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

Abstract

The purpose of the present invention is to create a pseudo-personality having unique features based on the values of a real (or imaginary) person, and to make it possible for a mentee to converse with the pseudo-personality as a mentor. An information processing device according to the present invention inputs instruction text including a quotation extracted from text data related to a predetermined person and input data input from a user into a language model and acquires a response text including the quotation from the language model (see fig. 7).

Description

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

[0001] The present invention relates to a technique for implementing mentoring through dialogue with a user.

[0002] In recent years, chatbots that output natural conversational sentences as if they were conversing with humans have been on the rise. By connecting to large-scale language models (LLMs), the naturalness of conversations and the ability to find answers close to the correct answer have improved dramatically. Furthermore, there is a technology called Retrieval Augmentation Generation (RAG) that connects to knowledge databases that store corporate documents and the like, and responds to questions that require specialized knowledge (Patent Document 1).

[0003] Patent No. 7349219

[0004] As social change accelerates, supporting employee growth is essential for companies to create value. Adult development theory states that people continue to grow even after they reach adulthood, and that the more advanced their developmental stage, the more flexible they become in accepting their own limitations and differences in values. However, it is difficult for people to grow alone; they grow through encounters with others. While many companies have introduced mentoring systems, it is uncertain whether appropriate support can be provided due to issues such as compatibility between mentors and mentees (those receiving mentoring) and the time and space constraints of both parties. Meanwhile, while books offer the opportunity to learn from a wide range of predecessors, it has been impossible to tailor dialogue to suit the mentee.

[0005] Meanwhile, technologies that use chatbots that use large-scale language models (LLMs) to answer user questions in natural language are becoming widespread (see, for example, Patent Literature 1). Because LLMs are trained by collecting information from the Internet, augmentation and generation (RAG) technology is also known that enables them to connect to knowledge databases and use information specific to specific fields or industries. These existing technologies are designed to search for the information users are looking for and return answers that are closer to the correct answer.

[0006] However, what is important for a mentee's growth is not acquiring new information, but the values ​​based on the mentor's experience. They grow because they apply that information to themselves, interpret it, and put it into practice, which leads to self-transformation. This is clear from the fact that even if the same advice is given, whether it is accepted depends on the relationship between the speaker and the listener. Therefore, even if a chatbot that acts as a mentor is implemented using conventional technology, it may only return mediocre answers and may not be effective as a mentor.

[0007] The present invention was made in consideration of the above-mentioned problems, and aims to create a pseudo-personality with unique characteristics based on the values ​​of a real (or fictional) person, and to enable this pseudo-personality to act as a mentor and have a dialogue with a mentee.

[0008] The information processing device of the present invention inputs an instruction sentence including a quotation extracted from text data related to a specified person and input data entered by a user into a language model, and obtains an answer sentence including the quotation from the language model.

[0009] According to the present invention, it is possible to generate a response sentence with unique characteristics based on the values ​​of a real or fictional person in response to text data entered by a user, and to provide advice to the user. Problems, configurations, and effects other than those described above will become clear from the description of the following embodiments.

[0010] FIG. 1 is an explanatory diagram showing an overview of a system according to a first embodiment of the present invention. FIG. 2 is a block diagram showing an example of a system configuration according to the first embodiment of the present invention. FIG. 3 is a sequence diagram showing an example of a processing procedure when setting a mentor, which is executed in the first embodiment of the present invention. FIG. 4 is a sequence diagram showing an example of a processing procedure when generating an answer sentence, which is executed in the first embodiment of the present invention. FIG. 5 is a sequence diagram showing an example of a processing procedure when generating an answer sentence, which is executed in the first embodiment of the present invention. FIG. 6 is a sequence diagram showing an example of a processing procedure when generating an answer sentence, which is executed in the first embodiment of the present invention. FIG. 7 is an explanatory diagram showing an example of a configuration of an instruction sentence related to answer sentence generation to be sent to the LLM server in the first embodiment of the present invention. FIG. 8 is an explanatory diagram showing an example of a configuration of an answer sentence output by the LLM server in the first embodiment of the present invention. FIG. 9 is an explanatory diagram showing an example of a screen when setting a mentor, which is displayed on the display of a client according to the first embodiment of the present invention. FIG. 10 is an explanatory diagram showing an example of a screen when setting a mentor, which is displayed on the display of a client according to the first embodiment of the present invention. FIG. 1 is an explanatory diagram showing an example of a table format of citation type information stored in the mentor management server according to embodiment 1 of the present invention. FIG. 2 is an explanatory diagram showing an example of a table format of answer pattern information stored in the mentor management server according to embodiment 1 of the present invention. FIG. 3 is an explanatory diagram showing an example of a table format of dialogue history information stored in the mentor management server according to embodiment 1 of the present invention. FIG. 4 is an explanatory diagram showing an example of a table format of mentor book information stored in the mentor management server according to embodiment 2 of the present invention. FIG. 5 is an explanatory diagram showing an example of the structure of an instruction statement relating to quotations extraction to be sent to the LLM server according to embodiment 2 of the present invention.

[0011] One example of the present invention is a system that creates a pseudo-personality of a person who will become a mentor and enables dialogue. The system uses text data written by the mentor and outputs advice and guidelines based on the mentor's values ​​using a large-scale language model (LLM). The following definitions apply to terms used in explaining the present invention. A "problem statement" refers to text data written by a user (human) who will become a mentee. An "answer statement" refers to text data generated and output by the large-scale language model (LLM). An "instruction statement" refers to all or part of the text data input into the large-scale language model (LLM). In particular, when the term "task instruction statement" is used, it refers to text data that describes detailed specifications for generating an "answer statement" to a "problem statement."

[0012] <First Embodiment> <Fig. 1: System Overview> Fig. 1 is an explanatory diagram showing an overview of a system according to a first embodiment of the present invention. This system is a mentoring support system that enables an administrator to easily set the characteristics of a person who will become a mentor, and generates a response based on the characteristics of the mentor when a user who will become a mentee posts a question.

[0013] In a first embodiment of the present invention, the mentoring support system includes a mentor management server (MS) and an application server (AS). However, these servers may be configured by a common computer, or either or both may be configured by multiple computers. Furthermore, the mentoring support system may include an LLM server (LS) and a client (CL). The application server (AS), mentor management server (MS), LLM server (LS), and client (CL) are connected via a network (NW). A user (US) operates the client (CL) and communicates with the application server (AS). An administrator (AD) operates the client (CL) and communicates with the application server (AS) and the mentor management server (MS). The administrator (AD) may also directly operate the application server (AS) and the mentor management server (MS).

[0014] <Fig. 2: System Configuration Diagram> Fig. 2 is a block diagram showing an example of the configuration of each of the client (CL), application server (AS), mentor management server (MS), and LLM server (LS) according to the first embodiment of the present invention. Each of these has a transmitting / receiving unit and can be connected to each other via a network (NW). The clients (CL) operated by the user (US) and the administrator (AD) are not distinguished in Fig. 2.

[0015] Although the components are shown separately for convenience of illustration, the processes shown are executed in cooperation with each other. Furthermore, each function in the diagram is realized by cooperation between hardware and software. As is clear from the diagram, each of these components has a control unit, a memory unit, and a transmission / reception unit. The control unit is composed of a central processing unit (CPU), which is a processing unit of a normal computer, etc., the memory unit is composed of a memory device such as a semiconductor memory device or a magnetic memory device, and the transmission / reception unit is composed of a wired or wireless network interface. In addition, each component is equipped with a clock, etc. as necessary.

[0016] The client (CL) is a terminal such as a personal computer, smartphone, tablet, etc. operated by a user (US) or an administrator. The client (CL) has an input / output unit (CLIO), a control unit (CLCO), and a transmission / reception unit (CLSR).

[0017] The input / output unit (CLIO) includes a keyboard (CLIK) for inputting characters, a display (CLOD) for outputting images and characters, and can also be connected to other input / output devices, such as a mouse (not shown), a touch panel (not shown), or a microphone (not shown) for voice input, via an external terminal (not shown).

[0018] The control unit (CLCO) performs input / output control (CLCC) and screen display (CLDD).

[0019] The input / output control (CLCC) receives text and operation instructions from the user (US) input through the input / output unit (CLIO). When it receives a screen drawing command (for example, (ASDD) or (MSDD)) from another server, it draws it in a size and manner suitable for the display (CLOD) by the screen display (CLDD) and outputs it to the display (CLOD) through the input / output control (CLCC).

[0020] The application server (AS) comprises a storage unit (ASME), a control unit (ASCO), and a transmission / reception unit (ASSR), and accepts textual submissions of consultation messages from users (US) via a screen such as the application screen (AS00) in Fig. 12, and transmits the text of the consultation message to the mentor management server, requesting a reply text. Fig. 12 shows a mentor dialogue screen displayed to the user on the application.

[0021] The storage unit (ASME) includes user information (ASUS), a question list (ASQT), a clock (ASCK), and the like, and serves to store information necessary for applications.

[0022] User information (ASUS) is a table stored in the database that manages information about users who are authorized to use the application (not shown). It stores user IDs, names, email addresses, and encrypted login passwords. If necessary, information about each user's contract plan may also be stored. By storing contract plan information, it is possible to specify the contract plan (MS23) that can be used for each mentor setting, as shown on the setting confirmation screen (MS20) in FIG. 11, thereby limiting the users who can use the mentor.

[0023] The question list (ASQT) is a table stored in a database that stores questions (not shown) to guide the direction of the consultation content so that the user (US) can casually ask the mentor for advice on the application. The question list (ASQT) is input by the administrator of the application server (AS) and includes items such as a question ID and question content. The content of the question content is, for example, the question (AS03) displayed on the application screen (AS00) in FIG. 12. The question (AS03) may be configured to select different content depending on a specified date or day of the week. In this case, the question list (ASQT) may include items such as a target period and a target day of the week. It is also possible to allow the user (US) to post any consultation content without presenting the question (AS03), in which case the question list (ASQT) is not necessary.

[0024] FIG. 12 shows an example of the application screen (AS00). The application screen (AS00) uses screen drawing (ASDD) to display mentor information and dialogue logs available to the currently logged-in user (US), and also includes an interface for the user (US) to input new text. Typical items displayed are the mentor name (AS01), question (AS03), consultation text (AS06), and answer text (AS07). The mentor name (AS01) is information about the mentor currently selected as the consultation partner. A face photo or icon image representing the mentor may be displayed next to the mentor name. Additionally, detailed information about the mentor (such as an introduction or book title) may be displayed by clicking or tapping the mentor name. To change the mentor, press Change Mentor (AS02). One or more questions from the Question List (ASQT) are displayed as questions. To change the question, press Change Question (AS04). Once the user (US) has decided on a question to ask the mentor, they can click or tap on Text Input (AS05) to input a question (AS06). The input question (AS06) is sent to the application server (AS), and after a while, a reply (AS07) is displayed. The user (US) can continue the conversation, or if they want to start a new consultation, they can click or tap on Get Question (AS08) to display and select the question (AS03) again.

[0025] The control unit (ASCO) receives a request from the client (CL) and performs user authentication (ASUC), usage data acquisition (ASGD), question list acquisition (ASGQ), answer acquisition (ASGA), screen drawing (ASDD), etc.

[0026] When a user (US) attempts to start an application through a client (CL), a user authentication (ASUC) verifies whether or not input information matches an email address and password registered in user information (ASUS), thereby confirming that the user is the person registered in advance.

[0027] If the user authentication is successful, the Usage Data Acquisition (ASGD) sends a data acquisition request to the Mentor Management Server (MS) according to a predetermined protocol to acquire data such as mentor information and past dialogue logs available to the user (US).

[0028] The Get Question List (ASGQ) retrieves one or more questions from the Question List (ASQT) stored in the Storage Unit (ASME). If necessary, conditions may be specified to retrieve only questions that match the conditions.

[0029] The answer acquisition (ASGA) sends a request to acquire an answer (AS07) to the mentor management server (MS) according to a predetermined protocol based on the mentor name (AS01) selected by the user (US), the question (AS03), and the consultation text (AS06) entered.

[0030] The screen drawing (ASDD) then draws the application screen (AS00) based on the various data received in the response from the mentor management server (MS) and the data extracted from the question list (ASQT).

[0031] The transmitting / receiving unit (ASSR) transmits and receives data to and from the mentor management server (MS), client (CL), etc. via the network (NW), and performs communication control for this purpose.

[0032] The mentor management server (MS) acts as an intermediary between the application server (AS) and the LLM server (LS), and allows the administrator (AD) to register mentor settings and create instructions to request the LLM server (LS) to generate answers based on the registered mentor settings.

[0033] The mentor management server (MS) has a storage unit (MSME), a control unit (MSCO), and a transmitting / receiving unit (MSSR).

[0034] The memory unit (MSME) stores task instructions (MSPT) for requesting the LLM server (LS) to generate a response. These instructions are composed of a combination of static instructions (MSPS) and dynamic instructions (MSPD). The static instructions (MSPS) are the basic instructions used when requesting the LLM server (LS) to generate a response, and are commonly used for any user (US) and any consultation content. On the other hand, the dynamic instructions (MSPD) contain information specific to each mentor and vary the instructions probabilistically, allowing instructions to vary depending on the conditions at the time. This allows the task instructions (MSPT) to flexibly change the instructions depending on various conditions, while maintaining a uniform general policy and format for response generation, thereby enabling diversity in the response content.

[0035] Template directives are stored in the memory unit (MSME), and the dynamic directives (MSPD) to be used are created by embedding appropriate data in the specified locations within the directives. Alternatively, multiple types of directives are created in advance and stored in a table, and the directives that meet specified conditions are extracted and used each time.

[0036] FIG. 7 shows an example of a task directive (MSPT). The task directive (MSPT) is a directive that sets the information and specifications necessary to generate a response to a question from a user (US). In this example, the task directive (MSPT) is composed of a feature definition part (P10) and a response directive part (P20). The feature definition part (P10) is composed of items such as a character definition (P11) and a quotation definition (P12), while the response directive part (P20) is composed of items such as a basic response directive (P21), a metaphor response directive (P22), a response pattern directive (P23), and references (P24).

[0037] All or part of the above items may be adopted, combined in any order, and sent to the LLM server (LS) as a single sentence (MSSP). Items other than those listed above may also be added to the task instructions (MSPT). Alternatively, each item may be sent to the LLM server (LS) in order, and based on the responses, the next element may be slightly modified to create the next instruction, and this may be repeated interactively.

[0038] The characteristic definition part (P10) is composed of items intended to define unique characteristics based on the values ​​of the mentor. The person definition (P11) is an item for describing basic information about the mentor and is a dynamic directive (MSPD) whose contents vary for each mentor. The mentor's name, introduction, book information, etc. may be described. This information may be entered by the administrator (AD), for example, on the setting input screen (MS10) of Figure 10. The quotation definition (P12) is an item for describing information about quotations related to the mentor's knowledge and values. This dynamic directive (MSPD) whose contents vary for each mentor. Here, it is assumed that a list of quotations extracted from books written by the mentor will be entered. This information may also be uploaded by the administrator (AD), for example, on the setting input screen (MS10). It is desirable for the quotations to reflect the knowledge and values ​​of the mentor author. A feature of the present invention is that the quotations are used to generate responses to the user (US). Details regarding the creation of the quotation definition (P12) will be discussed later. The characteristic definition part (P10) also serves as information defining the role of the pseudo-personality who will be the mentor and the unique characteristics of the person. This allows the LLM to be instructed to output a response from the perspective of the person receiving the consultation. In addition, by specifying the name of the mentor, the user can identify that they are speaking to the mentor.

[0039] Next, the answer instruction part (P20) is composed of items intended to define how to respond to a question from the user (US). The symbols and metaphor answer instructions (P22) that appear in the instructions will be discussed later. The basic answer instruction (P21) is a static instruction (MSPS) that describes basic answering methods and the format of the answer. The basic answer instruction (P21) instructs the user (US) to select a quote appropriate to the question from the list of quotes included in the quotation definition (P12). The basic answer instruction (P21) then instructs the system to output the selected quote as the first line of the answer, and further instructs the system to generate an answer to the question based on the content of the quote. Due to its nature, LLM calculates the probability of the most suitable word following a previously output word or sentence, so the answer is generated to be consistent with the previously output sentence. Therefore, by including a quote that reflects the mentor's values ​​at the beginning of the answer, it is expected to generate an answer that is in line with the mentor's values. However, outputting the quotation in the first line of the answer is merely an example, and the quotation may be output anywhere in the answer. The important point about the basic answer instruction (P21) is that the instructions include the inclusion of a quotation in the answer and the creation of an answer to the consultation content based on the content of the quotation.

[0040] Next, the answer pattern instruction (P23) is an item that describes further instructions regarding the answering method, and various patterns of answering methods are assumed. Therefore, the answer pattern instruction (P23) may be configured as a dynamic instruction statement (MSPD) that provides different instruction content each time by preparing an instruction statement describing multiple answer patterns and selecting one of them randomly or according to arbitrary conditions. In the example of FIG. 7, an instruction is given to present a question to the user, but the administrator (AD) may also prepare a different answer pattern. For example, the setting input screen (MS10) of FIG. 10 assumes a situation in which multiple answer patterns are prepared, and the administrator (AD) can freely set a distribution value representing the selection probability of each answer pattern. In this way, by selecting different answer pattern instructions (P23) based on probabilistic differences, it is possible to present the user (US) with an answer using a different approach each time, which is expected to give them new insights and a sense of satisfaction.

[0041] Next, the references (P24) is an item that instructs the user to specify information about the book or other source of the citation. In this example, it is a static directive (MSPS) that instructs the user to enter the information defined in the person definition (P11) in a predetermined format. The book information can also be entered directly in the directive, making it a dynamic directive (MSPD).

[0042] Finally, we will explain the quotations definition (P12) and the symbols and metaphors response prompts (P22). These are extension elements that aim to create more diversity in the responses generated by the LLM and to help improve the user's (US's) understanding and satisfaction.

[0043] The quotation definition (P12) may contain a list of all quotations for the mentor, or each quotation may be further classified into types based on some viewpoint, and a quotation definition (P12) may be prepared for each type. In other words, the directives in the quotation definition (P12) may be dynamic directives (MSPD) whose contents differ for each combination of mentor and type.

[0044] The quotations contained in the quotation definition (P12) are first classified into types (MSTQ3, MSTT3), one of which is selected by probability (MSTT4), and then the most relevant quotation is selected from among them. Probabilistic type classification prevents repeated output of similar answers and allows the mentor's psychological and personality traits to be reflected in the bias in quotation frequency. If a mentor were human, even if their core values ​​were consistent, it would be natural for them to rephrase or give different answers depending on the situation. This variation is human nature and allows the recipient to reassess their situation from multiple perspectives and gain new insights. Probabilistic type classification can be implemented for this purpose. Rather than listing all quotations in the instruction and having the LLM select them, preparing multiple instructions for each type of quotation and implementing a probabilistic selection step in the early stages ensures reliable filtering of candidate quotations, making it easier to achieve the above-mentioned goal.

[0045] A symbol is information associated with each quotation, and the administrator (AD) may optionally prepare one or more symbols. For example, a symbol may be set for each quotation on the UI shown in the setting confirmation screen (MS20) of FIG. 11 . As an example of a symbol, in the quotation definition (P12) shown in FIG. 7 , the symbol #rainbow# is associated with the quotation, "If you think it sounds interesting, that's the starting signal." Note that <> and ## are formats that indicate the quotation and symbol, respectively, within the instruction, and the format itself may be freely determined by the administrator (AD). The basic answer instruction (P21) instructs the user (US) to select a pair of quotation and symbol, and the selected symbol is used in the metaphor answer instruction (P22). The metaphor answer instruction (P22), for example, as shown in the example, instructs the user (US) to add a metaphorical explanation to the content of the quotation and answer presented to the user (US) using the meaning of the symbol. By making such an instruction, a reply sentence expressed as, for example, the metaphor reply (A03) shown in Fig. 8 is generated. The metaphor reply instruction (P22) may be a static instruction sentence (MSPS), or may be a dynamic instruction sentence (MSPD) by preparing multiple patterns as in the reply pattern instruction (P23).

[0046] Simply providing an instruction to generate a response based on a quotation, such as the basic response instruction (P21), can generate a response that is in line with the mentor's values ​​and provide guidance to the user (US). However, by specifying such symbols and adding instructions to further enhance the expression of the response content, it is expected that the user (US) will be more satisfied and understandable, and that their initiative will be enhanced through active interpretation of the image. The metaphor response instruction (P22) is not limited to the example sentence shown in Figure 7, but may also instruct the user to respond with a sentence that metaphorically explains the quotation. Alternatively, it may be an instruction to entrust the LLM with the specific content of a sentence that expands on the quotation according to the metaphorical image represented by the symbol.

[0047] 8 shows an example of a possible output configuration of a reply sentence when the example task instruction sentence (MSPT) shown in FIG. 7 is used in response to a consultation sentence written by a user (US). The question (U00) and the consultation sentence (U10) are examples of the question content selected and the consultation content entered by the user (US) on the application screen (AS00), respectively. The reply sentence (A00) is an example of text data generated when answer generation is performed for this consultation content based on the contents of the task instruction sentence (MSPT) in FIG. 7.

[0048] The correspondence between the components of the answer sentence (A00) and the task instruction sentence (MSPT) is as follows: The opening message (A01) is the result of the LLM selecting one quotation from the list of quotation sentences described in the quotation sentence definition (P12) in accordance with the instructions in the basic answer instruction (P21) and outputting it according to the specified format. In the example of Figure 8, the quotation selected is, "If you think it sounds interesting, that's the start signal." Similarly, the association between the quotation and the consultation content (A02) is the output result of combining the quotation selected from the quotation sentence definition (P12) with the instructions in the basic answer instruction (P21). The answer is generated to the content of the consultation sentence (U10) so that it is consistent in content with the quotation presented in the opening message (A01). In this way, by linking the quote that reflects the mentor's values ​​with the content of the consultation, it is possible to present the user (US) with an answer that is based on the mentor's unique values, rather than a mundane answer as in the prior art.

[0049] Furthermore, the metaphor answer (A03) is the result of outputting in accordance with the metaphor answer instruction (P22) using the symbol #rainbow# associated with the selected quotation. In the example of metaphor answer (A03), a sentence is generated that skillfully uses the symbol "rainbow" to expand on the image of "meeting various people," which is part of the consultation content, and the "idea" associated with the quote "seems interesting."

[0050] The question (A04) is a sentence generated according to the answer pattern instruction (P23), and the citation source information (A05) is a sentence generated according to the reference (P24).

[0051] A task instruction (MSPT) that instructs the generation of an answer sentence with such an output configuration is stored in the memory unit (MSME), and the contents of the dynamic instruction (MSPD) in the task instruction (MSPT) are determined based on the mentor information selected by the user and various probabilistic selection processes, and an answer sentence based on a different instruction sentence is generated each time.

[0052] The memory unit (MSME) also includes mentor basic information (MSTB), quotation information (MSTQ), quotation type information (MSTT), answer pattern information (MSTP), dialogue history information (MSTD), and a clock (MSCK).

[0053] FIG. 13 shows an example of the structure of the Mentor Basic Information (MSTB). The Mentor Basic Information (MSTB) is a table stored in a database that manages basic information about people who will be mentors. It contains fields such as Mentor ID (MSTB1), Name (MSTB2), Introduction (MSTB3), and Books (MSTB4). Other fields, such as remarks about the mentor setting (e.g., notes) and fields specifying users who can use the mentor (e.g., available contract plans), may also be added as needed. The information stored in the Name (MSTB2), Introduction (MSTB3), and Books (MSTB4) fields is used to create instructions that correspond to the Person Definition (P11) in the Task Instructions (MSPT).

[0054] FIG. 14 shows an example of the structure of quotation information (MSTQ). The quotation information (MSTQ) is a table stored in a database that manages various information linked to quotations. It contains fields such as quotation ID (MSTQ1), mentor ID (MSTQ2), type (MSTQ3), quotation (MSTQ4), and symbol (MSTQ5). The information stored in the type (MSTQ3), quotation (MSTQ4), and symbol (MSTQ5) fields is used to create instructions that correspond to the quotation definition (P12) of the task instruction (MSPT).

[0055] FIG. 15 shows an example of the structure of the citation type information (MSTT). The citation type information (MSTT) is a table stored in a database that manages probability values ​​for determining the citation list used in the citation definition (P12). It contains fields such as citation type ID (MSTT1), mentor ID (MSTT2), type (MSTT3), and probability (MSTT4). When a request for answer acquisition (ASGA) arrives from the application server (AS), one type value is selected from multiple type values ​​associated with the target mentor ID value. At this time, the value stored in probability (MSTT4) represents the degree of likelihood of each type being selected. The citation list for the selected type value and the associated symbol list are written in a specified format in the directive of the citation definition (P12).

[0056] FIG. 16 shows an example of the structure of answer pattern information (MSTP). Answer pattern information (MSTP) is a table stored in a database that manages probability values ​​for determining the content of the answer pattern directive used in the answer pattern directive (P23). It contains fields such as answer pattern ID (MSTP1), mentor ID (MSTP2), pattern (MSTP3), probability (MSTP4), and answer pattern directive (MSTP5). When a request for answer sentence acquisition (ASGA) arrives from the application server (AS), one pattern value is selected from multiple pattern values ​​associated with the target mentor ID value. At this time, the value stored in probability (MSTP4) represents the likelihood of each pattern being selected. The directive stored in the answer pattern directive (MSTP5) field in the record of the selected pattern value is used as the directive for the answer pattern directive (P23) in the task directive (MSPT). The format shown in FIG. 16 is an example, and the table may be normalized and the information may be divided into multiple tables.

[0057] FIG. 17 shows an example of the structure of the dialogue history information (MSTD). The dialogue history information (MSTD) is a table stored in a database that manages the dialogue history between the user (US) and the mentor. When the user (US) inputs a consultation into the application and a response is generated from the mentor, a history ID (MSTD1) is automatically assigned, and the date and time (MSTD2) are also recorded by referencing the clock (MSCK). Other items recorded at the same time include the user ID (MSTD3), the selected question ID (MSTD4), the input consultation text (MSTD5), the task instruction text used to generate the response (MSTD6), and the generated response text (MSTD7). The information recorded here is used, for example, to display the dialogue log from the previous interaction with the mentor when the user (US) successfully logs in to the application. In this case, the system receives a usage data acquisition (ASGD) request from the application server (AS), searches for the user ID value of the user (US), and sends text data of past consultations and answers to the application server (AS). Alternatively, when requesting the LLM server (LS) to generate an answer, past dialogue history data may be extracted from this table and sent together. The large-scale language model (LLM) may also generate an answer taking into account the past dialogue history. By generating an answer that takes into account the content of previous consultations, the user (US) can consult with the mentor more interactively. The past dialogue history may include the question (AS03) immediately before the user inputs the question.

[0058] The control unit (MSCO) of the mentor management server (MS) performs data exchange with the application server (AS), data exchange with the LLM server (LS), drawing of screens for executing mentor settings for the administrator (AD), etc. The control unit (MSCO) includes the following functional units for performing these processes: usage data acquisition (MSGD), screen drawing (MSDD), mentor setting registration (MSRM), and LLM input / output control (MSCL).

[0059] The LLM Input / Output Control (MSCL) is responsible for controlling the exchange of data with the LLM Server (LS), and includes thread control (MSCT), feature definition part creation (MSCF), response instruction part creation (MSCI), instruction generation (MSGP), instruction transmission (MSSP), response reception (MSRA), and dialogue storage (MSSD).

[0060] The transmitting / receiving unit (MSSR) transmits and receives data to and from the application server (AS), client (CL), LLM server (LS), etc. via the network (NW), and controls the communications for this purpose.

[0061] The LLM server (LS) is a server that includes a large-scale language model (LLM) and is composed of a transmission / reception unit (LSSR) for receiving input and output from external terminals, a memory unit (LSME) for storing the large-scale language model (LLM), and a control unit (LSCO) for controlling input and output and calculations.

[0062] The control unit (LSCO) includes an instruction reception unit (LSRP) that receives an instruction from the mentor management server (MS) to the large-scale language model (LLM), controls the answer generation process (LSGA) performed in the large-scale language model (LLM), and an answer transmission unit (LSSA) that sends the obtained answer to the mentor management server (MS).

[0063] The large-scale language model (LLM) in the memory unit (LSME) is a pre-trained model that has been machine-learned using large amounts of natural language data. The large-scale language model (LLM) is a type of artificial intelligence that pre-trains a huge amount of text data and uses it to perform natural language processing tasks such as text classification, generation, text summarization, and question answering, thereby generating sentences. By using the large-scale language model (LLM), it is possible to process large amounts of text in a short time and convert it into a shorter form (numbers or summary sentences) that explains the writer's situation.

[0064] Regarding the large-scale language model (LLM), a known model may be used, such as a so-called generative AI of natural language systems that supports input and output in natural language. Specific examples include, but are not limited to, ChatGPT, GPT3, GPT4, etc. Since the use of these models is well known to those skilled in the art, they will not be described in detail in the embodiments of the present invention.

[0065] 3 to 6 are sequence diagrams showing an example of a processing procedure executed in embodiment 1. Of these, Fig. 3 shows the mentor setting process by the administrator (AD), and Figs. 4 to 6 show the process of generating a reply to the user (US).

[0066] The sequence diagram in Figure 3 will now be explained. The administrator (AD) starts up the client (CL) (CLST) and starts up the application of the mentor management server (MS) (MSST). At this time, the administrator (AD) may authenticate whether he has the authority to issue instructions to the mentor management server (MS) (not shown). After starting up, the mentor management server (MS) draws (MSDD) a setting screen like those shown in Figures 9 to 11, and the client (CL) displays it on the display (CLOD).

[0067] FIG. 9 is an example of a settings list screen (MS00) that displays registered mentor setting information. The administrator (AD) can check the mentor setting list (MS01) on this screen and can change or delete the setting contents as necessary. Pressing the new creation button (MS02) transitions the screen to the setting input screen (MS10) shown in FIG. 10. Note that multiple settings can be registered for the same person as different mentors. Mentor settings can be set flexibly according to the purpose.

[0068] 10 shows an example of a setting input screen (MS10). The setting input screen (MS10) is an example of a UI for registering new mentor settings. The administrator (AD) performs mentor information input (CLIM), quotation list input (CLIQ), symbol setting input (CLIS), quotation type input (CLIT), and probability value setting input (CLIP) on the setting input screen (MS10) displayed on the client (CL).

[0069] First, in the mentor information input (CLIM) step, basic information about the person who will be the mentor is entered, such as name (MS11), memo (MS12), introduction (MS13), icon (MS14), etc. When setting up a real person as a mentor, it is desirable that the icon be a photograph or portrait of the person.

[0070] Next, as the citation list input (CLIQ), a list of citations to be used in answer generation is set by uploading (MS15) a list of citations in a specified format, such as CSV data. The list of citations may be prepared manually by a human, or may be extracted using data obtained using a function such as popular highlights in an e-book. Furthermore, book text data may be input into a large-scale language model (LLM), which may then be automatically extracted based on user instructions. When setting citations, information about the books and references from which they were extracted (MS16) must also be input and set.

[0071] Next, the symbol setting input (CLIS) and quotation type input (CLIT) are processes for inputting related information linked to each quotation, and are entered together in the CSV data by the administrator (AD). The symbol and type may also be set manually by a human, or, for example, the table of contents information of the book from which the quotation was extracted may be used as the type. Alternatively, the symbol and type may be left blank and mechanically assigned, for example, by randomly assigning candidate values ​​prepared in advance in the memory unit (MSME).

[0072] Finally, in the probability value setting input (CLIP), a probability value indicating the likelihood of each value being selected is set through the UIs for the quotation type allocation setting (MS17) and the answer pattern allocation setting (MS18), which are shown as examples. For example, in the quotation type allocation setting (MS17) example, the quotation types are set to four types (Hope, Efficacy, Resilience, and Optimism), which are components of psychological capital, and each quotation is classified into one of these types. In this example, the quotation with the Hope type is the most likely to be selected, and when a user (US) inputs a question, the probability of a reply being generated using a Hope type quotation is the highest. The probability values ​​may be set freely by a human, or, for example, the number of quotations of each type may be tallied and the ratio may be set as the probability value. Alternatively, to more accurately reproduce the pseudo-personality of the person who will be the mentor, the mentor's emotional capital could be measured using a questionnaire or the like, or the emotional capital value could be estimated using a large-scale language model (LLM) from the text data of the book written by the mentor, and the result could be used as a probability value. In this way, the psychological and personality characteristics of the mentor can be reflected in the likelihood of a quote type being selected, making it easier to reflect this tendency in the generation of answers.

[0073] After completing the various settings, pressing the confirmation screen button (MS19) transitions to the setting confirmation screen (MS20) shown in FIG. 11 . On this screen, the contents of the registered quotation list (MS21) can be confirmed and, if necessary, changes or deletions can be made. A function such as a sample answer preview (MS22) can also be provided to allow users (US) to check examples of responses to sample consultation data before completing the settings. Furthermore, the users (US) who can conduct dialogue consultations with the set mentor can be restricted. For example, by adding a checkbox UI for available contract plans (MS23), a setting item can be added so that only users (US) subscribed to the target contract plan can conduct dialogue consultations with the mentor. For example, when a consultation is entered by a user who does not have a contract, the consultation can be avoided from being sent to the language model, and a message indicating that there is no contract can be sent instead. Alternatively, it may be possible to prevent users from selecting mentors who do not have a contract from the start. After completing the confirmation, pressing the complete button (MS24) sends the mentor setting information to the mentor management server (MS).

[0074] The mentor management server (MS) that receives the information executes the mentor setting registration (MSRM) process and saves the data in various tables in the memory unit (MSME). When the data saving is complete, the mentor management server (MS) sends a completion message (MSMC) to the client (CL), and the client (CL) displays the completion message (CLMC) to the administrator (AD).

[0075] <Figures 4 to 6: Sequence Diagrams> Figure 4 is a sequence diagram showing the steps from when a user (US) inputs a question on an application to when a response is generated. The user (US) launches the client (CL) (CLST) and logs in to the application (CLAL) by entering information such as an ID and password. The application server (AS) performs user authentication (ASUC), and if authentication is successful, requests the mentor management server (MS) to acquire usage data (ASGD) for the user (US). Based on the user ID information received with the request, the mentor management server (MS) acquires usage data (MSGD), such as mentors available to the user (US) and past conversation logs, from the memory unit (MSME) and transmits the data in response to the application server (AS). The application server (AS) also acquires a question list (ASGQ) and uses the usage data and question list to perform screen drawing (ASDD), which the client (CL) displays on the display (CLOD).

[0076] The user (US) selects a mentor to consult with on the application (CLSM), and then selects a question (CLSQ). The user then confirms the question and inputs a question for the selected mentor (CLIC). The selected mentor and question information, along with the text data for the question, are sent together as consultation conditions to the application server (AS). Upon receiving the consultation conditions, the application server (AS) sends a wait message (ASMW), and the client (CL) displays the wait message until the answer is received (CLMW). If the wait time is not expected to be long, steps (ASMW) and (CLMW) may be omitted.

[0077] Figures 5 and 6 are sequence diagrams showing the processing steps within the LLM input / output control (MSCL) for generating answer sentences. LLM servers (LSs) often have several constraints. For example, there are constraints on the number of characters in the instruction sentence that the LLM can receive, the number of executions, and response time. This can result in excessively long wait times until answer generation. This is addressed by generating threads for each answer generation task and processing them in parallel. The LLM input / output control (MSCL) performs all of this LLM input / output control, determining the number of threads to generate based on the number of answer generation requests.

[0078] The application server (AS) sends a request for answer acquisition (ASGA) along with the consultation conditions (data related to the target user, target mentor, target question, and consultation text) to the mentor management server (MS). During this time, the client (CL) is in a standby state (CLWW).

[0079] The mentor management server (MS) first executes thread generation (MSCT) to launch thread i (MS_i) (MSST), and allocates threads for each consultation condition unit.

[0080] Next, the feature definition part (MSCF) is created based on the target mentor information. The feature definition part (P10) contains the person definition (P11) and quotation definition (P12) directives. Both are dynamic directives (MSPD), so directives must be created. The former is created by extracting the target mentor's data from the Mentor Basic Information (MSTB) table and embedding information such as name (MSTB2), introduction (MSTB3), and book (MSTB4) in the template directives for the person definition (P11). The latter is created by narrowing down the table data in the quotation type information (MSTT) by the target mentor ID, probabilistically selecting the type to use, extracting a list of target quotations from the quotation information (MSTQ) table using the selected type information and the target mentor ID, and embedding the extracted list of pairs of quotations and symbols in the template directives for the quotation definition (P12).

[0081] Next, the mentor management server (MS) executes the answer instruction part creation (MSCI). Since the remaining directives in the answer instruction part (P20), except for the answer pattern directive (P23), are static directives (MSPS), the directives can simply be extracted from the memory unit (MSME). Since the answer pattern directive (P23) is a dynamic directive (MSPD), the answer pattern information (MSTP) table data is first narrowed down by the target mentor ID, a pattern to be used is probabilistically selected, and the answer pattern directive is extracted.

[0082] The created and extracted instructions are then appropriately concatenated in the instruction generation (MSGP) process to produce the actual task instruction (MSPT). Furthermore, in the instruction generation (MSGP) process, the task instruction (MSPT), the text data of the consultation, the text data of the question, and the past dialogue history data between the user (US) and the mentor are merged to produce a complete instruction to be input into the large-scale language model (LLM). Here, the question text data refers to the text data of the question actually selected by the user (US) when selecting a question (CLSQ) in the application. While not necessarily required as a merge target, inputting this data into the large-scale language model (LLM) along with the consultation text can be used as information related to the user's consultation content, which is expected to influence the generation of high-quality answers. Similarly, while past dialogue history data between the user (US) and the mentor is not required to be merged, inputting past questions, inquiries, task instructions, and answers into the large-scale language model (LLM) together can be expected to generate answers that take into account the content of past dialogues. The merging method differs depending on the API specifications defined on the LLM server (LS), but one possible method is to explicitly specify the task instructions, question and inquiry text data, and past dialogue history text data in JSON format. Once the complete instruction is complete, it is sent to thread i (MS_i).

[0083] Thread i (MS_i) sends a complete directive to the LLM server (LS) (MSSP) and waits for a response (MSWW). In this way, the control unit (MSCO) inputs the complete directive into the LLM. The mentor management server (MS) starts a new thread and sends another directive. If thread i (MS_i) does not receive a response within a predetermined period of time (MSCL02), it sends an error (MSSE) to the mentor management server (MS). In response to the error, the mentor management server (MS) determines whether to have the thread resend, stop, or instruct another thread, and if necessary, sends a resend command (MSCL10) to thread i (MS_i).

[0084] When the LLM server (LS) receives the instruction (LSRP), it performs a response generation process (LSGA) and sends a response (LSSA).

[0085] When thread i (MS_i) receives the reply (MSRA), it transfers it to the mentor management server (MS) and stops its own thread (MSET).

[0086] When the mentor management server (MS) receives the answer, it stores the consultation conditions (data related to the target user, target mentor, target question, and consultation text), task instructions, and answer text in the dialogue history information (MSTD) (MSDD). When storage is complete, it terminates the LLM input / output control (MSCL) (SCL20).

[0087] The mentor management server (MS) sends the reply as a response to the application server (AS). The application server (AS) performs screen drawing (ASDD) as necessary, and sends the reply to the client (CL) for display on the screen (CLDD).

[0088] Summary of First Embodiment The mentor management server (MS) (information processing device) according to the first embodiment is configured such that a control unit (MSCO) (serving as an input unit, processing unit, and output unit) receives a consultation message (input data) from a user, acquires a response message from an LLM, and returns the response message to the user. The control unit (MSCO) generates an instruction message (task instruction message (MSPT)) including a quote extracted from text data related to a mentor, a pseudo-personality, and the consultation message from the user, and inputs this to the LLM to acquire a response message for the user from the LLM. This allows the user to receive the response message to the consultation message along with a quote corresponding to the mentor's unique characteristics. Therefore, the user can experience consulting with a pseudo-mentor from the LLM.

[0089] The mentor management server (MS) according to the first embodiment includes a memory unit (MSME) for storing pseudo-personality information describing the characteristics of the pseudo-mentor (mentor basic information (MSTB), quotation information (MSTQ), answer pattern information (MSTP), and quotation type information (MSTT)). The pseudo-personality information includes information defining the pseudo-mentor's unique characteristics of the sentences used by the pseudo-mentor when generating answers. The control unit (MSCO) references the pseudo-personality information to generate an instruction statement instructing the LLM to respond with a sentence having the pseudo-mentor's unique characteristics. This allows the LLM to be instructed to reflect the pseudo-mentor's individuality in the answer, allowing the user to have a consultation experience in which they feel as if they are receiving a unique answer from each pseudo-mentor.

[0090] <Embodiment 2> Embodiment 2 of the present invention improves on the part of embodiment 1 where a quote list was manually prepared in advance. That is, instead of an administrator (AD) registering a list of quotes extracted from books, book data from which quotes are extracted is registered, and a large-scale language model (LLM) extracts quotes instead. Embodiment 2 will be described below, but explanations of parts common to embodiment 1 will be omitted.

[0091] In the second embodiment, in addition to the system configuration of the mentor management server (MS) of the first embodiment, the memory unit (MSME) includes mentor book information (MSTA) and quotation extraction instruction text (MSPQ), and the control unit (MSCO) includes quotation extraction control (MSGQ).

[0092] FIG. 18 shows an example of the structure of Mentor Book Information (MSTA). Mentor Book Information (MSTA) is a table stored in a database that manages text data such as books written by mentors. It contains items such as Book Text ID (MSTA1), Book (MSTA2), Mentor ID (MSTA3), Category (MSTA4), and Text (MSTA5). Category (MSTA4) is a label for each divided text when the entire text data of a book is divided into several partial data, and Text (MSTA5) refers to that partial data. Since there may be a limit to the amount of text data that can be input to a large-scale language model (LLM) at one time, text data is input in units of these categories (MSTA4). In the example of FIG. 18, Category (MSTA4) is defined as each chapter of a book. However, in practice, it is desirable to keep the category (MSTA4) within the size of the text data that can be input. Therefore, if the text size per chapter is large, it may be divided into smaller units. As in the quoted text of the first embodiment, these data may be uploaded by the administrator (AD) from the setting input screen (MS10).

[0093] 19 shows an example of the structure of a quotation extraction directive (MSPQ). The quotation extraction directive (MSPQ) is a directive that requests a large-scale language model (LLM) to extract quotations from the text data of a book. The quotation extraction directive (MSPQ) includes a citation condition (MSPQ1) and a citation instruction (MSPQ2), but other elements may also be added to the structure as needed.

[0094] The quotation conditions (MSPQ1) list appropriate conditions for a quotation to be used by a mentor when responding to a consultation request from a user (US). It is desirable for the quotation to have a broad meaning that can respond to any question from the client and to support the client's spiritual growth. Therefore, in the example of Figure 19, conditions 1 through 4 are listed to extract such quotation, and item 5 is also listed to exclude quotation that does not fit any of the conditions. Note that these items are merely examples and may be freely changed according to the purpose.

[0095] Next, the citation instructions (MSPQ2) provide instructions on how to extract quotations. In the example of Figure 19, a two-step procedure is used: first, multiple quotations that appear to convey a message based on the author's values ​​and experience are extracted, and then each quotation is inferred to fall into one of items 1 through 5. This is because the intermediate inference step allows for complex inference. The first priority is to extract quotations based on the author's experience and values, and then the appropriateness of these quotations for use in responding to the client is individually determined, and quotations other than item 5 are selected. However, depending on the situation, the extraction process may be performed in one step (i.e., directly extracting quotations that fall into items 1 through 4) or in three or more steps. The order of the steps may also be freely changed. Other instructions for the quotation extraction method may include specifying the number of quotations to extract, the length of the quotations to extract, and the output format for the quotations and the item judgment results. To accurately extract a quotation, instructions such as "word for word" may be included.

[0096] The quotation extraction control (MSGQ) controls the entire process of extracting quotations using book data and quotation extraction directives (MSPQ). When book data is received from the client (CL) operated by the administrator (AD), it is stored in the Menta Book Information (MSTA) in the memory unit (MSME). After reading the quotation extraction directives (MSPQ), it reads one text (MSTA5) for each category (MSTA4) from the Menta Book Information (MSTA) and merges the quotation extraction directives (MSPQ) and the text (MSTA5). The directive sender (MSSP) then sends the merged complete directive to the LMM server (LS). The large-scale language model (LLM) then outputs the quotation extraction results in the specified output format, following the same procedure as for answer generation. When the quotation extraction control (MSGQ) receives the answer (MSRA), it stores the quotations determined to fall into any of items 1 to 4 in the quotation information (MSTQ). The stored quotations may be displayed on the client (CL) screen in a format similar to the quotations list (MS21) on the setting confirmation screen (MS20), allowing the administrator (AD) to confirm and modify the extraction results. Then, the administrator (AD) performs symbol setting input (CLIS), quote type input (CLIT), probability value setting input (CLIP), etc. to complete the settings for the quotations. Symbols, quote types, and probability values ​​may be randomly assigned to the quotations.

[0097] This reduces the work of the administrator (AD) manually extracting quotations from book data, and makes it possible to extract quotations using uniform standards that are not dependent on the quality of each administrator's (AD's) work.

[0098] <Embodiment 2: Summary> In the mentor management server (MS) according to embodiment 2, the control unit (MSCO) generates a quotation extraction instruction (MSPQ) that instructs the LLM to generate a quotation by extracting text from book data (mentor book information (MSTA)). This transfers the task of extracting quotations from books to the LLM, thereby reducing the workload of the administrator and the like.

[0099] The quotation extraction directive (MSPQ) instructs the LLM to extract the author's message from the book data and then extract text from the extracted text that meets the characteristics specified by the quotation condition (MSPQ1). This allows the LLM to perform intermediate inference steps when extracting quotations, enabling complex inference by the LLM. In other words, it can automatically extract quotations suitable for use by pseudo-mentors.

[0100] <Modifications of the Present Invention> Although the embodiments of the present invention have been described above, the present invention is not limited to the above embodiments and can be implemented in various modifications. For example, the above embodiments have been described in detail to provide a better understanding of the present invention, and are not necessarily limited to those having all of the described configurations. To achieve the object of the present invention, other existing technologies related to language models, such as RAG and fine tuning, may be used. Furthermore, other models, such as image recognition models and voice recognition models, may be combined to enable the handling of input and output data in formats other than text data, such as image data and voice data.

[0101] In the above embodiments, ChatGPT and the like are given as examples of systems that employ LLM, but other language models may also be used. For example, any language model may be used that is configured to output an answer to a question in natural language when the question is input in natural language by learning the probability that the arrangement of words that make up a sentence is natural in natural language.

[0102] The configurations, functions, processing units, processing means, etc. in the above embodiments may be partially or entirely implemented in hardware, for example, by designing them as integrated circuits. Furthermore, the configurations, functions, etc. may be implemented in software by a processor interpreting and executing programs that implement the respective functions. Information such as programs, tables, and files that implement the respective functions may be stored in storage devices such as nonvolatile semiconductor memory, hard disk drives, and solid-state drives (SSDs), or in computer-readable non-transitory data storage media such as IC cards, SD cards, and DVDs. For example, the control unit (MSCO) may be implemented by hardware such as a circuit device that implements the functions, or by a computing device such as a central processing unit (CPU) that executes software (information processing programs) that implements the functions.

[0103] The control lines and information lines in the above embodiments are those that are considered necessary for the explanation, and do not necessarily show all the control lines and information lines in the product. In reality, it can be considered that almost all components are connected to each other.

[0104] AD Administrator AS Application Server ASCO Control Unit ASME Storage Unit CL Client LS LLM Server MS Mentor Management Server MSCO Control Unit MSME Storage Unit NW Network US User

Claims

1. An information processing device comprising an input unit, a processing unit, and an output unit, wherein the input unit accepts a plurality of quotations extracted from text data related to a specific person and user input data; the processing unit generates an instruction sentence for a language model, the instruction sentence including at least one quote selected from the plurality of quotations and the user input data; the processing unit inputs the instruction sentence into the language model, thereby obtaining a response sentence for the user that includes the at least one quote; and the output unit outputs the response sentence.

2. The information processing device according to claim 1, further comprising: a memory unit for storing pseudo-personality information describing the characteristics of a pseudo-personality that generates the answer sentence; the pseudo-personality information includes information defining characteristics specific to the pseudo-personality of a sentence used by the pseudo-personality when generating the answer sentence; the processing unit, upon receiving a question sentence from the user, references the pseudo-personality information to generate the instruction sentence that instructs the language model to output a sentence having the characteristics specific to the pseudo-personality as the answer sentence; and the output unit outputs the answer sentence based on the response output by the language model in response to the instruction sentence generated by reference to the pseudo-personality information.

3. The information processing device according to claim 2, characterized in that the pseudo-personality information describes the role of the pseudo-personality, and the processing unit generates the instruction sentence that instructs the language model to behave in accordance with the role by referring to the pseudo-personality information.

4. The information processing device according to claim 2, characterized in that the pseudo-personality information describes the quotation and the type of quotation that the pseudo-personality quotes when generating the answer, and the processing unit, by referring to the pseudo-personality information, generates the instruction sentence that instructs the language model to output a sentence using the quotation corresponding to the pseudo-personality as the answer.

5. The information processing device according to claim 4, characterized in that the pseudo-personality information describes one or more of the quotations and describes the type of each of the quotations, the pseudo-personality information describes a first probability that a quotation belonging to the type will be included in the answer, and the processing unit selects one of the quotations in accordance with the first probability by referring to the pseudo-personality information, and generates the instruction statement that instructs the language model to output a sentence using the selected quotation as the answer.

6. The information processing device according to claim 2, characterized in that the pseudo-personality information describes one or more sentence patterns that can be used in the answer sentence and describes a second probability that the sentence pattern will be included in the answer sentence, and the processing unit refers to the pseudo-personality information to select one of the sentence patterns in accordance with the second probability and generate the instruction sentence that instructs the language model to output a sentence using the selected sentence pattern as the answer sentence.

7. The information processing device according to claim 1, characterized in that the information processing device includes a memory unit that stores data describing a dialogue history between the user and the language model, and the processing unit inputs the dialogue history and the instruction sentence together into the language model, thereby obtaining the response sentence that correlates with the dialogue history from the language model.

8. The information processing device according to claim 4, characterized in that the pseudo-personality information describes symbols that metaphorically express impressions of the quotation, and the processing unit, by referring to the pseudo-personality information, generates the instruction sentence that instructs the language model to output, as the answer sentence, a sentence that metaphorically explains the quotation in accordance with the symbol corresponding to the quotation.

9. The information processing device according to claim 2, characterized in that the processing unit is configured to generate, as the instruction sentences, dynamic instruction sentences having content corresponding to the characteristics of the pseudo-personality and static instruction sentences having content common to all the pseudo-personality's characteristics, the memory unit stores templates of the static instruction sentences, and the processing unit generates the static instruction sentences according to the templates and also generates the dynamic instruction sentences according to the pseudo-personality information.

10. The information processing device according to claim 2, characterized in that the pseudo-personality information describes contract information that identifies the users who can interact with the pseudo-personality, and the processing unit allows only the users who are permitted by the contract information to interact with the pseudo-personality to interact with the pseudo-personality.

11. The information processing device described in claim 1, characterized in that the processing unit generates an extraction instruction statement that instructs the language model to generate the quotation by extracting text from book data, and the processing unit uses the text extracted from the book data by the language model in accordance with the extraction instruction statement as the quotation.

12. The information processing device described in claim 11, characterized in that the extraction instruction statement is configured to instruct the language model to extract, from the text described in the book data, the portion of the text that represents a message sent by the author of the book data, and then to extract, from the extracted text, text having one or more specified characteristics.

13. The information processing device according to claim 2, wherein the processing unit provides a user interface for inputting the pseudo-personality information.

14. The information processing device described in claim 1, characterized in that the language model is configured to output an answer to a question in natural language when a question in natural language is input by learning the probability that the arrangement of words that make up a sentence is natural as natural language, and the processing unit uses the answer that the language model outputs in natural language by receiving the instruction sentence in natural language as the answer sentence.

15. The information processing device according to claim 1, characterized in that the processing unit generates the instruction sentence that instructs the user to write the quoted sentence at the beginning of the answer sentence.

16. An information processing method comprising: a step of accepting a plurality of quotations extracted from text data related to a specified person; a step of accepting user input data; a step of generating an instruction sentence for a language model, the instruction sentence including at least one quote selected from the plurality of quotations and the user input data; a step of inputting the instruction sentence into the language model to obtain a response sentence for the user, the response sentence including the at least one quote; and a step of outputting the response sentence.

17. An information processing program that causes a computer to execute the following steps: accepting a plurality of quotations extracted from text data related to a specified person; accepting user input data; generating an instruction sentence for a language model that includes at least one quote selected from the plurality of quotations and the user input data; inputting the instruction sentence into the language model to obtain a response sentence for the user that includes the at least one quote; and outputting the response sentence.

Images