processes. 【0051】 The recording medium 24 is a non-volatile, non-temporary recording medium such as a disk-shaped recording medium or semiconductor memory, and is configured to be detachable from the information providing device 20. The recording medium 24 stores various programs that the processor 22 executes. When the information providing device 20 performs various processes, the programs stored in the recording medium 24 are loaded into the memory 23 and executed by the processor 22. 【0052】 DB25 stores the clone AI of each subject. DB25 also stores the disclosure restriction information dictionary, which will be described later. 【0053】 (Functional Configuration) Figure 6 is a block diagram showing the functional configuration of the information providing device 20 of the first embodiment. Functionally, the information providing device 20 comprises an instruction acquisition unit 201, a clone AI storage unit 202, an answer creation unit 203, and an output unit 204. 【0054】 The clone AI memory unit 202 is implemented by the DB 25 shown in Figure 5. The data acquisition unit 101, the instruction acquisition unit 201, the response creation unit 203, and the output unit 204 are all composed of the processor 22 shown in Figure 5. 【0055】 The user operates the terminal device 30 to specify which target person's clone AI to use. The user also operates the terminal device 30 to create an instruction message for the clone AI. The terminal device 30 transmits the specified target person (hereinafter also referred to as the "specified target person") and the instruction message to the information providing device 20. 【0056】 The instruction acquisition unit 201 acquires the designated person and instruction from the terminal device 30 and outputs them to the response creation unit 203. 【0057】The clone AI storage unit 202 stores the clone AI received from the terminal device 10 (i.e., a combination of the generated AI and the accumulated experience data). It should be assumed that there are multiple subjects and terminal devices 10, and the clone AI storage unit 202 stores a clone AI for each subject. 【0058】 The response generation unit 203 uses a clone AI of the designated person to create a response to the instruction. At this time, the response generation unit 203 may also use the so-called RAG (Retrievable-Augmented Generation) technique, which involves searching for information from accumulated experience data and having the generating AI create a response based on the searched information. 【0059】 Specifically, in the RAG search phase, the response generation unit 203 extracts information related to the instruction from the clone AI memory unit 202. For example, first, the response generation unit 203 converts the instruction into a vector using an embedding model and calculates the similarity between the instruction vector and the conceptual vector of each experience data. The conceptual vector of each experience data belongs to the designated person and is obtained from the clone AI memory unit 202. Cosine similarity or Euclidean distance can be used as an indicator of similarity. Next, the response generation unit 203 selects conceptual vectors related to the instruction vector based on the calculated similarity. For example, the response generation unit 203 may select conceptual vectors whose similarity to the instruction vector is above a predetermined threshold, or it may select a predetermined number of conceptual vectors in descending order of similarity to the instruction vector. The response generation unit 203 extracts the experiences corresponding to the selected conceptual vectors as information related to the instruction. 【0060】 In the RAG generation phase, the response creation unit 203 attaches the extracted experience to the instruction text and inputs it into the generation AI of the specified target. The response creation unit 203 then obtains the response to the instruction text from the generation AI. 【0061】In the above example, the response generation unit 203 uses the experience data stored in the clone AI memory unit 202 as external information for RAG, but the method of response generation is not limited to this. For example, the response generation unit 203 may use only the generated AI of the specified subject to create a response to the instruction statement. 【0062】 The response generation unit 203 outputs the response to the instruction to the output unit 204. The output unit 204 transmits the response to the instruction to the terminal device 30. 【0063】 (Processing Flow) Next, the information provision process by the information provision device 20 will be described. Figure 7 is a flowchart of the information provision process by the information provision device 20. This process is realized when the processor 22 shown in Figure 5 executes a pre-prepared program and operates as each element shown in Figure 6. 【0064】 First, the instruction acquisition unit 201 acquires the designated person and the instruction from the terminal device 30 (step S201). Next, the response creation unit 203 creates a response to the instruction using a clone AI of the designated person (step S202). The response creation unit 203 outputs the response to the instruction to the output unit 204. Next, the output unit 204 transmits the response to the instruction to the terminal device 30 (step S203). Then, the process ends. 【0065】 [Restrictions on Information Disclosure] The experiences of the subjects may include information that the subjects or related parties may wish to keep confidential from users, such as personal information of the subjects or related parties, trade secrets held by the subjects, and secrets of the subjects' private lives. Furthermore, the subjects' experiences are also primary information of the subjects and should be managed as part of their identity. Primary information refers to information obtained directly by the subject, such as information gained through their own experiences or information obtained through research or experiments they have conducted. Identity management involves managing which identities are disclosed to which parties. 【0066】 Therefore, this section describes methods for restricting the disclosure of experience data from the perspectives of confidentiality, privacy protection, and identity management. The methods described in this section can be used in combination as appropriate. 【0067】 (Method 1) Restrict the disclosure of experience data using classification tags. In Method 1, the disclosure of experience data is restricted based on the classification tags assigned to the experience data. 【0068】 (1) Explanation of Classification Tags First, let's explain the classification tags. The terminal device 10 assigns classification tags to the experience data stored in the data storage unit 102 and transmits them to the information providing device 20. Classification tags are tags used to classify experiences, and examples include school, club activities, hobbies, family, company A, company B, etc. Classification tags may be prepared in advance or created by the subject. In addition, one or more classification tags may be assigned to each piece of experience data. 【0069】 The data acquisition unit 101 of the terminal device 10 searches for experience data that meets predetermined search conditions from the experience data stored in the data storage unit 102, and assigns predetermined classification tags to the retrieved experience data. The data acquisition unit 101 can specify metadata such as date and time, location, and people involved as search conditions. The classification tags are predetermined according to the search conditions. For example, if the search conditions of the data acquisition unit 101 are "the date and time is during the daytime and the location is a school," the classification tag "school" will be assigned to the experience data that meets the search conditions. The classification tags are added (appended) to the metadata of the experience data. 【0070】 Classification tags may be assigned at predetermined intervals, such as every six months or every year, or at significant life events such as graduating from school or starting a job. In addition, although the data acquisition unit 101 assigns the classification tags in the above example, the subject may also assign the classification tags by operating the terminal device 10. 【0071】 (2) Explanation of disclosure restrictions based on classification tags Next, we will explain how to restrict the disclosure of experience data. Here, as examples, we will explain the method using a binary option of disclosure / non-disclosure and the method using cost. 【0072】(2)-1. Method using a binary of disclosure / non-disclosure First, the terminal device 10 sets the disclosure / non-disclosure setting for the classification tag. For example, if the terminal device 10 is willing to disclose the experience data including the classification tag "school", it sets the classification tag "school" to "disclose". On the other hand, if the terminal device 10 wants to keep the experience data including the classification tag "school" confidential, it sets the classification tag "school" to "non-disclosure". The terminal device 10 performs the above settings for each user according to the instructions of the target person. The terminal device 10 transmits the disclosure / non-disclosure setting information to the information providing device 20. 【0073】 Next, the response creation unit 203 of the information provision device 20 extracts information related to the instruction from the disclosed experience data during the RAG search phase. Disclosed experience data is, for example, experience data stored in the clone AI storage unit 202 that does not have the "non-disclosed" classification tag. The response creation unit 203 can determine whether or not the experience data is disclosable based on the disclosure / non-disclosure setting information. The response creation unit 203 attaches the extracted information to the instruction and inputs it into the generating AI of the designated person. The response creation unit 203 obtains the response to the instruction from the generating AI. This prevents non-disclosed experience data (experience to be kept confidential) from being used in the generating AI's response creation. 【0074】 (2)-2. Cost-based method This method controls the disclosure of experience data according to the amount of compensation the user pays to the clone AI. The more compensation the user pays, the more confidential the experience data they can access. 【0075】 First, the terminal device 10 sets a cost for each classification tag. The terminal device 10 sets the cost of classification tags assigned to highly confidential experience data to be higher than that of other classification tags. The terminal device 10 transmits cost setting information, which indicates the classification tag and its cost, to the information providing device 20. 【0076】The user's terminal device 30 transmits the designated target person and instruction text, along with the disclosure request value, to the information providing device 20. The disclosure request value is set, for example, based on the amount of compensation the user pays to the clone AI. The terminal device 30 sets a higher disclosure request value the higher the compensation amount. 【0077】 Next, the response creation unit 203 of the information providing device 20 searches for a combination of classification tags from among multiple classification tags such that the sum of the costs does not exceed the disclosure request value. Then, the response creation unit 203 determines the experience data that can be disclosed based on the search results. For example, the response creation unit 203 makes the classification tags obtained through the search available for disclosure, and the other classification tags not available for disclosure. The response creation unit 203 then selects the experience data stored in the clone AI memory unit 202 that does not contain the non-disclosed classification tags as the experience data that can be disclosed. The experience data that can be disclosed is used in the RAG search phase, as in (2)-1. In this way, the cost-based method allows the disclosure / non-disclosure of experience data to be changed according to the reward amount. 【0078】 (Method 2) Prepare multiple generative AIs and use them according to the user. In Method 2, multiple generative AIs trained on different datasets are prepared and used according to the user. These different datasets are created by classifying accumulated experience data based on classification tags. 【0079】 For example, terminal device 10 generates multiple generative AIs tailored to specific applications using a fine-tuning technique. Specifically, terminal device 10 creates basic learning experience data and additional learning experience data from accumulated experience data. Then, terminal device 10 generates basic learning data for the generative AI from the basic learning experience data, and generates additional learning data for the generative AI from the additional learning experience data. Terminal device 10 pre-trains the generative AI with the basic learning data and further trains the pre-trained generative AI with the additional learning data. 【0080】Figure 8 shows an example of fine-tuning. In Figure 8, experience data up to high school graduation is used as the experience data for basic learning, and experience data from university, part-time jobs, Company A, Company B, and private life are used as the experience data for additional learning. The learning data generation unit 103 of the terminal device 10 can classify the accumulated experience data into experience data for basic learning and experience data for additional learning, respectively, based on the classification tags mentioned above. 【0081】 In Figure 8, the learning unit 104 of the terminal device 10 generates a university knowledge model and a part-time job knowledge model by further training the generated AI (high school graduation model), which was pre-trained using experience data up to high school graduation, with university experience data and part-time job experience data. Furthermore, the terminal device 10 generates a company A business knowledge model and a company B business knowledge model by further training the university knowledge model with experience data from company A and company B. In addition, the terminal device 10 generates a private knowledge model by further training the part-time job knowledge model with private experience data. 【0082】 In this way, the terminal device 10 generates multiple generating AIs by branching a pre-trained generating AI. The terminal device 10 then sets whether each generating AI can be used for each user. The terminal device 10 sets generating AIs trained on experience data that do not contain information to be kept secret to be usable, and sets generating AIs trained on experience data that contain information to be kept secret to be unusable. The terminal device 10 can, for example, set whether a generating AI can be used based on the input of the target person. The information providing device 20 can select the generating AI to use from among the multiple generating AIs according to the above settings. 【0083】 This method prevents generative AI trained on non-disclosed experience data (experience data that you want to keep confidential) from being used to generate answers. 【0084】(Method 3) Modify the response to comply with disclosure restrictions. In Method 3, the response created by the clone AI is modified so that it does not include the information that you want to restrict disclosure of (hereinafter also referred to as "disclosure-restricted information"). Below, we will explain the method using classification tags and the method using dictionaries as examples. The first response of the clone AI to a given instruction will also be referred to as the "initial response" below. 【0085】 (1) Method using classification tags The information providing device 20 uses the experience data stored in the clone AI storage unit 202, and the experience data that includes the classification tag "non-disclosure" as information with disclosure restrictions. 【0086】 The information providing device 20 determines whether the initial response contains information that is conceptually similar to the information subject to disclosure restrictions. If the initial response contains the above information, the information providing device 20 creates an instruction to conceal the relevant part of the initial response. The information providing device 20 inputs the created instruction into the generating AI and obtains a modified response from the generating AI. 【0087】 For example, the response creation unit 203 of the information providing device 20 divides the primary response into chunks and generates a vector for each chunk using an embedding model. The response creation unit 203 also retrieves experience data containing the classification tag "non-disclosed" from the clone AI storage unit 202. The response creation unit 203 then calculates the similarity between the vector of each chunk and the conceptual vector of the retrieved experience data, and searches for chunks whose similarity to the experience data is above a predetermined threshold. If there are chunks whose similarity to the experience data is above a predetermined threshold, the response creation unit 203 creates an instruction statement instructing the corresponding chunk to be kept secret. The response creation unit 203 inputs the created instruction statement into the generating AI and retrieves the corrected response from the generating AI. 【0088】 The response creation unit 203 repeats the above process until the response no longer contains information that is conceptually similar to the information subject to disclosure restrictions, and then creates the final response. 【0089】(2) Method using a dictionary The target user shall prepare a dictionary of restricted disclosure information in advance for each user. The dictionary of restricted disclosure information shall contain NG words and other words that indicate information to be restricted from disclosure. NG words include, for example, confidential information obtained from other companies, such as customer names and raw material composition ratios, as well as personal information such as hobbies and friendships. The information providing device 20 modifies the clone AI's response using the dictionary of restricted disclosure information. The dictionary of restricted disclosure information shall hereafter also be simply referred to as the "dictionary". 【0090】 The response creation unit 203 of the information providing device 20 performs filtering on the initial response using a dictionary. For example, the response creation unit 203 compares the initial response with the dictionary and determines whether the initial response contains any forbidden words from the dictionary. If the initial response contains any forbidden words from the dictionary, the response creation unit 203 creates an instruction sentence that instructs the system to filter the relevant section. The response creation unit 203 inputs the created instruction sentence into the generation AI and obtains a corrected response from the generation AI. 【0091】 The response generation unit 203 repeats the above process until the response no longer contains any forbidden words, and then generates the final response. 【0092】 Furthermore, similarity may be used to compare the initial response with the dictionary. The response generation unit 203 may determine whether the initial response contains information that is conceptually similar to the dictionary's NG words, based on the similarity between the initial response and the NG words in the dictionary. 【0093】 Through the above method, the information providing device 20 can retrospectively restrict the disclosure of experience data. 【0094】 In the above configuration, the data acquisition unit 101 and the data storage unit 102 are examples of storage means, the learning data generation unit 103 and the learning unit 104 are examples of learning means, and the output unit 105 is an example of output means. 【0095】 Furthermore, in the above configuration, the instruction acquisition unit 201 is an example of an instruction acquisition means, and the clone AI storage unit 202, the answer creation unit 203, and the output unit 204 are examples of an answer creation means, a determination means, and an answer correction means. 【0096】[Modifications] Next, modifications of the first embodiment will be described. The following modifications can be combined as appropriate and applied to the first embodiment. 【0097】 (Modification 1) The information providing device 20 may create a response to the instruction by referring to the subject's digital ID wallet. The digital ID wallet stores digitized identity information such as academic transcripts and professional qualification certificates. The subject registers their digital ID wallet with the terminal device 10, and the terminal device 10 transmits the digital ID wallet to the information providing device 20. If the user requests proof of qualifications, the information providing device 20 can search the digital ID wallet for the relevant proof of qualifications and include it in the response before transmitting it to the terminal device 30. 【0098】 (Modification 2) In Method 1 of the first embodiment, the terminal device 10 restricted the disclosure of experience data using classification tags. In addition to this, the terminal device 10 may also restrict the disclosure of the digital ID wallet using classification tags. 【0099】 For example, when a high school graduation certificate is registered as a digital ID wallet in terminal device 10, terminal device 10 assigns a classification tag such as "school" to the registered high school graduation certificate. Note that the classification tags assigned to the digital ID wallet and the classification tags assigned to the experience data are the same. Terminal device 10 transmits the digital ID wallet with the assigned classification tag to information provider device 20. Terminal device 10 also sets disclosure / non-disclosure settings or costs for the classification tag and transmits the setting information to information provider device 20. 【0100】 The information provider 20 can restrict the disclosure of experience data and digital ID wallets by referring to the configuration information. For example, if the classification tag "school" is set to "non-disclosure", the information provider 20 can restrict the disclosure of both experience data related to schools and graduation certificates from schools. 【0101】By using the method described above, the terminal device 10 can restrict the disclosure of academic credentials, qualification certificates, and other documents that it wishes to keep confidential. Furthermore, by unifying the classification tags for experience data and digital ID wallets, the disclosure restrictions for experience data and digital ID wallets become linked, eliminating the need for individual settings. 【0102】 (Variation 3) In addition to a business-use clone AI that performs tasks on behalf of the subject, the information providing device 20 may also provide users with clone AIs representing different stages of the subject's development, such as an elementary school student clone AI or a junior high school student clone AI. In this case, the terminal device 10 appropriately saves the clone AI at different stages of learning and transmits the clone AI at each stage to the information providing device 20. The information providing device 20 then provides the clone AI at a predetermined stage according to the user's request. For example, a user can use an elementary school student clone AI to hear the opinion of an elementary school student about a certain product or service. The information providing device 20 may also provide the clone AI at an intermediate stage of development to the subject or their family. The information providing device 20 can reproduce the subject's words and actions at the time, as well as images from that time, and provide them to the subject or their family. 【0103】 (Modification 4) Method 3 of the first embodiment describes a method for modifying the response created by the clone AI so that it does not contain information subject to disclosure restrictions (a method of restricting disclosure retrospectively). However, this method can be applied not only to the clone AI but also to other AIs. Other AIs include, for example, a generative AI that has learned internal company information (hereinafter also referred to as "corporate AI"). 【0104】 Corporate information refers to information held by a company, including organizational information, customer information, and technical information. Corporate information may be stored on external devices such as the company's file server, or it may be stored in the DB 25 of the information provision device 20. Corporate information is assumed to be pre-associated with classification tags for classifying it. Classification tags may include department names, customer names, and confidentiality categories. Furthermore, classification tags are assumed to be set to be disclosed or not disclosed for each user. 【0105】Specifically, the information provider 20 uses corporate AI and internal corporate information to create a response to a user's instruction. At this time, the information provider 20 modifies the response created by the corporate AI so as not to include information subject to disclosure restrictions, and creates the final response. The information provider 20 can use internal corporate information that includes the classification tag "non-disclosure" as information subject to disclosure restrictions. For example, the information provider 20 determines whether the initial response contains information conceptually similar to the information subject to disclosure restrictions, and if the initial response contains such information, it creates an instruction to conceal the relevant part of the initial response. The information provider 20 inputs the created instruction to the corporate AI and obtains the modified response from the corporate AI. The information provider 20 repeats the above process until the response no longer contains information conceptually similar to the information subject to disclosure restrictions, and creates the final response. 【0106】 Furthermore, the information providing device 20 may modify the response created by the corporate AI based on a pre-prepared dictionary of restricted disclosure information. 【0107】 As described above, the information providing device 20 can retrospectively restrict the disclosure of information not only from clone AIs but also from responses created by other AIs. 【0108】 <Second Embodiment> Figure 9 is a block diagram showing the functional configuration of the information providing device according to the second embodiment. The information providing device 300 includes an instruction text acquisition means 301, an answer creation means 302, a determination means 303, and an answer correction means 304. 【0109】 Figure 10 is a flowchart of the processing by the information providing device of the second embodiment. Instruction acquisition means 301 acquires an instruction from the user (step S301). Answer creation means 302 creates an answer to the instruction using an AI model that includes a trained generating AI learned based on the information to be learned and a storage device (step S302). Determination means 303 determines whether the answer contains information that is conceptually similar to the disclosure restriction information (step S303). Answer modification means 304 modifies the answer by inputting an instruction to the AI ​​model instructing it to conceal the conceptually similar information if the answer contains information that is conceptually similar to the disclosure restriction information (step S304). 【0110】 According to the information provision device 300 of the second embodiment, it becomes possible to restrict the disclosure of information to users of the AI ​​service. 【0111】 <Third Embodiment> Figure 11 is a block diagram showing the functional configuration of the learning device according to the third embodiment. The learning device 400 comprises a storage means 401, a learning means 402, and an output means 403. 【0112】 Figure 12 is a flowchart of the processing performed by the learning device of the third embodiment. The storage means 401 stores experience data that associates the subject's experience, the conceptual vector of the experience, and metadata including classification tags for classifying the experience (step S401). The learning means 402 estimates the result for the experience from the experience and uses the pairs of the experience and the result for the experience as learning data to train the generating AI (step S402). The output means 403 outputs a set of the experience data and the trained generating AI (step S403). 【0113】 According to the learning device 400 of the third embodiment, it becomes possible to provide a clone AI to a user. 【0114】 Some or all of the above embodiments may also be described as follows, but are not limited to the following: 【0115】 (Note 1) An information providing device comprising: an instruction acquisition means for acquiring an instruction from a user; an answer creation means for creating an answer to the instruction using an AI model that includes a trained generating AI that has learned based on information to be learned and a storage device; a determination means for determining whether the answer contains information that is conceptually similar to the information that is restricted from disclosure; and an answer modification means for modifying the answer by inputting an instruction to the AI ​​model that instructs it to conceal the information that is conceptually similar to the information that is restricted from disclosure if the answer contains such information. 【0116】 (Note 2) The information providing device described in Note 1, wherein the determination means determines whether the response contains information that is conceptually similar to the disclosure restriction information, based on the degree of similarity between the response and the disclosure restriction information. 【0117】 (Note 3) The disclosure restriction information is a disclosure restriction information dictionary in which information to be restricted from disclosure is registered for each user, and the determination means is an information providing device according to Note 1 or 2 that determines whether or not the response includes information registered in the disclosure restriction information dictionary. 【0118】 (Note 4) The aforementioned disclosure restriction information dictionary is the information provision device described in Note 3, which includes personal information and confidential information. 【0119】 (Note 5) The information providing device according to Note 1 or 2, wherein the storage device stores the information to be learned and classification tags for classifying the information to be learned in association, and the determination means acquires the disclosure restriction information from the storage device based on setting information in which disclosure restrictions of classification tags are set for each user. 【0120】 (Note 6) The setting information is disclosure / non-disclosure setting information for classification tags, and the determination means is the information provision device described in Note 5, in which the learning target information including non-disclosure classification tags is the disclosure restriction information. 【0121】 (Note 7) The information to be learned is the subject's experience, and the information providing device described in Note 6 stores experience data relating the subject's experience, a conceptual vector of the experience, and classification tags for classifying the experience. 【0122】 (Note 8) The aforementioned classification tags include the learning device described in Note 7, which includes tags such as school, club activities, hobbies, and home. 【0123】 (Note 9) The information to be learned is corporate information, and the information providing device described in Note 6 stores the corporate information and classification tags for classifying the corporate information in association. 【0124】 (Note 10) The classification tag is the learning device described in Note 9, which includes tags such as department name, customer name, and confidentiality category. 【0125】(Note 11) An information provision method which involves obtaining an instruction from a user, creating a response to the instruction using an AI model that includes a trained generating AI that has learned based on the information to be learned and a storage device, determining whether the response contains information that is conceptually similar to the information that is restricted from disclosure, and if the response contains information that is conceptually similar to the information that is restricted from disclosure, modifying the response by inputting an instruction into the AI ​​model that instructs the conceptually similar information to be kept secret. 【0126】 (Note 12) A recording medium that records a program that causes a computer to execute a process in which it obtains an instruction from a user, creates a response to the instruction using an AI model that includes a trained generating AI that has learned based on the information to be learned, and a storage device, determines whether the response contains information that is conceptually similar to the information that is restricted from disclosure, and if the response contains information that is conceptually similar to the information that is restricted from disclosure, inputs an instruction to the AI ​​model instructing it to conceal the information that is conceptually similar, thereby correcting the response. 【0127】 (Note 13) A learning device comprising: a storage means for storing experience data that associates the experience of a subject, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; a learning means for estimating the result for the experience from the experience and training a generative AI using pairs of the experience and the result for the experience as training data; and an output means for outputting a set of the experience data and a trained generative AI. 【0128】 (Note 14) A learning method that stores experience data relating the subject's experience, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; estimates the result for the experience from the experience; trains a generative AI using pairs of the experience and the result for the experience as training data; and outputs a set of the experience data and the trained generative AI. 【0129】(Note 15) A recording medium that stores experience data relating a subject's experience, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; estimates the result for the experience from the experience; trains a generative AI using pairs of the experience and the result for the experience as training data; and causes a computer to execute a process that outputs a set of the experience data and the trained generative AI. 【0130】 Although the present disclosure has been described above with reference to embodiments and examples, the present disclosure is not limited to the above embodiments and examples. Various modifications to the structure and details of the present disclosure can be understood by those skilled in the art within the scope of the present disclosure. 【0131】 10 Terminal device 20 Information provision device 30 Terminal device 101 Data acquisition unit 102 Data storage unit 103 Learning data generation unit 104 Learning unit 105 Output unit 201 Instruction acquisition unit 202 Clone AI storage unit 203 Answer creation unit 204 Output unit

Claims

1. An information providing device comprising: an instruction acquisition means for acquiring instruction text from a user; an answer creation means for creating an answer to the instruction text using an AI model that includes a trained generating AI that has learned based on information to be learned and a storage device; a determination means for determining whether the answer contains information that is conceptually similar to the information that is restricted from disclosure; and an answer modification means for modifying the answer by inputting an instruction text into the AI ​​model that instructs the conceptually similar information to be kept secret if the answer contains information that is conceptually similar to the information that is restricted from disclosure.

2. The information providing device according to claim 1, wherein the determination means determines whether the response contains information that is conceptually similar to the disclosure restriction information, based on the degree of similarity between the response and the disclosure restriction information.

3. The information providing device according to claim 1 or 2, wherein the disclosure restriction information is a disclosure restriction information dictionary in which information to be restricted from disclosure is registered for each user, and the determination means determines whether or not the response includes information registered in the disclosure restriction information dictionary.

4. The information providing device according to claim 3, wherein the disclosure restriction information dictionary includes personal information and confidential information.

5. The information providing device according to claim 1 or 2, wherein the storage device stores the information to be learned and classification tags for classifying the information to be learned in association with each other, and the determination means acquires the disclosure restriction information from the storage device based on setting information in which disclosure restrictions of classification tags are set for each user.

6. The information providing device according to claim 5, wherein the setting information is disclosure / non-disclosure setting information for classification tags, and the determination means uses information to be learned, including non-disclosure classification tags, as disclosure restriction information.

7. The information providing device according to claim 6, wherein the information to be learned is the subject's experience, and the storage device stores experience data relating the subject's experience, a conceptual vector of the experience, and classification tags for classifying the experience.

8. The information providing device according to claim 7, wherein the classification tags include tags such as school, club activities, hobbies, and home.

9. The information providing device according to claim 6, wherein the information to be learned is corporate information, and the storage device stores the corporate information and classification tags for classifying the corporate information in association with each other.

10. The information providing device according to claim 9, wherein the classification tag includes tags such as department name, customer name, and confidentiality category.

11. An information provision method comprising: obtaining an instruction from a user; creating a response to the instruction using an AI model that includes a trained generating AI that has learned based on the information to be learned, and a storage device; determining whether the response contains information that is conceptually similar to the information that is restricted from disclosure; and, if the response contains information that is conceptually similar to the information that is restricted from disclosure, modifying the response by inputting an instruction into the AI ​​model that instructs the conceptually similar information to be kept confidential.

12. A recording medium that records a program that causes a computer to execute a process in which it obtains an instruction from a user, creates a response to the instruction using an AI model that includes a trained generating AI that has learned based on the information to be learned, and a storage device, determines whether the response contains information that is conceptually similar to the information that is restricted from disclosure, and if the response contains information that is conceptually similar to the information that is restricted from disclosure, inputs an instruction to the AI ​​model instructing it to conceal the information that is conceptually similar, thereby correcting the response.

13. A learning device comprising: a storage means for storing experience data that associates the experience of a subject, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; a learning means for estimating the result for the experience from the experience and training a generative AI using pairs of the experience and the result for the experience as training data; and an output means for outputting a set of the experience data and a trained generative AI.

14. A learning method that stores experience data relating a subject's experience, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; estimates the result for the experience from the experience; trains a generative AI using pairs of the experience and the result for the experience as training data; and outputs a set of the experience data and the trained generative AI.

15. A recording medium that stores experience data relating a subject's experience, a conceptual vector of the experience, and metadata including classification tags for classifying the experience; estimates the results for the experience from the experience; trains a generative AI using pairs of the experience and the results as training data; and causes a computer to execute a process that outputs a set of the experience data and the trained generative AI.

Images