Multiple Generation AI Utilization System and Program

Abstract

This system and program provides a multiple-generating AI utilization system that assigns virtual personalities to multiple generating AIs, protects input data, understands context based on the user's emotions and role, and enables interaction through the personality selected by the user. [Solution] The multiple generation AI control server 1 has an AI dialogue mediation function and sequentially performs processing by a data protection filter function (S11), processing by an emotion analysis engine function (S12), processing by a role-based optimization engine function (S13), and processing by an artificial personality filter function (S14). This is a multiple generation AI utilization system and program.

Description

Technical Field

[0006] , , , , , ,

[0001] The present invention relates to a system that utilizes a plurality of interactive generative AI (Artificial Intelligence) systems (hereinafter referred to as "generative AI"), and more particularly, to a system and program that can assign virtual personalities (virtual personas) to a plurality of generative AI to obtain optimal answers.

Background Art

[0002] [Prior Art] Conventionally, there are a plurality of generative AI, and users utilize each generative AI according to its application. [[ID=!16]]

[0003] [Related Art] In addition, as a related prior art, there is Japanese Patent No. 7503700 "Device, Method, and Program for Creating Technical Documents of Software" (Patent Document 1).

[0004] In Patent Document 1, it is shown that for creating technical documents, a first generative AI model generates a plurality of use cases, a second generative AI model generates a plurality of business logics available for the use cases, and a third generative AI model generates API documents corresponding to each of the business logics.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] Note: There seems to be a misspelling in the original text where "尚、関連する先行技術として" should probably be "また、関連する先行技術として". The translation is adjusted accordingly. Also, the "!16" in the original text might be an error, and it's left as is in the translation as it's not clear what it's supposed to be.However, the conventional methods of using generation AI described above had problems: it was difficult to select the appropriate generation AI for the application, and it was also difficult to determine whether the response from the selected generation AI was appropriate or not, resulting in the ineffective use of multiple generation AIs.

[0007] Furthermore, there were issues with the interaction between specific generative AIs and users, such as the lack of data protection to ensure security, the inability to understand context through the user's emotions and role, and the inability to interact through the user's chosen personality.

[0008] Furthermore, Patent Document 1 does not describe how data protection ensures security, how context can be understood through the user's emotions and roles, or how interaction can be conducted through a personality chosen by the user.

[0009] This invention has been made in view of the above circumstances, and aims to provide a system and program for utilizing multiple generating AIs that assigns virtual personalities to multiple generating AIs, protects data regarding input data, understands context through the user's emotions and role, and enables interaction through the personality selected by the user. [Means for solving the problem]

[0010] To solve the problems of the above-mentioned conventional example, the present invention is a multiple-generation AI utilization system having a multiple-generation AI control server that controls a system of multiple generation AIs, characterized in that the multiple-generation AI control server evaluates the characteristics of multiple generation AIs, identifies a generation AI according to the content of the dialogue, and includes a data protection filter that masks confidential information in the data to be input to the identified generation AI; an emotion analysis engine that analyzes emotions from the data, extracts emotion labels, and incorporates them into prompts for the identified generation AI; a role-based optimization engine that obtains role information from stored user information and incorporates it into prompts for the identified generation AI; and an artificial personality filter that generates prompts to respond with a persona based on a set personality profile.

[0011] The present invention is characterized in that, in the above-mentioned multiple generation AI utilization system, the data protection filter masks personal information as confidential information using a general-purpose named entity recognition model and extracts specific technical terms using a specialized named entity recognition model specialized for specific technical terms.

[0012] The present invention is characterized in that, in the above-mentioned multiple generation AI utilization system, the emotion analysis engine and the role-based optimization engine generate emotion labels and role information as contextual metadata structured in a predetermined format, and incorporate them into prompts for the generation AI together with the text of the input data.

[0013] The present invention is characterized in that, in the above-mentioned multiple generation AI utilization system, the personality filter generates prompts to the generation AI so that it responds with the persona specified by the configured personality profile.

[0014] The present invention relates to a processing program that operates on a multiple generation AI control server that controls a system of multiple generation AIs, characterized in that the multiple generation AI control server evaluates the characteristics of multiple generation AIs, identifies a generation AI corresponding to the content of the dialogue, masks confidential information with a data protection filter for data to be input to the identified generation AI, analyzes emotions from the data with an emotion analysis engine to extract emotion levels and incorporate them into prompts for the identified generation AI, obtains role information from stored user information with a role-based optimization engine and incorporates it into prompts for the identified generation AI, and generates prompts to respond with a persona based on a set personality profile using an artificial personality filter.

[0015] The present invention is characterized in that, in the above program, the data protection filter functions to mask personal information as confidential information using a general-purpose named entity recognition model, and to extract specific technical terms using a specialized named entity recognition model specialized for specific technical terms.

[0016] The present invention is characterized in that, in the above program, the emotion analysis engine and the role-based optimization engine generate emotion labels and role information as contextual metadata structured in a predetermined format, and incorporate them into prompts for the generating AI together with the text of the input data.

[0017] The present invention is characterized in that, in the above program, the personality filter is configured to generate prompts to the generating AI so that it responds with the persona specified by the configured personality profile.

[0018] The present invention relates to a multiple generation AI utilization system having a multiple generation AI control server that controls a system of multiple generation AIs, characterized in that the multiple generation AI control server evaluates the characteristics of multiple generation AIs, selects an initial conversational generation AI from the multiple generation AIs whose characteristics have been evaluated, prepares a normal AI by having it read an initial normal conversation prompt to form a persona for normal conversation and engages in dialogue with a user terminal, and when it extracts specialized knowledge request information from the user terminal during the dialogue, it selects a specialized conversational generation AI from the multiple generation AIs based on its characteristics, prepares a specialized AI by having it read a specialized conversation prompt to form a persona for specialized conversation and takes over the dialogue with the user terminal.

[0019] The present invention is characterized in that, in the above-mentioned system for utilizing multiple AI generation, the multiple AI generation control server registers and stores user profile data from the user terminal before interaction, stores user profile data acquired during interaction, and takes over the stored user profile data to the specialized AI that takes over.

[0020] The present invention relates to a processing program that operates on a multiple generation AI control server that controls a system of multiple generation AIs, characterized in that the multiple generation AI control server evaluates the characteristics of multiple generation AIs, selects an initial conversational generation AI from the multiple generation AIs whose characteristics have been evaluated, reads an initial normal conversation prompt to form a persona for normal conversation and prepares a normal AI to engage in conversation with a user terminal, and when it extracts expert knowledge request information from the user terminal during the conversation, it selects a specialized conversational generation AI from the multiple generation AIs based on its characteristics, reads a specialized conversation prompt to form a persona for specialized conversation and prepares a specialized AI to take over the conversation with the user terminal.

[0021] The present invention is characterized in that, in the above program, the multiple generation AI control server registers and stores user profile data from the user terminal before interaction, stores user profile data acquired during interaction, and functions to transfer the stored user profile data to the specialized AI that takes over. [Effects of the Invention]

[0022] According to the present invention, the multiple generation AI control server is a multiple generation AI utilization system that includes a data protection filter that masks confidential information in the data input to the generation AI, an emotion analysis engine that analyzes emotions from the data, extracts emotion labels, and incorporates them into prompts for the identified generation AI, a role-based optimization engine that obtains role information from stored user information and incorporates it into prompts for the identified generation AI, and an artificial personality filter that generates prompts to respond with a persona based on a set personality profile. Therefore, it has the effect of protecting user data, analyzing emotions to understand the context according to the user's role, and responding with the persona desired by the user.

[0023] According to the present invention, a plurality of generation AI control servers evaluate the characteristics of a plurality of generation AIs, select an initial generation AI for dialogue from the plurality of generation AIs whose characteristics have been evaluated, load an initial prompt for normal dialogue to form a persona for normal dialogue, prepare a normal AI, and conduct a dialogue with a user terminal. When extracting specialized knowledge request information from the user terminal during the dialogue, a specialized generation AI for dialogue is selected from the plurality of generation AIs based on the characteristics, a specialized prompt for dialogue is loaded to form a persona for specialized dialogue, prepare a specialized AI, and continue the dialogue with the user terminal. As a system that utilizes a plurality of generation AIs, the user can effectively utilize the plurality of generation AIs in cooperation to obtain an optimal answer.

Brief Description of the Drawings

[0024] [Figure 1] It is a schematic configuration diagram of this system. [Figure 2] It is a flowchart showing the processing of this system. [Figure 3] It is a flowchart showing the processing of a plurality of generation AI control servers. [Figure 4] It is a flowchart showing the processing of a plurality of generation AI control servers in another application example.

Modes for Carrying Out the Invention

[0025] Embodiments of the present invention will be described with reference to the drawings. [Overview of the Embodiment] In the multiple AI generation system (this system) of the embodiment of the present invention, the multiple AI generation control server evaluates the characteristics of multiple AI generation systems, selects an AI generation system to perform an initial normal conversation, prepares a normal AI with a normal persona (personality) by having the selected AI generation system read prompts for the initial normal conversation, and engages in a conversation with the user. If the system extracts specialized knowledge request information from the user during the conversation, it selects an AI generation system corresponding to that specialized knowledge request information, prepares a specialized AI with a specialized persona by having the selected AI generation system read prompts for the specialized conversation, and takes over the conversation with the user, while also taking over the user's profile data and conversation history data. As a result, the user can effectively utilize multiple AI generation systems in coordination to obtain the optimal answer.

[0026] Here, a hypothetical personality (a fictional personality) refers to a personality formed by imagining a person based on factors such as age, gender, occupation, etc. Thus, the "virtual personality architecture," which assigns a virtual personality to the generated AI, gives users the feeling of security as if they were chatting with a trusted friend or expert.

[0027] [This system: Figure 1] This system will be explained with reference to Figure 1. Figure 1 is a schematic diagram of the system's configuration. As shown in Figure 1, this system consists of a multiple generation AI control server 1, a user terminal 2, multiple interactive generation AIs 3a, 3b, and 3c, and a network 4.

[0028] In this system, we used three multiple AI generators, but any number of them is acceptable as long as there are multiple AIs. Furthermore, it is assumed that there are multiple user terminals 2. The multiple generation AI control server 1, user terminal 2, and multiple generation AIs 3a, 3b, and 3c are connected to network 4.

[0029] [Each part of this system] This section explains each part of this system. [Multiple Generation AI Control Server 1] The multiple generation AI control server 1 is a computer device comprising a control unit 11, a storage unit 12, and an interface unit 13. The control unit 11 implements the processing functions described below by executing the processing program stored in the storage unit 12. The processing functions to be implemented include "means for creating personas for multiple generated AIs," "means for managing user profiles," "means for managing dialogue history," and "means for transferring generated AIs."

[0030] [User Terminal 2] User terminal 2 is a smartphone, tablet, personal computer (PC), etc., which accesses the multiple generation AI control server 1 via network 4, engages in chat (dialogue / conversation) with multiple generation AIs 3a, 3b, and 3c, and displays the results.

[0031] [Multiple generated AI3a, 3b, 3c] The multiple generative AIs 3a, 3b, and 3c are different types of generative AIs provided via the cloud, and are given different prompts (prompts for forming virtual personalities) from the multiple generative AI control server 1, so that they can form different virtual personalities (personas).

[0032] Specifically, for example, a normal AI persona is created in Generator AI3a to engage in initial general conversations, a specialist AI persona is created in Generator AI3b to engage in specialized conversations in a specific field (field x), and a specialist AI persona is created in Generator AI3c to engage in specialized conversations in a specific field (field y).

[0033] [Means for implementing the functions of the multiple generation AI control server 1] The means for realizing the functions of the multiple generation AI control server 1 will be described below. The means for realizing the functions include "means for forming personas for multiple generation AIs," "means for managing user profiles," "means for managing dialogue history," and "means for transferring generated AIs."

[0034] [Methods for creating personas for multiple AI generation] The multiple generation AI control server 1 asks multiple generation AIs 3a, 3b, and 3c the same question multiple times and evaluates the characteristics of each generation AI based on the answers obtained. Then, the multiple generation AI control server 1 selects a generation AI that can engage in general conversation based on the evaluation results of the characteristics of each generation AI, and loads prompts (prompts for virtual personality formation) written to realize general conversation (normal conversation) to form an initial "normal AI persona". A generated AI in which a standard AI persona has been formed is called a "standard AI." In this embodiment, a standard AI persona has been formed in generated AI 3a, and it is considered a standard AI.

[0035] Furthermore, the multiple AI generation control server 1, based on the evaluation results of the characteristics of the generated AIs, causes the generated AI 3b to read prompts written to enable expert conversations in the specialized field x, thereby forming an "expert AI persona". Similarly, the multiple generation AI control server 1 causes the generation AI 3c to read prompts written to enable expert conversations in the field of expertise y, thereby forming another "expert AI persona". A generative AI that has formed a specialized AI persona is called a "specialized AI," a specialized AI in specialized field x is called the "first specialized AI," and a specialized AI in specialized field y is called the "second specialized AI."

[0036] While the above describes creating a specialist AI persona in advance, it is also possible to extract information requesting specialized knowledge from the user (specialized knowledge request information) during a conversation between the user terminal 2 and the normal AI generation AI 3a, generate prompts for specialized conversation from this specialized knowledge request information, and feed them to, for example, generation AI 3b to create a specialist AI persona in real time, thereby turning generation AI 3b into a specialist AI. Creating a specialized AI persona for the generated AI midway through the conversation allows for a smoother transition from the general AI to the specialized AI.

[0037] Alternatively, the multiple generation AI control server 1 may prepare specialized AIs in advance, each with a specialized AI persona. In response to the user's request for specialized knowledge during a conversation, the server may select the most suitable specialized AI for that request, have that AI read prompts for specialized conversation from the requested information, and evolve and form the specialized AI persona.

[0038] Furthermore, the multiple generation AI control server 1 may utilize the Retrieval-Augmented Generation (RAG) function to have a specialized AI provide answers based on documents such as PDFs stored in a specific cloud storage. The specific cloud storage may be set by the system administrator or by the user.

[0039] This enables the system to provide highly customized information based on documents within specific cloud storage, allowing academic researchers and professionals to retrieve necessary knowledge using their own materials, papers, or related materials and papers without spending time or effort, thereby improving user satisfaction.

[0040] [User profile management method] The multiple generation AI control server 1 stores the user profile data initially registered from the user terminal 2 in the storage unit 12, and also stores and updates user profile data such as hobbies and preferences obtained later through interaction with the generation AIs 3a, 3b, and 3c in the storage unit 12. Alternatively, the user profile data may be stored in a user profile database (user profile DB) instead of the storage unit 12. The user profile data stored in the memory unit 12 is shared between the user terminal 2 and the generated AIs 3a, 3b, and 3c.

[0041] Furthermore, the multiple generation AI control server 1 identifies and manages users using user identifiers, so even if the user terminal 2 changes, for example, from a PC to a smartphone, the information is shared and continues to be maintained.

[0042] [Dialogue history management means] The multiple generation AI control server 1 stores the content of the dialogue between the user terminal 2 and the generated AIs 3a, 3b, and 3c as dialogue history data in the storage unit 12. Alternatively, the dialogue history data may be stored in a dialogue history database (dialogue history DB) instead of the storage unit 12. The dialogue history data is loaded and used for learning by the new generative AI when switching from one AI to another.

[0043] [Method for transferring generated AI] In the interaction between the user terminal 2 and the normal AI 3a, the multiple AI generation control server 1, upon receiving expert knowledge request information (first expert knowledge request information) from the user terminal 2, extracts the expert knowledge request information, analyzes its content, selects the optimal generation AI based on the evaluation results of the characteristics of each generation AI performed by the persona formation means for the multiple AI generation, generates prompts for expert conversation from the expert knowledge request information, and prepares an expert AI (first expert AI) with an expert AI persona formed by having the persona formation means for the multiple AI generation read these prompts. Here, the generation AI 3b is referred to as expert AI 3b with an expert AI persona formed.

[0044] Then, the multiple generation AI control server 1 switches from normal AI 3a to specialized AI 3b. During this switch, specialized AI 3b reads and learns the user profile data from the memory unit 12, and also reads and learns the dialogue history data from the memory unit 12. In the above state, the expert AI 3b engages in expert dialogue with the user terminal 2. The content of this dialogue is stored in the storage unit 12 as dialogue history data, and user profile data subsequently obtained through this dialogue is also stored in the storage unit 12.

[0045] Furthermore, in the interaction between the user terminal 2 and the expert AI 3b, when the multiple AI generation control server 1 receives additional expert knowledge request information (second expert knowledge request information) from the user terminal 2, it extracts the additional expert knowledge request information, analyzes its content, selects the optimal generation AI based on the evaluation results of the characteristics of each generation AI performed by the persona formation means for the multiple AI generation, generates a prompt for expert conversation from the additional expert knowledge request information, and prepares an expert AI (second expert AI) with an expert AI persona formed by having the persona formation means for the multiple AI generation read the prompt. Here, the generation AI 3c is referred to as expert AI 3c with an expert AI persona formed. Then, the multiple generation AI control server 1 switches from specialized AI 3b to specialized AI 3c.

[0046] Furthermore, in the interaction between the user terminal 2 and the specialized AI 3b or specialized AI 3c, when general conversation request information (normal conversation request information) is input from the user terminal 2, the multiple generation AI control server 1 extracts the normal conversation request information, analyzes the content of the normal conversation request information, and switches from the specialized AI 3b or specialized AI 3c to the normal AI 3a. During this switch, the normal AI 3a reads and learns the user profile data from the memory unit 12, and also reads and learns the dialogue history data from the memory unit 12.

[0047] Furthermore, information requiring specialized knowledge can be identified by the technical terms used during the conversation. Furthermore, conversational request information can be identified as request information when it is expressed using common terminology used during the conversation, or when technical terms cease to be used within a specific time frame during the conversation.

[0048] [Processing of this system: Figure 2] Next, the processing in this system will be explained with reference to Figure 2. Figure 2 is a flowchart showing the processing of this system. As shown in Figure 2, the multiple generation AI server 1 selects a generation AI to initially interact with based on the characteristics of the generation AIs that have been evaluated, and loads a prompt for initial normal conversation. In Figure 2, generation AI 3a is selected, prompt a is provided to train it, and a normal AI for normal conversation is prepared. Next, the multiple generation AI server 1 outputs question 1 from user terminal 2 to generation AI 3a, inputs the answer a1 generated by generation AI 3a, and outputs it to user terminal 2 (normal conversation).

[0049] Then, when the multiple generation AI server 1 receives question 2 from the user terminal 2, if question 2 contains specialized knowledge request information, it evaluates the characteristics and selects the generation AI best suited to that specialized knowledge, and has it read a prompt based on that specialized knowledge request information. In Figure 2, generation AI 3b is selected, prompt b is provided, and the AI ​​is trained to prepare a specialized AI for specialized conversation. Furthermore, the multiple generation AI server 1 outputs question 2 from user terminal 2 to generation AI 3b, inputs the answer b2 generated by generation AI 3b, and outputs it to user terminal 2 (specialized conversation).

[0050] Then, when the multiple generation AI server 1 receives question 3 from the user terminal 2, if question 3 contains other specialized knowledge request information, it evaluates the characteristics and selects the generation AI best suited to that specialized knowledge, and has it read a prompt based on that other specialized knowledge request information. In Figure 2, generation AI 3c is selected, prompt c is provided, and it is trained to prepare another specialized AI for another specialized conversation. Furthermore, the multiple generation AI server 1 outputs question 3 from user terminal 2 to generation AI 3c, and inputs the answer c3 generated by generation AI 3c, outputting it back to user terminal 2 (another specialized conversation).

[0051] In the above process, the user profile data and dialogue content related to the questions from user terminal 2 are shared among the generating AIs 3a to 3c. Therefore, instead of simply switching which generating AI is interacting with the user, the dialogue can be conducted while maintaining the continuity of the user profile data and dialogue history.

[0052] [Processing by the multiple generation AI control server 1: Figure 3] Next, we will explain the processing of the multiple generation AI control server 1 with reference to Figure 3. Figure 3 is a flowchart showing the processing of the multiple generation AI control server. As shown in Figure 3, the multiple generation AI control server 1 evaluates the characteristics of each generation AI from the answers obtained by asking common questions about the multiple generation AIs using the persona formation means for the multiple generation AIs (S1).

[0053] The multiple AI generation control server 1 selects an AI suitable for general conversation (normal conversation) based on the results of the characteristic evaluation as the AI ​​for normal conversation (S2). Furthermore, the multiple generation AI control server 1 reads a prompt for normal conversation and forms a normal AI with a normal AI persona (S3). Then, it transitions to a normal conversation (S4).

[0054] Next, the multiple generation AI control server 1 determines whether or not it has extracted expert knowledge request information from the user terminal 2 (S5). If it has not extracted the information (No), it returns to process S4. If the multiple AI generation control server 1 extracts expert knowledge request information (if yes), it selects an AI to generate the expert conversation (S6). Then, it has the selected AI read a prompt for the expert conversation to form an expert AI equipped with an expert AI persona (S7). Then, it proceeds to the expert conversation (S8).

[0055] Furthermore, the multiple AI generation control server 1 determines whether or not it has extracted information requesting other specialized knowledge from the user terminal 2 (S9). If it has extracted information requesting other specialized knowledge (Yes), it returns to process S6, loads a prompt for a different specialized conversation, and forms another specialized AI with a different specialized AI persona.

[0056] If the multiple generation AI control server 1 has not extracted any other specialized knowledge request information (No), it determines whether or not it has extracted normal conversation request information from the user terminal 2 (S10). If it has not extracted any information (No), it returns to processing S8. If the multiple generation AI control server 1 extracts normal conversation request information (if Yes), it returns to process S4 and proceeds to normal conversation.

[0057] [Application Examples] This system uses calendar and GPS (Global Positioning System) location information to remind users of their next appointments and initiate conversations to encourage rest if they are traveling long distances, such as while driving. Conversely, if no movement is detected on a holiday, the system proactively initiates conversations to encourage interaction and acquire user profile data.

[0058] [Examples of application] This system can be used in applications such as chat for medical professionals, educational support chat, and chat for government agency users. User terminal 2 typically uses AI to smoothly conduct general conversations, and by combining multiple specialized AIs and taking into account the user's profile data and conversation history data, it provides expert answers. As a result, users can receive highly accurate and relevant answers from the beginning of the conversation.

[0059] [Specific application cases] Specific application cases for providing this system to companies, local governments, and public institutions include "Social Welfare Consultation AI," "Harassment Consultation AI for Companies," and "Multilingual AI for the Tourism and Inbound Market." The specifications of this system (prompts, etc.) are adjusted to accommodate these AI applications.

[0060] [Social Welfare Consultation AI] The "Social Welfare Consultation AI," designed for people suffering from loneliness and mental health issues, is offered to local governments and public institutions. In particular, it enables 24-hour automated responses by AI in the field of social welfare, realizing the digitalization of consultation services that do not rely on traditional human resources.

[0061] [AI-powered harassment consultation service for businesses] The "AI-powered harassment consultation service for businesses," which provides support for harassment issues, internal reporting, and mental health issues related to interpersonal relationships within companies, creates an environment where employees can consult anonymously and safely in accordance with the Whistleblower Protection Act. It also provides anonymous reports to management from the AI, enabling prompt action against potential troubles and misconduct within the company and providing added value in compliance management.

[0062] [Multilingual AI for the Tourism and Inbound Market] In the railway, hotel, and tourism industries, a multilingual AI capable of handling foreign languages ​​for the tourism and inbound market will provide guidance and medical support for tourists. In particular, in the Japanese market, where an increase in foreign visitors is expected, it will meet the multilingual needs of the tourism industry.

[0063] [Advantages of this system] The advantages offered by this system include improved emotional resonance and trust, increased motivation and engagement, reduced stress and resistance, and the ability to build and collect continuous relationships.

[0064] [Emotional resonance and increased trust] AI that possesses a virtual personality and behaves like a human can more easily build emotional connections with users through its human-like behavior. People tend to trust those who appear to have emotions more than those who simply respond mechanically, and interactions with AI tend to be smoother. In particular, AI personas that provide emotional support are effective in the fields of mental health and counseling. This reduces feelings of loneliness, makes users feel more familiar, and lowers their psychological resistance to problem-solving.

[0065] [Improvement of learning motivation and engagement] In the fields of education and training, AI with a virtual personality can enhance learner motivation and improve long-term learning effectiveness. Learners are more likely to become interested in learning when they feel they are receiving human support rather than just mechanical instruction. AI with a virtual personality can provide personalized support and offer feedback and encouragement based on user responses, enabling more effective learning assistance.

[0066] [Reduction of stress and resistance] By giving AI a virtual personality, users are more likely to perceive the AI ​​not merely as a "machine" or "supervisor," but as a "partner" with whom they can converse and seek advice. This allows users to engage in conversations more easily and without stress. In the medical field in particular, it is expected to be especially useful in that patients will be more likely to consult with the AI ​​about their health problems and mental health concerns, leading to better outcomes.

[0067] [Continuous relationship building and data collection] When AI can build long-term relationships through its virtual personality, users tend to use that AI more consistently. This continued use allows the AI ​​to accumulate data based on user needs and behavioral patterns, enabling it to provide even more personalized responses. As a result, it can provide more accurate services, leading to improved user satisfaction. This translates into business-beneficial features such as driving advertising and purchases, and reducing churn rates.

[0068] [Another application example: Figure 4] Next, we will explain another application example of this system with reference to Figure 4. Figure 4 is a flowchart showing the processing of the multiple generation AI server 1 in another application example. In this system, the multiple generation AI server 1 functions as an AI agent. However, in another application example, instead of AI agent functionality, it possesses AI dialogue mediation functionality, and combines multiple filter functions and multiple engine functions, as described below, to realize an AI dialogue mediation system. This AI dialogue mediation system is included in this system.

[0069] Firstly, the uniform and impersonal responses provided by external general-purpose AIs impose a psychological burden on users and are unsuitable for use in medical settings where specialized and emotional considerations are particularly required. Furthermore, directly transmitting personal information and sensitive medical information contained in user statements to an external AI (generating AI) 3 poses a significant security risk.

[0070] Therefore, as shown in Figure 4, the multiple generation AI control server 1 is equipped with the function of a dialogue mediation server and has a series of pipelines that process the input text from the user terminal 2 in the following order: first, processing by the data protection filter function (S11), second, processing by the sentiment analysis engine function (S12), third, processing by the role-based optimization engine function (S13), and fourth, processing by the artificial personality filter function (S14).

[0071] This pipeline anonymizes the input text, assigns sentiment and role tags, and dynamically generates instruction prompts for the external AI3. Furthermore, in response to the external AI3, the text representation is converted (regenerated) based on the specified personality before being returned to the user terminal 2.

[0072] The novel hierarchical architecture combines multiple information processing filters—each with different objectives: security (data protection), context (emotions / roles), and expression (personality)—in a specific order. Rather than relying on a single function, it integrates and controls these filters, making it useful in systems that dynamically transform external general-purpose AI (external AI)3 into a "safe, context-aware, and personality-driven" conversational partner.

[0073] [Description of functionality in another application example] The following sections will provide a detailed explanation of the data protection filter, sentiment analysis engine, role-based optimization engine, and artificial personality filter, which are the functions that enable the series of processes in the pipeline.

[0074] [Data protection filter] As a prerequisite, conversations in medical settings contain a mixture of personally identifiable information (PII) and protected health information (PHI), and it is necessary to comprehensively and accurately identify these and anonymize confidential information. In another application example, a general-purpose named entity recognition model (AI-based extraction model) is used to detect names and place names from the input text, and then a specialized named entity recognition model (AI-based extraction model) fine-tuned specifically for medical terminology is used to detect disease names and drug names. This is a two-stage detection process.

[0075] The detected generic proper nouns are replaced (masked) with category names (e.g., "Tanaka" → [Patient T]), and at the same time, a flag tag "Concerning Information" is generated and passed along with the masked text to subsequent processing.

[0076] In another application example, a hybrid NER (Named Entity Recognition) method combining a general-purpose model (general-purpose named entity recognition model) and a specialized model (specialized named entity recognition model) can improve the accuracy of information detection in the specific domain of medicine. While masking and protecting the information, the fact that "sensitive information was included" is communicated to an external AI3 as a tag, allowing the external AI3 to adjust the caution of its response.

[0077] [Emotion Analysis Engine] As a premise, conventional AI has the problem of giving responses that are inappropriate to the situation because it cannot understand who the user is (their role) or what emotional state they are in (their emotions). In another application example system, the emotion analysis engine processes input text from user terminal 2 using natural language processing, extracts emotion labels such as "joy," "anger," and "anxiety," and performs emotion analysis. These emotion labels are then incorporated into prompts for external AI 3.

[0078] [Role-based optimization engine] As a role-based optimization engine, it stores user data associated with user IDs in the storage unit 12, retrieves role information such as the user's occupation from that user data, generates "context metadata" structured in a predetermined format (e.g., JSON tags) using the extracted emotion labels and role information, and incorporates this into a prompt for the external AI3 along with the original text.

[0079] The operation of the emotion analysis engine and role-based optimization engine described above allows for the systematic integration of two different types of information—the user's internal state (emotions) and external attributes (roles)—and provides this as context for the external AI3 to generate responses, thereby enabling the external AI3 to make more sophisticated and human-like situational judgments.

[0080] [Artificial Personality Filter] As a prerequisite, the responses of external AI3 depend on the underlying training data, making it difficult to provide a flexible personality tailored to specific organizational cultures or user preferences. A "personality profile" (data defining a persona) consisting of multiple parameters such as gender, age, dialect, level of politeness, and level of expertise is set and selected by the user or administrator and stored in the memory unit 12.

[0081] Upon receiving a response (primary generated text) from an external AI3, the system dynamically generates a secondary prompt that instructs the AI3 to maintain the content of the primary generated text while converting it to the tone and expressions of the persona specified by the personality profile, based on the configured and selected personality profile. This prompt is then input back into the language model (LLM) to obtain the final response.

[0082] The idea is to assign (wrap) a "personality" as a post-processing step, without relying on the performance of the external AI3. The unique aspect lies in the specific prompt engineering technical process that dynamically generates instruction prompts to the LLM from a combination of multiple parameters, thereby enabling the provision of diverse personas in a scalable manner.

[0083] [Effects of the embodiment] According to this system, the multiple generation AI control server 1 evaluates the characteristics of multiple generation AIs 3a, 3b, and 3c, selects a generation AI to perform initial normal dialogue, prepares a normal AI with a normal persona (personality) by having the selected generation AI read prompts for initial normal dialogue, and engages in dialogue with the user. If the system extracts specialized knowledge request information from the user during the dialogue, it selects a generation AI corresponding to that specialized knowledge request information, prepares a specialized AI with a specialized persona by having the selected generation AI read prompts for specialized dialogue, and takes over the dialogue with the user, also taking over the user's profile data and dialogue history data. As a result, the user can effectively utilize multiple generation AIs 3a, 3b, and 3c in coordination to obtain the optimal answer.

[0084] In another application example system, the multiple generation AI control server 1 is equipped with an AI dialogue mediation function, and processes data using a data protection filter function, emotion analysis engine function, role-based optimization engine function, and artificial personality filter function in sequence. This protects user data, improves responses in specialized areas, analyzes emotions, and enables responses in the user's desired persona according to the user's role. [Industrial applicability]

[0085] The present invention is suitable for a system and program that utilizes multiple generating AIs by assigning virtual personalities to multiple generating AIs and effectively utilizing those multiple generating AIs. [Explanation of Symbols]

[0086] 1…Multiple generation AI control server, 2…User terminal, 3a,3b,3c…Generation AI system (Generation AI), 4…Network, 11…Control unit, 12…Storage unit, 13…Interface unit

Claims

1. A system utilizing multiple generation AIs, having a multiple generation AI control server that controls a system of multiple generation AIs, The aforementioned multi-generation AI control server is The characteristics of the aforementioned multiple generation AIs are evaluated, and a generation AI is identified that corresponds to the content of the dialogue. A data protection filter that masks confidential information in the data input to the aforementioned identified generating AI, An emotion analysis engine that analyzes emotions from the aforementioned data, extracts emotion labels, and incorporates them into prompts for the identified generative AI, A role-based optimization engine that obtains role information from the user's information to be memorized and incorporates it into the prompts for the identified generating AI, A multiple-generation AI utilization system characterized by having an artificial personality filter that generates the prompt to respond with a persona based on a set personality profile.

2. The data protection filter is characterized in that it masks personal information as confidential information using a general-purpose named entity recognition model and extracts the specific technical terms using a specialized named entity recognition model specialized for specific technical terms, as described in claim 1.

3. The multiple generation AI utilization system according to claim 1 or 2, characterized in that the emotion analysis engine and the role-based optimization engine generate the emotion labels and the role information as context metadata structured in a predetermined format, and incorporate them into prompts for the generation AI together with the text of the input data.

4. The multiple AI generation system according to claim 1 or 2, characterized in that the personality filter generates prompts to the generating AI to respond with the persona specified by the configured personality profile.

5. A processing program that operates on a multiple generation AI control server that controls a system of multiple generation AIs, The aforementioned multiple generation AI control server is The characteristics of the aforementioned multiple generation AIs are evaluated, and a generation AI is identified that corresponds to the content of the dialogue. The confidential information is masked by a data protection filter for the data to be input to the identified generating AI. The emotion analysis engine analyzes the emotions from the data, extracts emotion labels, and incorporates them into prompts for the identified generative AI. The role-based optimization engine retrieves role information from the stored user information and incorporates it into the prompts for the identified generating AI. A program characterized by using an artificial personality filter to generate prompts that respond with a persona based on a configured personality profile.

6. The program according to claim 5, characterized in that the data protection filter is configured to mask personal information as confidential information using a general-purpose named entity recognition model, and to extract the specific technical terms using a specialized named entity recognition model specialized for specific technical terms.

7. The program according to claim 5 or 6, characterized in that the emotion analysis engine and the role-based optimization engine are configured to generate the emotion label and the role information as context metadata structured in a predetermined format, and to incorporate this into a prompt for the generating AI together with the text of the input data.

8. The program according to claim 5 or 6, characterized in that the personality filter is configured to generate prompts to the generating AI to respond with the persona specified by the configured personality profile.

9. A system utilizing multiple generation AIs, having a multiple generation AI control server that controls a system of multiple generation AIs, The aforementioned multi-generation AI control server is The characteristics of the aforementioned multiple generation AIs are evaluated, From the multiple generation AIs whose characteristics have been evaluated, an initial generation AI for dialogue is selected, an initial prompt for normal dialogue is loaded to form a persona for normal dialogue, a normal AI is prepared, and dialogue with the user terminal is performed. A system for utilizing multiple generating AIs, characterized in that, when specialized knowledge request information is extracted from the user terminal during a conversation, a generating AI for specialized conversation is selected from the multiple generating AIs based on the characteristics, a prompt for specialized conversation is read, a persona for specialized conversation is formed, the specialized AI is prepared, and the conversation with the user terminal is taken over.

10. The multiple generation AI utilization system according to claim 9, characterized in that the multiple generation AI control server registers and stores user profile data from the user terminal before interaction, stores user profile data acquired during interaction, and passes the stored user profile data to the specialized AI that takes over.

11. A processing program that operates on a multiple generation AI control server that controls a system of multiple generation AIs, The aforementioned multiple generation AI control server is The characteristics of the aforementioned multiple generation AIs are evaluated, From the multiple generation AIs whose characteristics have been evaluated, an initial generation AI for dialogue is selected, an initial prompt for normal dialogue is loaded to form a persona for normal dialogue, a normal AI is prepared, and dialogue with the user terminal is performed. A program characterized by extracting expert knowledge request information from the user terminal during a conversation, selecting a specialized conversational generation AI from the multiple generation AIs based on the characteristics, having it read a specialized conversational prompt to form a persona for the specialized conversation, preparing the specialized AI, and then taking over the conversation with the user terminal.

12. The program according to claim 11, characterized in that the multiple generation AI control server registers and stores user profile data from the user terminal before interaction, stores user profile data acquired during interaction, and functions to transfer the stored user profile data to the specialized AI that takes over.

Images