Text generation device, text generation method, text generation system, and text generation program

Abstract

To support the creation of appropriate documents in fields such as education. [Solution] The text generation device receives observation facts about a subject and a request to generate text for a predetermined purpose from a terminal device used by an observer, outputs a generation prompt to the generation AI instructing it to generate one or more draft texts for the predetermined purpose about the subject based on the received observation facts, retrieves one or more draft texts from the generation AI, outputs a modification prompt to the generation AI instructing it to modify each of the retrieved draft texts to meet predetermined safety standards, retrieves each of the modified draft texts from the generation AI, outputs an evaluation prompt to the generation AI instructing it to evaluate each of the retrieved draft texts according to predetermined evaluation standards, retrieves the evaluation results from the generation AI, and outputs each of the retrieved draft texts and the evaluation results to the terminal device.

Description

Technical Field

[0001] The present disclosure relates to a text generation device, a text generation method, a text generation system, and a text generation program.

Background Art

[0002] For example, in the field of education, teachers need to create a variety of documents. Examples of documents include, for example, attendance records, educational support plans, guidance plans, class newsletters, event guides, report cards, learning guidance records, recommendation letters for exams, etc. These documents have columns where teachers should fill in the text. Each time a teacher creates a document, they need to consider the situation and feelings of children or students (hereinafter referred to as students) and create appropriate text.

[0003] Such work of teachers is necessary for each student at timings such as daily, weekly, monthly, and semesterly, and it is a heavy burden for teachers who are usually busy.

[0004] Regarding this, Patent Document 1 describes a system for generating text to be filled in an attendance record or the like using a text generation AI.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, in reality, texts read by students and parents in educational settings need to be created considering various requirements, such as ensuring quality from an expert's perspective, complying with various laws and regulations, and not causing offense or anxiety to the reader. Ideally, not only individual teachers but also the organizations to which those teachers belong should be able to create texts that take these various requirements into account.

[0007] As shown in Patent Document 1, even if a generative AI is used, it is not possible to generate text that takes into account the various requirements mentioned above.

[0008] Not only in education, but also in other fields such as healthcare and welfare, it is necessary to create documents that take these various requirements into consideration. Furthermore, ideally, not only the individual creating the document, but also the organization to which that individual belongs should be able to create documents that take these various requirements into consideration.

[0009] This disclosure aims to support the creation of appropriate documents in fields such as education. [Means for solving the problem]

[0010] One aspect of the present disclosure is a text generation device capable of communicating with a generative AI using a large-scale language model, comprising at least one processor and at least one memory resource, wherein the processor performs: reception processing to receive observation facts about a subject and a request to generate text for a predetermined purpose from a terminal device used by an observer; generation processing to output a generation prompt to the generative AI instructing it to generate one or more draft texts for the predetermined purpose about the subject based on the received observation facts, and to obtain one or more draft texts from the generative AI; modification processing to output a modification prompt to the generative AI instructing it to modify each of the obtained draft texts to meet predetermined safety standards, and to obtain each of the modified draft texts from the generative AI; evaluation processing to output each of the obtained draft texts and evaluation results to the terminal device. [Effects of the Invention]

[0011] This disclosure can help in creating appropriate documents in fields such as education.

[0012] Other issues, configurations, and effects not mentioned above will be clarified by the following description of the embodiments. [Brief explanation of the drawing]

[0013] [Figure 1] Figure 1 shows an example of the configuration of a text generation system according to one embodiment. [Figure 2A] Figure 2A shows an example of a UI screen (generation request). [Figure 2B] Figure 2B shows an example of a UI screen (approval). [Figure 2C] Figure 2C shows an example of a UI screen (generation report). [Figure 3] Figure 3 is a flowchart (part 1) showing an example of the text generation process. [Figure 4] FIG. 4 is a flowchart showing an example of the text generation process (Part 2). [Figure 5] FIG. 5 is a diagram showing an example of a prompt and output related to text generation. [Figure 6] FIG. 6 is a diagram showing an example of a prompt and output related to text correction. [Figure 7] FIG. 7 is a diagram showing an example of a prompt and output related to text refinement. [Figure 8] FIG. 8 is a diagram showing an example of a prompt and output related to text evaluation. [Figure 9] FIG. 9 is a flowchart showing an example of a text generation process including translation processing. [Figure 10] FIG. 10 is a diagram showing an example of a prompt and output related to text translation. [Figure 11] FIG. 11 is a flowchart showing an example of a text generation process including accessibility refinement processing. [Figure 12] FIG. 12 is a diagram showing an example of a prompt and output related to accessibility refinement. [Figure 13] FIG. 13 is a flowchart showing an example of a log monitoring process.

DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, an embodiment according to the present disclosure will be described based on the drawings. In all the drawings for explaining the embodiment, the same components are generally denoted by the same reference numerals, and the repeated description thereof will be omitted. Further, in the following embodiments, the components (including element steps, etc.) are not necessarily essential except in cases where it is particularly specified or considered to be clearly essential in principle. Further, when it is said that "comprising A", "consisting of A", "having A", or "including A", other elements are not excluded unless it is particularly specified that only that element is involved. Similarly, in the following embodiments, when referring to the shape, positional relationship, etc. of components, etc., unless it is particularly specified or considered to be clearly otherwise in principle, those substantially approximating or similar to the shape, etc. are included.

[0015] FIG. 1 is a diagram showing a configuration example of a text generation system according to an embodiment. The text generation system 1 includes a text generation device 2, an AI (Artificial Intelligence) system 3, and one or more terminal devices 4, which are communicably connected via a network 5, respectively. Each terminal device 4 is used by a user. As will be described in detail later, when the roles are distinguished, the users are observers, distribution partners, approvers, administrators, etc.

[0016] The text generation system 1 assists in creating text for use by the user. In this embodiment, for the sake of clarity, we will consider an example scenario in which a school teacher communicates with the parents of a student under their supervision. Specifically, the teacher operates terminal device 4 to request the text generation device 2 to generate a draft text to be used for communication with the parents. Upon receiving the request, the text generation device 2 works in conjunction with AI system 3 to generate the draft text and returns it to the teacher's terminal device 4. This allows the teacher to create the text to be used for communication with the parents based on the draft text. The teacher can be described as the "observer" of the student, the student as the "subject" of the observation, and the parents as the "recipient" of the text. Note that the "subject" and the "recipient" may be the same person. For example, the teacher's superior may be the "approver," the system manager of the school to which the teacher belongs, or the business operator that provides the text generation service by the text generation device 2 to the teacher and school may be the "administrator."

[0017] Network 5 is a communication network such as a LAN (Local Area Network), WAN (Wide Area Network), the Internet, or a mobile phone network. Network 5 may also be a VPN (Virtual Private Network) on a communication network.

[0018] The text generation device 2 can be configured using a general-purpose computer such as a server computer, which includes, for example, a processor such as a CPU (Central Processing Unit), memory such as DRAM (Dynamic Random Access Memory), storage such as an HDD (Hard Disk Drive) or SSD (Solid State Drive), input devices such as a keyboard, mouse, or touch panel, output devices such as a display or audio output device, and communication modules such as a NIC (Network Interface Card) (all not shown). The text generation device 2 can also be configured using multiple computers. Memory and storage correspond to memory resources.

[0019] Terminal device 4 can consist of a general-purpose computer such as a personal computer, smartphone, tablet, or wearable device, which includes, for example, a processor such as a CPU, memory such as DRAM, storage such as an HDD or SSD, input devices such as a keyboard, mouse, or touch panel, output devices such as a display or audio output device, and a communication module such as a NIC (not shown). Terminal device 4 can also consist of multiple computers. Memory and storage correspond to memory resources.

[0020] AI System 3 is a collection of AI models that utilize machine learning algorithms, such as neural networks. Each AI model outputs a prediction result by performing a given task based on input information. The generative AI included in AI System 3 is a service that provides the functionality of so-called generative AIs (LLMs (Large Language Models)), such as GPT, Gemini, and Claude, via APIs (Application Programming Interfaces). AI System 3 provides the generative AI with natural language commands (prompts) to generate the desired results. The generative AI can generate specific information based on information containing various types of data, such as text, documents, images, videos, audio, and sensor information.

[0021] In this embodiment, when the AI ​​system 3 receives instructions, for example via an API, it causes the generating AI to generate response information and sends the result to the source of the instructions as the API's return value. Note that the AI ​​system 3 and the text generation device 2 may be an integrated system. For example, the text generation device 2 can be configured to include at least the generating AI from the AI ​​system 3.

[0022] The text generation device 2 has a predetermined program 211 (for example, a dedicated application in this embodiment) installed on it that runs on the OS (Operating System), and this program 211 allows the user to use the text generation service provided by the text generation device 2.

[0023] Terminal device 4 has a predetermined program 411 installed that runs on the OS (for example, a dedicated application for this embodiment or a general-purpose web browser), and this program 411 allows the user to use the text generation service provided by the text generation device 2. The following explanation will use the case of using a web browser as an example.

[0024] The text generation device 2 has functional blocks consisting of a processing unit 21, a storage unit 22, and a communication unit 23. The function of the processing unit 21 is realized by the computer's processor executing a predetermined program loaded into memory. Some or all of the functions of the processing unit 21 may be implemented as hardware using integrated circuits or the like. Furthermore, some or all of the functions of the processing unit 21 may be implemented by multiple computers arranged in a distributed manner.

[0025] The processing unit 21 controls the entire text generation device 2. The processing unit 21 provides and displays various UI (User Interface) screens to the terminal device 4 via the communication unit 23. The processing unit 21 also receives user operations and inputs for the various UI screens from the terminal device 4 via the communication unit 23.

[0026] The basic functions of the processing unit 21 are as follows: The processing unit 21 receives observational facts about the subject (student) and a request to generate a document (such as a contact letter) for a predetermined purpose from the terminal device 4 used by the observer (teacher). The processing unit 21 also outputs a generation prompt to the AI ​​system 3 instructing it to generate one or more draft documents for a predetermined purpose about the subject based on the received observational facts, and retrieves one or more draft documents from the AI ​​system 3. The processing unit 21 also outputs a modification prompt to the AI ​​system 3 instructing it to modify each retrieved draft document to meet predetermined safety standards, and retrieves each modified draft document from the AI ​​system 3. The processing unit 21 also outputs an evaluation prompt to the AI ​​system 3 instructing it to evaluate each retrieved draft document according to predetermined evaluation criteria, and retrieves the evaluation results from the AI ​​system 3. Finally, the processing unit 21 outputs each retrieved draft document and the evaluation results to the terminal device 4. In this way, the observer can select their preferred draft from the outputted drafts, edit it as needed, and send the selected or edited text to the parent or guardian via email or message. The detailed functions of the processing unit 21 and other functions will be described later using a flowchart.

[0027] The storage unit 22 can be implemented using the computer's memory and storage. The storage unit 22 may also be implemented using external memory resources from the document generation device 2. The storage unit 22 stores user information 221 and log information 222. User information 221 stores, for each user, the user's account (ID, password, etc.) for using the document generation device 2 (document generation service), the user's name, the user's type (observer, approver, administrator, etc.), the user ID of the user's approver, the user ID of the user's recipient, etc. Details of log information 222 will be described later. The storage unit 22 may also store information and data other than the information described above.

[0028] The communication unit 23 is implemented by a computer communication module. The communication unit 23 communicates various data and information via the network 5, connecting with the AI ​​system 3, each terminal device 4, and other computers.

[0029] The terminal device 4 has functional blocks consisting of a processing unit 41, a storage unit 42, a communication unit 43, an input unit 44, and an output unit 45.

[0030] The functions of the processing unit 41 are realized by the computer's processor executing a predetermined program loaded into memory. Some or all of the functions of the processing unit 41 may be implemented as hardware, such as by an integrated circuit. Furthermore, some or all of the functions of the processing unit 41 may be implemented by multiple computers arranged in a distributed manner.

[0031] The processing unit 41 displays various UI screens on the output unit 45, such as a display. The processing unit 41 also receives user input via the input unit 44, such as a keyboard or touch panel, and executes the corresponding processing.

[0032] Figure 2A shows an example of a UI screen (generation request). UI screen 500 receives a request from a teacher to generate a message about a student to be sent to a parent. UI screen 500 includes a student ID, which is the student's identifier, the student's name, the date on which the message is to be generated, the teacher's observations of the student's condition today, class activities, special notes, and a start button to begin generating the message. At a minimum, the items for today's condition, class activities, and special notes correspond to observed facts. In addition to setting text entered by the user in the fields for student ID, name, and date, input data may also be selected from a pre-prepared list using a pull-down menu or the like. In addition to setting text entered by the user as observed facts in the fields for today's condition, class activities, and special notes, the user may also set various types of data (text, documents, images, videos, audio, sensor information, etc. files) by operating a reference button. Observed facts that may be included can be extracted from non-text information such as images.

[0033] When the processing unit 41 receives a start button operation from the user (teacher), it acquires data set in student ID, name, date, today's activities, class activities, and special notes, and sends a request for text generation associated with this data to the text generation device 2 via the communication unit 43.

[0034] Figure 2B shows an example of a UI screen (approval). As will be described in detail later, UI screen 510 is displayed on the approver's terminal device 4 to approve the draft document generated by the document generation device 2. UI screen 510 accepts the approver's approval result for the draft document. The approval result is either approved or rejected for each draft document, and if approved, it may include revisions to the draft document. The approval result may also include the approver's comments. UI screen 510 includes, for each draft document (drafts 1 to 3 in Figure 2B), an Approve / Reject button to select whether to approve or reject, a Modify button to instruct revisions if approved, an explanation such as an evaluation of each draft document, a comment field for entering the approver's comments as text, and a Send button to send the approval result. In addition, a message indicating that a recommended draft document is recommended ("Recommended" in Figure 2B) may be displayed. By operating the Modify button, the draft document is displayed as editable text, and the approver can make revisions.

[0035] When the processing unit 41 receives a submission button operation from the user (approver), it retrieves, for each draft document, whether it has been approved or rejected, the revised draft document if it has been modified, and any comments from the approver if they have been entered. This data is then sent to the document generation device 2 via the communication unit 43 as the approval result.

[0036] Figure 2C shows an example of a UI screen (generation report). UI screen 520 displays and makes available for use one or more draft documents generated by the document generation device 2 (those not subject to approval request, those subject to approval request and approved, or those subject to approval request, approved, and modified). UI screen 520 includes, for each draft document (Draft 1 to Draft 2 in Figure 2C), an edit button for copying and pasting the draft document into the editing field described below, an explanation such as an evaluation of each draft document, an editing field for editing the draft document as text, a distribution button for instructing the distribution of the draft document in the editing field, a copy button for copying the draft document in the editing field to a storage area such as the clipboard, and an end button for ending document generation. In addition, a message indicating that a recommended draft document is recommended ("Recommended" in Figure 2C) may be displayed. When the user operates the edit button, the text of the draft document is displayed in the editing field and can be edited as needed. When a user clicks the copy button, the draft text in the editing field is copied to a storage area such as the clipboard, and can then be used in other applications such as email or messaging apps.

[0037] When the processing unit 41 receives a notification from the user (teacher) to press the distribute button, it retrieves the draft text from the editing field and sends a request to the text generation device 2 via the communication unit 43 to distribute it as the text to be distributed. The recipient of the text may be a parent or guardian associated with the student, based on the student ID etc. set at the time of the generation request, or it may be selectable by the user. When the processing unit 41 receives a notification from the user to press the end button, it terminates the series of processes for generating the text.

[0038] The memory unit 42 can be implemented by the computer's memory and storage. The memory unit 42 may also be implemented by external memory resources of the terminal device 4. The communication unit 43 is implemented by the computer's communication module. The communication unit 43 communicates various data and information by connecting to the document generation device 2 and other computers via the network 5. The input unit 44 is implemented by the computer's input device. The output unit 45 is implemented by the computer's output device.

[0039] Figures 3 and 4 are flowcharts (part 1 and part 2) illustrating an example of the text generation process.

[0040] As a prerequisite, the text generation device 2 provides a UI screen corresponding to the purpose of the text to be generated, and provides the user's terminal device 4 with a UI screen that matches the purpose selected by the user. The purpose of the text can be specified, for example, by the type of document in which the text is used (e.g., contact book, educational support plan, instruction plan, class newsletter, event announcement, report card, student guidance record, recommendation letter for entrance exams, etc.) and the type of text (e.g., a message in a contact book, a general comment in a report card, a recommendation letter in an entrance exam recommendation letter, etc.). Each UI screen, as in UI screen 500, includes the student ID, which is the student's identifier, the student's name, the date the text is generated, and items of observational facts corresponding to the type of text. This flowchart explains the process assuming the generation of a message in a contact book.

[0041] (Step S1) Terminal device 4 displays a UI screen. Specifically, when processing unit 41 receives a command from the user (observer) to display a UI screen for generating a message in the contact book, it accesses the text generation device 2 to obtain the UI screen and displays it on terminal device 4. For example, processing unit 41 sends a request for the URL (Uniform Resource Locator) of the UI screen to the text generation device 2, receives the web page of the UI screen, and displays the UI screen 500 (Figure 2A).

[0042] (Step S2) Terminal device 4 requests text generation. Specifically, when processing unit 41 receives an operation of the start button from the observer on the UI screen 500, it acquires data set in student ID, name, date, today's activities, class activities, and special notes, and sends a text generation request associated with this data to text generation device 2.

[0043] (Step S3) The text generation device 2 accepts requests. Specifically, the processing unit 21 receives text generation requests sent from the terminal device 4 and records them in the storage unit 22 as log information. For example, the date and time of the request, the user ID of the requesting user (observer), and the text generation request (including associated data) are recorded together.

[0044] (Step S4) The text generation device 2 instructs the generation of text. Specifically, first, the processing unit 21 generates a generation prompt that instructs the generation of one or more draft texts for communication messages in the student's communication notebook, based on the received observational facts. The processing unit 21 also outputs the generated generation prompt to the AI ​​system 3, causing the AI ​​system 3 to generate one or more draft texts, and then obtains the output from the AI ​​system 3.

[0045] Figure 5 shows an example of prompts and outputs related to text generation. The generation prompt includes, for example, instructions 101 for generating a draft of text, input 102, constraints 103, output format 104, etc.

[0046] In the example in Figure 5, instruction 101 specifies the role the generating AI should play as an expert (school teacher), the purpose of the document (a notice for parents), and, as a specific instruction, to generate three draft versions of the document based on observed facts. Input 102 contains the items included in the document generation request (date, student ID, observed facts). Observed facts are the text entered in the fields for "Today's situation," "Classroom situation," and "Special notes" on the UI screen 500. If files such as text, documents, images, videos, audio, or sensor information are entered by the user as observed facts on the UI screen 500, these files may also be made accessible in input 102 (multimodal). Constraints 103 set constraints on document generation, including prohibition of exaggeration, prohibition of future assertions, prohibition of excessive evaluation, prohibition of labeling (arbitrary assumptions), punctuation, character count, and output method (JSON format). Output format 104 contains the output format for the three draft versions of the document.

[0047] The character limit for constraint 103 may be set to a number of characters specified by the user on the UI screen. In addition, constraint 103 may include characteristics of each option (for example, "concise," "standard," "specific," etc.) to allow for variations in content. The character limit may also be set for each option. Furthermore, in addition to the items included in the text generation request, information about the subject that has been previously recorded in the memory unit 22 may be set in input 102. If non-text information such as an image is set in input 102, the generation prompt may include instructions to extract observational facts from the non-text information. Alternatively, the processing unit 21 may use a predetermined algorithm or AI system to extract observational facts from the non-text information as text data and set that text data in input 102.

[0048] (Step S5) AI system 3 generates text. Specifically, when the generation AI of AI system 3 receives a generation prompt, it generates one or more draft texts according to the generation prompt and outputs them to the text generation device 2 in the specified output format.

[0049] In the example in Figure 5, the three generated draft documents are set in the `drafts` directory. Each draft document may also have its characteristics defined (e.g., "concise," "standard," "specific," etc.).

[0050] (Step S6) The text generation device 2 instructs the text to be modified. Specifically, first, the processing unit 21 generates a modification prompt instructing the AI ​​system 3 to modify each draft text obtained from the AI ​​system 3 in the previous process (step S4) to meet predetermined safety standards. The processing unit 21 also outputs the generated modification prompt to the AI ​​system 3, causing the AI ​​system 3 to modify each draft text, and then obtains the output from the AI ​​system 3.

[0051] Figure 6 shows an example of prompts and outputs related to document revision. The revision prompt includes, for example, instructions 111 for revising the draft document, input 112, revision policy 113, output format 114, etc. In the example in Figure 6, prompts are shown for revising one draft document, but if multiple draft documents are to be revised, the input should be a list of multiple draft documents and the output format should be a list of multiple draft documents. It is desirable that the internal state of the generating AI after step S5 is executed (including information acquired or generated by the generating AI) be carried over within the generating AI.

[0052] In the example in Figure 6, instruction 111 is set as a specific instruction to revise the draft text according to the revision policy. Input 112 is set as the draft text obtained from the generation AI in the previous process. Revision policy 113 is set as the revision policy, which includes weakening of labeling expressions, deletion of inferential expressions, punctuation, and output method (recording the revised category (NG category)). Weakening of labeling expressions and deletion of inferential expressions correspond to the safety standards that must be met. Output format 114 is set as the output format for the revised draft text and the revised category. The revised category refers to the listed weakening of labeling expressions, deletion of inferential expressions, and punctuation.

[0053] Furthermore, revision policy 113 may include other safety standards such as prohibition of exaggeration, prohibition of future assertions, prohibition of overestimation, and prohibition of labeling. Users may specify which safety standards to set via the UI screen. Additionally, each draft document may be configured to output the corrected words and / or the number of corrected words.

[0054] (Step S7) AI system 3 modifies the text. Specifically, when the generating AI of AI system 3 receives a modification prompt, it modifies each draft text according to the modification prompt and outputs it to the text generation device 2 in the specified output format.

[0055] In the example in Figure 6, as a result of the correction, the corrected draft of the input text is set to `draft_after`, and the corrected category (predicted expression) is set to `guard_flags`. The corrected words and / or the number of corrected words may also be set.

[0056] (Step S8) The text generation device 2 instructs the processing of the text. Specifically, first, the processing unit 21 processes each draft text obtained from the AI ​​system 3 in the previous process (step S6) so that it does not contain sensitive information (sensitive information) about the subject, and generates a processing prompt that instructs the processing unit to output each draft text and the sensitive information before processing (sensitive information separated by processing). The processing unit 21 also outputs the generated processing prompt to the AI ​​system 3, causing the AI ​​system 3 to process each draft text and obtain the output from the AI ​​system 3. Sensitive information includes, for example, information about the subject's health or symptoms, or information about their privacy, such as their family. This information can pose a risk from an information security perspective, for example.

[0057] Figure 7 shows an example of prompts and outputs related to text processing. The processing prompts include, for example, instructions 121 for processing a draft text, input 122, processing policy 123, output format 124, etc. In the example in Figure 7, prompts for processing one draft text are shown, but if multiple draft texts are to be processed, the input should be a list of multiple draft texts and the output format should be a list of multiple draft texts. It is desirable that the internal state of the generating AI after step S7 is executed (including information acquired or generated by the generating AI) is carried over within the generating AI.

[0058] In the example in Figure 7, instruction 121 is set as a specific instruction to process the draft document according to the processing policy. Input 122 is set as the draft document obtained from the generation AI in the previous process. Processing policy 123 is set as the processing policy to separate sensitive information related to home and health from the main text, and the output method (record the separated sensitive information along with the category of sensitive information). Output format 124 is set as the output format for the processed draft document and the separated sensitive information and its category. The category of sensitive information refers to home, health, etc. The separated sensitive information and its category do not necessarily have to be set as output.

[0059] (Step S9) AI system 3 processes the text. Specifically, when the generating AI of AI system 3 receives a processing prompt, it processes each draft text according to the processing prompt and outputs it to the text generation device 2 in the specified output format.

[0060] In the example in Figure 7, the processed draft text is set as safe_text, and the separated sensitive information category (home_context) and sensitive information are set as sensitive_meta. The separated sensitive information and its category do not necessarily need to be set.

[0061] Steps S8 and S9 may be omitted.

[0062] (Step S10) The text generation device 2 instructs the evaluation of the text. Specifically, first, the processing unit 21 generates an evaluation prompt that instructs the AI ​​system 3 to evaluate each draft text obtained in the previous process (step S8, or step S6 if steps S8 and S9 are omitted) according to predetermined evaluation criteria. The processing unit 21 also outputs the generated evaluation prompt to the AI ​​system 3, causing the AI ​​system 3 to evaluate each draft text and obtains the output from the AI ​​system 3.

[0063] Figure 8 shows an example of prompts and outputs related to the evaluation of a document. The evaluation prompt includes, for example, instructions 131 for evaluating a draft document, input 132, evaluation criteria 133, evaluation policy 134, output format 135, etc. In the example in Figure 8, prompts for evaluating three draft documents are shown. It is desirable that the internal state of the generating AI (including information acquired or generated by the generating AI) after the execution of step S9 (or step S7 if the processing of steps S8 and S9 is omitted) is carried over within the generating AI.

[0064] In the example in Figure 8, instruction 131 is set as a specific instruction to evaluate each draft document according to the evaluation criteria and evaluation policy. Input 132 contains each draft document obtained from the generation AI in the previous process, and the legal compliance mode selected by the user via the UI screen or set in advance on the system. The legal compliance mode can be selected, for example, "off" (no evaluation of legal compliance), "light" (mild evaluation of legal compliance), or "strict" (strict evaluation of legal compliance). Evaluation criteria 133 is set as evaluation criteria, including specificity, consistency, psychological stability, word count standard (whether the word count specified in the generation prompt is followed), and legal compliance. Evaluation criteria correspond to evaluation standards. More detailed explanations and definitions may be set for each item of the evaluation criteria. For example, legal compliance may include the prohibition of exaggeration, prohibition of future assertions, prohibition of excessive evaluation, prohibition of labeling, and other rules and laws related to education, as specified in the generation prompt. Evaluation policy 134 sets the following as evaluation policies: select one recommended sentence, output the reason for the recommendation, output the evaluation criteria used as the basis for the recommendation, output whether or not it complies with the law, and the output method (record the recommended sentence, the reason for the recommendation, and the basis for the recommendation, and record the compliance mode and the judgment result). Output format 135 sets the output format of the evaluation result, which includes the number of the recommended sentence, the reason for the recommendation, the basis for the recommendation, the compliance mode and the judgment result.

[0065] You may also configure the prompt to output two or more recommended draft documents. In this case, it is advisable to configure it to assign a ranking to them. You may also configure the prompt to output the compliance mode and judgment result for draft documents other than the recommended ones. Furthermore, you may configure the prompt to assign a rank of three or more levels to the judgment result, rather than a binary value (true or false).

[0066] (Step S11) AI system 3 evaluates the text. Specifically, when the generating AI of AI system 3 receives an evaluation prompt, it evaluates each draft text according to the evaluation prompt and outputs it to the text generation device 2 in the specified output format.

[0067] In the example in Figure 8, the evaluation results show that the number of the recommended draft document is set to "recommended," the reason for the recommendation is set to "reason," the basis for the recommendation is set to "notes," and the compliance mode (strict) and judgment result (true) are set to "compliance." More than one recommended draft document may be output. In this case, a ranking may be assigned. In addition, the compliance mode and judgment result may also be output for draft documents other than the recommended draft document. The judgment result may not be a binary value (true or false), but may be assigned a rank of three or more levels.

[0068] (Step S12) The text generation device 2 performs a risk assessment of the text. Specifically, as an example, the processing unit 21 determines, for each draft text obtained from the AI ​​system 3 in the previous process (step S8, or step S6 if steps S8 and S9 are omitted), whether the degree of modification output from the AI ​​system 3 in step S6 exceeds a predetermined standard. The degree of modification can be, for example, the number of modified words or phrases. The number of modified words or phrases can be determined by comparing the text before and after modification, counting the modified words or phrases output from the AI ​​system 3, or using the number of modified words or phrases output from the AI ​​system 3. If the degree of modification exceeds the predetermined standard, it is determined that there is a risk. Multiple standards may be set to determine not only the presence or absence of a risk, but also the level of the risk.

[0069] As another example, the processing unit 21 determines whether the evaluation result (judgment result) output from the AI ​​system 3 is below a predetermined standard for each draft document acquired from the AI ​​system 3. The evaluation result is, for example, a value indicating a rank. If the evaluation result is below the predetermined standard, it is determined that there is a risk. Multiple standards may be established to determine not only the presence or absence of a risk, but also the level of the risk.

[0070] As another example, the processing unit 21 determines, for each draft document acquired from the AI ​​system 3, whether the number of separated sensitive information pieces output from the AI ​​system 3 in step S8 exceeds a predetermined standard. If the number of separated sensitive information pieces exceeds the predetermined standard, it determines that there is a risk. Multiple standards may be established to determine not only whether there is a risk, but also the level of the risk.

[0071] (Step S13) and (Step S14) Steps S13 and S14 are executed for draft documents that were determined to have a risk in step S12. The document generation device 2 requests approval. Specifically, the processing unit 21 issues and notifies the approver associated with the user (observer) a URL to access, for example, the UI screen 510 (Figure 2B). When the approver accesses the URL via the terminal device 4, the processing unit 21 displays the UI screen 510, which shows the draft documents that were determined to have a risk, on the approver's terminal device 4. The approver, via the terminal device 4, selects to approve or reject each draft document, makes necessary revisions if approved, enters comments if necessary, and operates the submit button. For each draft document, the processing unit 21 obtains the approval or rejection status, the revised draft document if revised, and the approver's comments if entered, and obtains this data as the approval result. Note that there may be more than one approver. The number of approvers may be determined according to the risk level (for example, the number of approvers increases as the level increases).

[0072] (Step S15) and (Step S16) The text generation device 2 reports the generated text. Specifically, the processing unit 21 issues and notifies the user (observer) a URL to access, for example, the UI screen 520 (Figure 2C). When the observer accesses the URL via the terminal device 4, the processing unit 21 displays the UI screen 520 on the observer's terminal device 4, showing one or more generated text drafts (those not subject to approval request, those subject to approval request and approved, or those subject to approval request, approved, and modified). For text drafts selected as recommendations in the evaluation results, a message indicating that they are recommendations is displayed.

[0073] (Step S17) The observer's terminal device 4 uses or distributes text. Specifically, when the processing unit 41 receives an operation of the edit button corresponding to the observer's preferred text draft, it displays the text draft in the editing field. The processing unit 41 also accepts edits to the text draft in the editing field from the observer as needed. When the observer operates the copy button, the text draft in the editing field is copied to a storage area such as the clipboard and used in other applications such as email or messaging. When the processing unit 41 receives an operation of the distribute button from the observer, it retrieves the text draft in the editing field and sends a request to the text generation device 2 to distribute it as a text to be distributed.

[0074] (Step S18) and (Step S19) Steps S18 and S19 are executed if the distribution button is operated in step S17. The text generation device 2 distributes the text. Specifically, the processing unit 21 receives a text distribution request sent from the terminal device 4, issues a URL to the recipient associated with the observer (for example, the student's guardian) to access a UI screen (not shown) that displays the text, and notifies them. When the recipient accesses the URL via the terminal device 4, the processing unit 21 displays the UI screen displaying the text on the recipient's terminal device 4. Alternatively, when the processing unit 21 receives a text distribution request, it may generate a correction prompt (Figure 6) instructing the AI ​​system 3 to modify the text to be distributed so that it meets predetermined safety standards, and have the AI ​​system 3 modify the text to be distributed. The modified text to be distributed may then be distributed to the recipient.

[0075] Next, modified examples of the above-described embodiments will be explained. In the following, each modified example will be described focusing on the differences from the above-described embodiments, and the commonalities will be omitted.

[0076] In one modified example, the text generation device 2 may provide a translation function. The translation function is useful, for example, when a teacher whose native language is a first language (e.g., Japanese) creates text for use by a parent whose native language is a second language (e.g., English).

[0077] Figure 9 is a flowchart showing an example of a text generation process that includes translation. In this modified example, steps S20 and S21 are added after steps S10 and S11 and before step S12.

[0078] As a prerequisite, in step S2, the observer's terminal device 4 receives a specification of the second language to which the draft text will be translated on the UI screen 500 (Figure 2A). The processing unit 21 receives a request from the terminal device 4 for text generation associated with the specification of the second language.

[0079] (Step S20) The text generation device 2 instructs the translation of the text. Specifically, first, the processing unit 21 generates a translation prompt that instructs the AI ​​system 3 to translate each draft text obtained from the AI ​​system 3 in the previous process (step S8, or step S6 if steps S8 and S9 are omitted) into the specified second language and verify its accuracy. The processing unit 21 also outputs the generated translation prompt to the AI ​​system 3, causing the AI ​​system 3 to translate each draft text and obtains the output from the AI ​​system 3.

[0080] Figure 10 shows an example of prompts and outputs related to text translation. Translation prompts include, for example, instructions 141 for translating a draft text, input 142, translation policy 143, output format 144, etc. In the example in Figure 10, prompts are shown for translating one draft text, but if multiple draft texts are to be translated, the multiple draft texts should be listed as input and the multiple draft texts should be listed as output format. It is desirable that the internal state of the generating AI after step S11 (including information acquired or generated by the generating AI) is carried over within the generating AI.

[0081] In the example in Figure 10, instruction 141 is set as a specific instruction to translate the draft text according to the translation policy. Input 142 contains the draft text obtained from the generation AI in the previous process and the target language selected by the user via the UI screen or set in advance on the system. Translation policy 143 is set as the translation policy, including the terminology to be used, the style of the translation, the verification method (comparing the back translation of the original and translated texts to verify the accuracy of the translation and calculating a score indicating accuracy), and the output method (recording the translated text and recording the score indicating accuracy). The score indicating accuracy corresponds to the verification result. More detailed explanations and definitions may be set for each item of the translation policy. For example, the method for verifying the accuracy of the translation and the method for calculating the score indicating accuracy may be included. The verification method is not limited to the back translation method, but may be other methods as well. Output format 144 contains the output format for the translated draft text and the score indicating accuracy.

[0082] (Step S21) AI system 3 translates text. Specifically, when the generating AI of AI system 3 receives a translation prompt, it translates each draft text into a second language according to the translation prompt, verifies its accuracy through back-translation, and outputs it to the text generation device 2 in the specified output format.

[0083] In the example in Figure 10, the translation result shows that the translated draft of the input draft is set to safe_text_translated, and the accuracy score is set to bt_score.

[0084] The processes in steps S12 to S19 are performed on the draft document in the second language.

[0085] In step S12, for example, the processing unit 21 determines whether the verification result (accuracy score) output from the AI ​​system 3 for each draft of a second-language document obtained from the AI ​​system 3 in step S20 is below a predetermined standard. If the verification result is below the predetermined standard, it is determined that there is a risk. Multiple standards may be established to determine not only whether there is a risk, but also the level of the risk.

[0086] In other variations, the text generation device 2 may provide accessibility features (also abbreviated as a11y). Accessibility features are useful, for example, when teachers create documents for use by people with disabilities such as visual impairments, hearing impairments, or intellectual disabilities.

[0087] Figure 11 is a flowchart showing an example of a text generation process that includes accessibility processing. In this modified example, steps S30 and S31 are added after steps S10 and S11 and before step S12.

[0088] As a prerequisite, in step S2, the observer's terminal device 4 accepts the observer's specification of the accessibility processing to be applied on the UI screen 500 (Figure 2A). The specification of accessibility processing includes, for example, text-to-speech, simplified text, and high contrast. With "text-to-speech," the text is processed so that it is less likely to be misread by the software's text-to-speech function; with "simplified text," the text is processed so that it is easy for people of average age under adulthood to understand; and with "high contrast," the text is visually processed so that it is easy for people with visual impairments to read. Of course, accessibility processing is not limited to the above and other processing may be used.

[0089] (Step S30) The text generation device 2 instructs the accessibility processing of the text. Specifically, first, the processing unit 21 generates an accessibility processing prompt that instructs the AI ​​system 3 to process each draft text obtained in the previous process (step S8, or step S6 if steps S8 and S9 are omitted) to improve accessibility. The processing unit 21 also outputs the generated accessibility processing prompt to the AI ​​system 3, causing the AI ​​system 3 to perform accessibility processing on each draft text and obtains the output from the AI ​​system 3.

[0090] Figure 12 shows an example of prompts and outputs related to the accessibility processing of text. The accessibility processing prompt includes, for example, instructions 151 for processing the accessibility of a draft document, input 152, accessibility processing policy 153, output format 154, etc. In the example in Figure 12, prompts are shown for processing one draft document, but if multiple draft documents are to be processed, the input should be a list of multiple draft documents and the output format should be a list of multiple draft documents. It is desirable that the internal state of the generating AI after step S11 is carried over within the generating AI (including information acquired or generated by the generating AI).

[0091] In the example in Figure 12, instruction 151 is set as a specific instruction to process the draft document according to the accessibility processing policy. Input 152 contains the draft document obtained from the generating AI in the previous process and the specification of the accessibility processing to be applied, which is selected by the user via the UI screen or set in advance on the system. Accessibility processing policy 153 is set as the accessibility processing policy, specifying the accessibility processing and the output method (recording the profile before and after processing). The profile is, for example, the score before and after processing for the specified accessibility processing. More detailed explanations and definitions may be set for each item of the accessibility processing policy. For example, the method for calculating the profile may be included. Output format 154 contains the output format for the processed draft document and the profile.

[0092] (Step S31) AI system 3 processes the text to improve accessibility. Specifically, when the generating AI of AI system 3 receives an accessibility processing prompt, it processes each draft text according to the prompt and outputs it to the text generation device 2 in the specified output format.

[0093] In the example in Figure 12, as an accessibility processing result, the accessibility processed draft of the input text is set to safe_text_a11y, and the scores before and after processing (69, 84) are set to a11y_profile.

[0094] The processes in steps S12 to S19 are performed on the document draft that has been processed for accessibility.

[0095] In step S12, for example, the processing unit 21 determines whether the processed score in the profile output from the AI ​​system 3 is below a predetermined standard for each draft document acquired from the AI ​​system 3 in step S30. If the verification result is below the predetermined standard, it is determined that there is a risk. Multiple criteria may be established to determine not only the presence or absence of a risk, but also the level of the risk.

[0096] When combined with the translation function (Figure 9), steps S30 and S31 may be added after steps S20 and S21 and before step S12. In this case, accessibility processing will be performed on the translated draft. Alternatively, steps S30 and S31 may be added after steps S10 and S11 and before steps S20 and S21. In this case, translation will be performed on the accessibility processed draft.

[0097] In further variations, the text generation device 2 may provide a log monitoring function. That is, the text generation device 2 may record logs regarding input and output information with the user and input and output information with the AI ​​system 3, monitor whether a predetermined event has occurred by referring to the recorded logs, and notify the user of information such as alerts if a predetermined event occurs. In addition, for example, log statistics may be made available to the user for viewing in a form such as a dashboard screen. The log monitoring function is useful for verifying, for example, whether the text generation service using the generation AI is functioning properly, whether the user is using the service properly, etc., and for considering improvements to the service or improvements to the usage method.

[0098] Specifically, as described in the above embodiment, the processing unit 21 receives a text generation request sent from the terminal device 4 and records it in the storage unit 22 as log information. The processing unit 21 also records in the storage unit 22 as log information at least a portion of the information input to the AI ​​system 3 at each prompt and at least a portion of the information output from the AI ​​system 3 at each prompt, in relation to the text generation request. The processing unit 21 also records in the storage unit 22 as log information such as risk assessment results, approval results, usage results, and distribution results, in relation to the text generation request. Examples of log information that may be recorded in each step are shown below. (Step S3) The purpose of the text to be generated, the date and time of the request, the requesting user ID, and the text generation request (including associated data). (Step S4) Observational facts input into AI system 3, constraints input into AI system 3, and drafts of each document output from AI system 3. (Step S6) The following information is provided: each draft document entered into AI System 3 (i.e., before revision), the revision policy entered into AI System 3, each draft document output from AI System 3 (i.e., after revision), and the details of the revisions made by AI System 3 (revisioned category, revised word / phrase, number of revised words / phrases). (Step S8) Each draft document input into AI System 3 (i.e., before processing), processing policy input into AI System 3, each draft document output from AI System 3 (i.e., after processing), sensitive information separated by AI System 3 and its category (Step S10) Each draft document entered into AI System 3, evaluation criteria entered into AI System 3, evaluation policy entered into AI System 3, and evaluation results for each draft document by AI System 3 (recommended sentence, reason for recommendation, basis for recommendation, mode of compliance with regulations, and judgment result). (Step S20) The following are the drafts of each document input into AI System 3 (i.e., before translation), the translation policy input into AI System 3, the drafts of each document output from AI System 3 (i.e., after translation), and the verification results of the translation of each draft document output from AI System 3 (score indicating accuracy). (Step S30) Each draft document input into AI System 3 (i.e., before accessibility processing), the accessibility processing policy input into AI System 3, each draft document output from AI System 3 (i.e., after accessibility processing), and the profile of each draft document output from AI System 3. (Step S12) The draft documents to be evaluated, whether each draft document has risks, the level of risk for each draft document, and the indicators used for evaluation (degree of revision, evaluation result, number of separated sensitive information items, etc.) (Step S13) Each draft document to be approved, the approval result for each draft document (approved or rejected, revisions to the draft document, approver's comments), and the approver's user ID. (Step S17) Text used (text copied from the editing field) (Step S18) The distributed text (the text to be distributed), the recipient's user ID.

[0099] Figure 13 is a flowchart illustrating an example of log monitoring processing. Log monitoring processing is performed at predetermined intervals, for example, periodically.

[0100] As a premise, we assume a situation where multiple users (observers) are using the text generation service (i.e., text generation is performed repeatedly). In this situation, log information is accumulated in association with each of the multiple text generation requests, and it is possible to aggregate and analyze it with respect to predetermined alert items and predetermined statistical items.

[0101] (Step S50) The text generation device 2 monitors the logs. Specifically, the processing unit 21 refers to the log information stored in the storage unit 22 and performs aggregations regarding predetermined alert items and predetermined statistical items. For example, the processing unit 21 uses log information from multiple correction prompts to aggregate the number of words that have been corrected so far. For example, it uses log information from multiple evaluation prompts to aggregate the number of times predetermined evaluation criteria have not been met so far. For example, it uses log information from multiple risk judgments to aggregate the number of times a risk has been determined to exist so far. Of course, other items and aggregation methods may be adopted, not limited to the above items and aggregation methods.

[0102] (Step S51), (Step S52), and (Step S53) The text generation device 2 determines whether or not an alert event has occurred. Specifically, the processing unit 21 determines whether or not to notify an alert based on aggregated data of a predetermined alert item and the alert criteria for that alert item. For example, the processing unit 21 determines whether the number of words that have been corrected so far exceeds a predetermined threshold, and if so, notifies the administrator of an alert indicating that there have been many word corrections. For example, it also determines whether the number of items that have not met a predetermined evaluation criterion has exceeded a predetermined threshold, and if so, notifies the administrator of an alert indicating that there have been many violations of the evaluation criterion. For example, it also determines whether the number of items that have been determined to be at risk has exceeded a predetermined threshold, and if so, notifies the administrator of an alert indicating that there have been many items at risk. The method of notifying the administrator is not particularly limited and may be, for example, a predetermined UI screen, email, or message. The administrator can check the content of the alert and consider taking action via the terminal device 4. Note that the recipient of the notification is not limited to the administrator, but may be other users.

[0103] Furthermore, the processing unit 21 may display various aggregated data on the user's terminal device 4 in a form such as a dashboard screen, in response to a request from a user such as an administrator. In this way, the user can check alert items and statistical items at any time.

[0104] In further variations, it goes without saying that the sender (the observer in this embodiment) and the receiver (the recipient in this embodiment) of the text may be swapped. That is, in the embodiment described above, a text was sent from the teacher to the student's guardian, but when creating a response to that text, the guardian may become the sender and the teacher the receiver, and the guardian may use the text generation device 2 to create and use the text. In this case, the text generation device 2 uses a UI screen, prompts, etc., that correspond to the purpose of the text to be generated (a message in response to a message from the teacher).

[0105] An embodiment of this disclosure and its various modifications have been described above. According to this disclosure, it is possible to assist in the creation of appropriate documents in fields such as education. The effects and applications of the embodiment of this disclosure and its various modifications, as well as the scope of application of this disclosure, will be described in detail below.

[0106] According to this disclosure, individuals or organizations using text generation services can build an integrated platform that enables secure, transparent, and consistent text generation, rather than simply using AI to generate text. For clarity, we will explain using an example from the education sector.

[0107] Each prompt sets roles, constraints, and policies appropriate to the expertise of the education field. This ensures that the quality of the text is appropriate for the expertise of the education field. Furthermore, revision prompts can improve the conformity of the generated text to various professional standards. Processing prompts can process sensitive expressions and descriptions in the generated text, improving conformity to professional standards and, in particular, enhancing psychological safety. Evaluation prompts allow for evaluation of whether the generated text is appropriate for the expertise of the education field. Translation prompts and accessibility processing prompts enable the generation of text suitable for diverse audiences (inclusive). Risk assessment and approval require human approval for texts deemed risky, improving the quality and safety of the text as an organization. Log monitoring allows for verification of whether the text generation is functioning properly, enabling consideration of improvements and contributing to external accountability. Furthermore, by enabling input of files such as text, documents, images, videos, audio, and sensor information (for example, content such as communication notebooks, student work, and classroom scenes), in addition to text input by teachers, it becomes possible to generate text based on multimodal information (observed facts).

[0108] While one embodiment and its variations in this disclosure describe the generation of communication documents in the field of education as an example, the scope of application of this disclosure is broad. That is, it can be applied to the generation of text included in documents requiring expertise. For example, this could apply to medical care, welfare, childcare, school education, counseling support, corporate training, etc. In any of these fields, there may be individuals with roles such as observer, target, recipient, approver, and administrator. The basic mechanism of the text generation system is common, but it goes without saying that UI screens and prompts will be provided according to the application field.

[0109] Typically, the system of this disclosure allows a text generation device to prepare various UI screens as web pages and display them in the web browser of a terminal device, as described above. However, the system of this disclosure is not limited to this embodiment. For example, a dedicated application equipped with various UI screens can be installed on the terminal device, and the text generation device can transmit some of the content to be displayed on the various UI screens to the terminal device, thereby allowing the dedicated application on the terminal device to display the various UI screens. In other words, the system of this disclosure can include various implementations that allow screens to be displayed on the terminal device.

[0110] This disclosure is not limited to the embodiments and variations described above, and many further modifications are possible. For example, the embodiments and variations described above are described in detail for the purpose of explaining this disclosure clearly, and are not necessarily limited to having all the configurations described. Furthermore, it is possible to replace parts of one variation with other variations or to combine variations.

[0111] Furthermore, each of the above configurations, functions, processing units, and processing means may be implemented in hardware, either partially or entirely, by designing them as integrated circuits, for example. Alternatively, each of the above configurations and functions may be implemented in software by a processor interpreting and executing programs that implement each function. Information such as programs, tables, and files that implement each function can be stored in memory, recording devices such as hard disks and SSDs, or recording media such as IC cards, SD cards, and DVDs. Also, control lines and information lines are shown only if deemed necessary for explanation, and not all control lines and information lines are necessarily shown in the actual product. In practice, it can be assumed that almost all configurations are interconnected.

[0112] One aspect of this disclosure is a text generation device that can communicate with a generative AI using a large-scale language model and comprises at least one processor and at least one memory resource. The aforementioned processor, A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents for the predetermined purpose relating to the subject, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards, and then obtaining the revised draft documents from the generating AI. An evaluation prompt instructing the generation AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generation AI, and an evaluation process is performed to obtain the evaluation results from the generation AI. The system performs output processing to output each of the acquired draft documents and evaluation results to the terminal device.

[0113] The aforementioned processor, After the correction process and before the evaluation process, the generation AI may be instructed to process each of the acquired draft documents so that it does not contain sensitive information about the subject, and to output the processed draft documents. The generation AI may then perform a processing process to obtain the processed draft documents.

[0114] The aforementioned processor, A determination process to determine whether or not there is a risk related to each of the aforementioned draft documents, The draft document that has been determined to have the aforementioned risk is output to the approver associated with the observer, and an approval process is performed to receive approval or rejection from the approver. In the approval process described above, the approved draft document can be modified by the approver. In the output process described above, each of the approved draft documents may be output to the observer's terminal device.

[0115] The aforementioned processor, In the aforementioned determination process, the presence or absence of risk in each draft document may be determined based on the number of words or phrases that have been modified for each draft document.

[0116] The aforementioned processor, In the aforementioned determination process, the presence or absence of risk in each draft document may be determined based on the evaluation result of each draft document.

[0117] The aforementioned processor, The observer may select one of the outputted draft documents, or edit the selected draft document, and send it to the recipient associated with the observer as the document to be distributed.

[0118] The aforementioned processor, For each of the multiple generation requests, a recording process is performed to log at least a portion of the input / output information between the generation AI and the memory resource, The system may perform a monitoring process that determines whether an alert is necessary based on the log recorded in the memory resource, and if an alert is necessary, outputs the alert to the administrator.

[0119] The aforementioned processor, In the recording process described above, with respect to the correction prompts executed for each of the multiple generation requests, the number of corrected phrases is recorded in the memory resource. In the monitoring process described above, the necessity of the alert may be determined based on the number of modified phrases recorded in the memory resource.

[0120] The aforementioned processor, In the recording process described above, with respect to the evaluation prompt executed for each of the multiple generation requests, the number of times the evaluation criteria were not met is recorded in the memory resource. In the monitoring process described above, the necessity of the alert may be determined based on the number of instances recorded in the memory resource that did not meet the evaluation criteria.

[0121] The aforementioned processor, For each of the multiple generation requests, a determination process is performed to determine whether or not there is a risk related to each of the draft documents, A recording process that logs the number of people determined to have the aforementioned risk to the memory resource, The system may perform a monitoring process that determines whether an alert is necessary based on the number of risks recorded in the memory resource, and if an alert is necessary, outputs the alert to the administrator.

[0122] The aforementioned processor, In the determination process described above, with respect to the correction prompt, the presence or absence of risk in each draft document may be determined based on the number of words or phrases that have been corrected for each draft document.

[0123] The aforementioned processor, In the determination process, with respect to the evaluation prompt, the presence or absence of risk in each draft document may be determined based on the evaluation result of each draft document.

[0124] The aforementioned processor, After the evaluation process, a translation prompt is output to the generating AI instructing it to translate each of the acquired drafts in the first language into the specified second language and verify their accuracy, and a translation process is performed to obtain each of the drafts in the second language and the verification results from the generating AI. Based on the verification results of each draft document in the second language, a determination process is performed to determine whether or not there is a risk, The draft document that has been determined to have the aforementioned risk is output to the approver associated with the observer, and an approval process is performed to receive approval or rejection from the approver. In the approval process described above, the approved draft document can be modified by the approver. In the output process described above, each of the approved draft documents may be output to the observer's terminal device.

[0125] The aforementioned processor, After the evaluation process, an accessibility processing prompt may be output to the generating AI instructing it to process each of the acquired draft documents to improve accessibility, and the generating AI may then perform an accessibility processing to obtain each of the processed draft documents.

[0126] Another aspect of this disclosure is a method for generating text using a text generation device that is capable of communicating with a generative AI using a large-scale language model and comprises at least one processor and at least one memory resource. The aforementioned processor, A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents for the predetermined purpose relating to the subject, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards, and then obtaining the revised draft documents from the generating AI. An evaluation prompt instructing the generation AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generation AI, and an evaluation process is performed to obtain the evaluation results from the generation AI. The system performs output processing to output each of the acquired draft documents and evaluation results to the terminal device.

[0127] Another aspect of this disclosure is a text generation system comprising a text generation device capable of communicating with a generative AI using a large-scale language model, and a terminal device used by an observer. The aforementioned text generation device, The terminal device receives a request for the generation of observational facts about the subject and a document for a predetermined purpose, and Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents for the predetermined purpose relating to the subject, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards, and then obtaining the revised draft documents from the generating AI. An evaluation prompt instructing the generation AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generation AI, and an evaluation process is performed to obtain the evaluation results from the generation AI. The system performs output processing to output each of the acquired draft documents and evaluation results to the terminal device.

[0128] Another aspect of this disclosure is a program that enables communication with a generative AI using a large-scale language model and causes a computer having at least one processor and at least one memory resource to function as a text generation device. A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents for the predetermined purpose relating to the subject, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards, and then obtaining the revised draft documents from the generating AI. An evaluation prompt instructing the generation AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generation AI, and an evaluation process is performed to obtain the evaluation results from the generation AI. Output process to output each of the acquired draft documents and evaluation results to the terminal device. Have the computer execute it. [Explanation of Symbols]

[0129] 1 Text generation system, 2 Text generation device, 3 AI system, 4 Terminal device, 5 Network, 21 Processing unit, 22 Memory unit, 23 Communication unit, 41 Processing unit, 42 Memory unit, 43 Communication unit, 44 Input unit, 45 Output unit, 101 Instructions, 102 Input, 103 Constraints, 104 Output format, 111 Instructions, 112 Input, 113 Correction policy, 114 Output format, 121 Instructions, 122 Input, 123 Processing policy, 124 Output format, 131 Instructions, 132 Input, 133 Evaluation criteria, 134 Evaluation policy, 135 Output format, 141 Instructions, 142 Input, 143 Translation policy, 144 Output format, 151 Instructions, 152 Input, 153 Accessibility processing policy, 154 Output format, 211 Program, 221 User information, 222 Log information, 411 Program, 500 UI screen, 510 UI screen, 520 UI screen

Claims

1. A text generation device capable of communicating with a generative AI using a large-scale language model, comprising at least one processor and at least one memory resource, The aforementioned processor, A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents relating to the subject for the predetermined purpose, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards regarding the expression in the documents, and then obtaining each of the revised draft documents from the generating AI. An evaluation prompt instructing the generating AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generating AI, and an evaluation process is performed to obtain the evaluation results from the generating AI. The process involves outputting each of the acquired draft documents and evaluation results to the terminal device. Sentence generator.

2. A text generation device according to claim 1, The aforementioned processor, After the correction process and before the evaluation process, a processing prompt is output to the generating AI instructing it to process each acquired draft document so that it does not contain sensitive information about the subject, and to output each processed draft document. A processing process is then executed to obtain each processed draft document from the generating AI. Sentence generator.

3. A text generation device according to claim 1, The aforementioned processor, A determination process to determine whether or not there is a risk related to each of the aforementioned draft documents, The draft document that has been determined to have the aforementioned risk is output to the approver associated with the observer, and an approval process is performed to receive approval or rejection from the approver. In the approval process described above, the approved draft document can be modified by the approver. In the output process described above, each of the approved draft documents is output to the observer's terminal device. Sentence generator.

4. A text generation device according to claim 3, The aforementioned processor, In the aforementioned determination process, the presence or absence of risk in each draft document is determined based on the number of words or phrases that have been modified for each draft document. Sentence generator.

5. A text generation device according to claim 3, The aforementioned processor, In the determination process, the presence or absence of risk in each draft document is determined based on the evaluation result of each draft document. Sentence generator.

6. A text generation device according to claim 1, The aforementioned processor, The observer selects one of the outputted draft documents, or edits the selected draft document, and then executes a distribution process to distribute it to the recipients associated with the observer as the document to be distributed. Sentence generator.

7. A text generation device according to claim 1, The aforementioned processor, For each of the multiple generation requests, a recording process is performed to log at least a portion of the input / output information between the generation AI and the memory resource, Based on the logs recorded in the memory resources, the system performs a monitoring process that determines whether an alert is necessary, and if an alert is necessary, outputs the alert to the administrator. Sentence generator.

8. A text generation device according to claim 7, The aforementioned processor, In the recording process described above, with respect to the correction prompts executed for each of the multiple generation requests, the number of corrected phrases is recorded in the memory resource. In the monitoring process, the necessity of the alert is determined based on the number of the modified phrases recorded in the memory resources. Sentence generator.

9. A text generation device according to claim 7, The aforementioned processor, In the recording process described above, with respect to the evaluation prompt executed for each of the multiple generation requests, the number of times the evaluation criteria were not met is recorded in the memory resource. In the monitoring process described above, the necessity of the alert is determined based on the number of instances recorded in the memory resource that did not meet the evaluation criteria. Sentence generator.

10. A text generation device according to claim 1, The aforementioned processor, For each of the multiple generation requests, a determination process is performed to determine whether or not there is a risk related to each of the draft documents, A recording process that logs the number of people determined to have the aforementioned risk to the memory resource, Based on the number of instances where the aforementioned risk is determined to exist, recorded in the aforementioned memory resources, the monitoring process determines whether an alert is necessary, and if an alert is necessary, it outputs the alert to the administrator. Sentence generator.

11. A text generation device according to claim 10, The aforementioned processor, In the aforementioned determination process, with respect to the correction prompt, the presence or absence of risk for each draft document is determined based on the number of words or phrases that have been corrected for each draft document. Sentence generator.

12. A text generation device according to claim 10, The aforementioned processor, In the determination process, with respect to the evaluation prompt, the presence or absence of risk in each draft document is determined based on the evaluation result of each draft document. Sentence generator.

13. A text generation device according to claim 1, The aforementioned processor, After the evaluation process, a translation prompt is output to the generating AI instructing it to translate each of the acquired drafts in the first language into the specified second language and verify their accuracy, and a translation process is performed to obtain each of the drafts in the second language and the verification results from the generating AI. Based on the verification results of each draft document in the second language, a determination process is performed to determine whether or not there is a risk, The draft document that has been determined to have the aforementioned risk is output to the approver associated with the observer, and an approval process is performed to receive approval or rejection from the approver. In the approval process described above, the approved draft document can be modified by the approver. In the output process described above, each of the approved draft documents is output to the observer's terminal device. Sentence generator.

14. A text generation device according to claim 1, The aforementioned processor, After the evaluation process, an accessibility processing prompt is output to the generating AI instructing it to process each of the acquired draft documents to improve accessibility, and an accessibility processing process is executed to obtain each of the processed draft documents from the generating AI. Sentence generator.

15. A text generation method using a text generation device that can communicate with a generative AI using a large-scale language model and comprises at least one processor and at least one memory resource, The aforementioned processor, A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents relating to the subject for the predetermined purpose, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards regarding the expression in the documents, and then obtaining each of the revised draft documents from the generating AI. An evaluation prompt instructing the generating AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generating AI, and an evaluation process is performed to obtain the evaluation results from the generating AI. The process involves outputting each of the acquired draft documents and evaluation results to the terminal device. Sentence generation method.

16. A text generation system comprising a text generation device capable of communicating with a generative AI using a large-scale language model, and a terminal device used by an observer, The aforementioned text generation device, The terminal device receives a request for the generation of observational facts about the subject and a document for a predetermined purpose, and Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents relating to the subject for the predetermined purpose, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards regarding the expression in the documents, and then obtaining each of the revised draft documents from the generating AI. An evaluation prompt instructing the generating AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generating AI, and an evaluation process is performed to obtain the evaluation results from the generating AI. The process involves outputting each of the acquired draft documents and evaluation results to the terminal device. Text generation system.

17. A program that enables a computer, which is capable of communicating with a generative AI using a large-scale language model and has at least one processor and at least one memory resource, to function as a text generation device, A reception process that receives requests from a terminal device used by the observer to generate observational facts about the subject and documents for a predetermined purpose, Based on the observed facts received, a generation prompt is output to the generation AI instructing it to generate one or more draft documents relating to the subject for the predetermined purpose, and a generation process is performed to obtain one or more draft documents from the generation AI. The process involves outputting a correction prompt to the generating AI instructing it to revise each of the acquired draft documents to meet predetermined safety standards regarding the expression in the documents, and then obtaining each of the revised draft documents from the generating AI. An evaluation prompt instructing the generating AI to evaluate each of the acquired draft documents according to predetermined evaluation criteria is output to the generating AI, and an evaluation process is performed to obtain the evaluation results from the generating AI. Output process to output each of the acquired draft documents and evaluation results to the terminal device. A program that causes a computer to execute something.

Images