A sentence generation system using security prompts.

Abstract

To provide a sentence creation system using confidentiality protection prompt for preventing leakage of confidentiality.SOLUTION: In the system according to the present invention, a sentence generation support apparatus connected through a communication network to a user terminal and a generative AI server has an input screen generating unit for generating an input screen to send it to the terminal, a screen input reception unit for receiving text data of an instruction sentence input on the input screen, a confidential information conversion unit for converting the text data of words input from the terminal and designated as confidential information into data subjected to hashing, a prompt generating unit for generating a prompt for input of large scale language model in which the designated words were replaced with hashed character strings, a sentence generation instruction unit for instructing generation of a sentence by transmitting the generated prompt to the server, a confidential information restoring unit for receiving the sentence generated from the server and restoring the hashed character strings to the words before hashing, a generated document output unit for transmitting the an output screen on which the restored sentence is displayed to the terminal, and a control unit for controlling the sentence generating apparatus generally.SELECTED DRAWING: Figure 3

Description

[Technical Field]

[0001] The present invention relates to a sentence generation system using secure prompts, and more particularly to a sentence generation system using secure prompts that prevents security leaks from prompts that are input to a large language model. [Background technology]

[0002] In recent years, the use of generative artificial intelligence (AI) that uses large-scale language models (LLMs) to respond to user questions and inquiries has been rapidly spreading. A large-scale language model is a natural language processing model trained using large amounts of text data, and when a question or instruction is input into a generative AI equipped with a large-scale language model, an answer is output from the large-scale language model. The text input into a large-scale language model is called a "prompt."

[0003] When using a generative AI capable of text processing, a user sends text data of a question via the Internet or a mobile communication network to a server computer that constitutes a text generation system consisting of a trained large-scale language model provided by the company that operates the generative AI, and receives text data of the answer generated by the server via the Internet or a mobile communication network. The sending and receiving of text data is performed using an API (Application Programming Interface) provided by the company that operates the generative AI, or can also be performed using common communication protocols such as HTTP and HTTPS, or can even be done via a web browser.

[0004] However, when sending text containing confidential information (for example, personal information including names and addresses, or information related to trade secrets) to a server computer of a text generation system that is outside of one's control, the leakage of confidential information becomes a problem.

[0005] With generative AI that is publicly available as described above, there is a risk that input information will be used as training data. To avoid this, some businesses allow users to select a setting that prevents input text data from being used for training, but the input text data itself is stored on the server computer that provides the generative AI, and confidentiality cannot be completely guaranteed.

[0006] Furthermore, when text containing confidential information is sent and received between the server computer that provides the generation AI and the text, there is a risk that the confidential information in the text may be leaked or stolen along the communication path.As a method for safely processing such text containing confidential information using external computer resources, Patent Document 1 discloses a method for encrypting the highly confidential text by replacing part of it with a string of characters that is difficult to decipher.

[0007] However, the method of Patent Document 1 requires the user to select the term to be encrypted in advance, associate a predetermined code (character string) with this term, and register it in a database, which forces the user to incur the effort and expense of registering and confirming the word. [Prior art documents] [Patent documents]

[0008] [Patent Document 1] Japanese Patent Publication No. 2020-77134 Summary of the Invention [Problem to be solved by the invention]

[0009] The present invention has been made in consideration of the above-mentioned problems of the conventional technology, and an object of the present invention is to provide a sentence generation system that uses security-protected prompts to prevent the leakage of confidential information from prompts input into an external computer equipped with a large-scale language model. [Means for solving the problem]

[0010] In order to achieve the above object, one aspect of the present invention provides a text generation system using a confidentiality protection prompt, which generates a prompt for inputting a large-scale language model while concealing confidential information contained in an input instruction sentence, the text generation system including a text generation support device connected to a user terminal and a generation AI (Generative Artificial Intelligence) server via a communication network, the text generation support device including an input screen generation unit that generates an input screen displaying a template form for inputting the instruction sentence and transmits the generated input screen to the user terminal, a screen input acceptance unit that receives character data of the instruction sentence input from the user terminal into the template form on the input screen, a confidential information conversion unit that converts character data of a wording designated as confidential information input from the user terminal via the screen input acceptance unit into a hashed character string, and a confidential information conversion unit that converts character data of the instruction sentence input into the template form that corresponds to the wording designated as confidential information into a hashed character string. The device is characterized by including: a prompt generation unit that generates a prompt for inputting a large-scale language model replaced with a hashed character string; a sentence generation instruction unit that sends the generated prompt for inputting a large-scale language model to the generation AI server to instruct sentence generation; a confidential information restoration unit that receives the generated sentence from the generation AI server, detects the hashed character string included in the received sentence, and restores the detected hashed character string to the character data before hashing; a generated document output unit that sends an output screen to the user terminal displaying the generated sentence with the hashed character string restored; and a control unit that controls the sentence generation device overall.

[0011] The confidential information conversion unit stores the hashed character string in a memory unit in association with character data of the wording specified in the confidential information before hashing, and the confidential information restoration unit can obtain and restore the character data of the wording specified in the confidential information before hashing that corresponds to the detected hashed character string from the memory unit. The wording to be designated as confidential information can be input in a confidential information designation field displayed on the input screen transmitted to the user terminal. The template form displayed on the input screen includes multiple input fields with different items, and the prompt generation unit dynamically generates a prompt for inputting a large-scale language model by combining character data entered into each of the multiple input fields of the template form, and generates the input screen on which the generated prompt for inputting a large-scale language model is displayed and transmits it to the user terminal. The standard form includes an input field for specifying the genre of the text to be generated, and the input screen generation unit can generate a prompt confirmation screen including a prompt template associated with the genre name entered in the input field for specifying the genre, and send it to the user terminal. The control unit can generate a usage history for each user that includes a prompt for inputting the generated large-scale language model and character data corresponding to the wording specified as the confidential information, and store the generated usage history in a memory unit. [Effects of the Invention]

[0012] According to the present invention, it is possible to easily generate prompts for inputting large-scale language models in which confidential information is concealed, thereby preventing the leakage of confidential information that may occur when using a generation AI. Furthermore, since prompts for inputting large-scale language models can be dynamically generated using templated input forms, users can easily obtain highly accurate sentence generation results. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a diagram illustrating the overall configuration of a sentence generation system using security prompts according to an embodiment of the present invention. [Figure 2] 1 is a block diagram showing an example of the hardware configuration of a writing generation assistance device and a generation AI server according to the present embodiment. FIG. [Figure 3] 1 is a block diagram showing an example of a functional configuration of a writing generation assistance device according to an embodiment of the present invention; [Figure 4] FIG. 2 is a block diagram showing an example of the functional configuration of a generation AI server according to this embodiment. [Figure 5] FIG. 2 is a diagram showing a document generation processing procedure of the text generation system according to the embodiment of the present invention. [Figure 6] FIG. 10 is a diagram showing an example of an input screen on which a template form for creating a prompt generated by the writing assistance device according to the present embodiment is displayed. [Figure 7] FIG. 10 is a diagram showing an example of a prompt confirmation screen generated by the writing assistance device according to the present embodiment. [Figure 8] FIG. 2 is a diagram showing an example of an original prompt generated by the writing support device according to the present embodiment. [Figure 9] FIG. 2 is a diagram showing an example of a generation result (answer sentence) generated by the writing generation assistance device according to the present embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, specific examples of embodiments of the present invention will be described in detail with reference to the drawings.

[0015] FIG. 1 is a diagram illustrating the overall configuration of a sentence generation system using security prompts according to one embodiment of the present invention. The sentence generation system 1 using security prompts according to this embodiment includes a user terminal 10, a sentence generation support device 20, and a generation AI server 30. The sentence generation support device 20 is communicatively connected to the user terminal 10 and the generation AI server 30 via a communications network. The communications network is, for example, the Internet, and is constructed using a public telephone network, a mobile communications network, a wireless communication path, Ethernet (registered trademark), or the like.

[0016] The user terminal 10 is a computer terminal operated by a user who wishes to create a sentence using a generation AI (sentence generation AI) equipped with a large-scale language model, and includes, for example, a smartphone, a tablet computer, a personal computer, etc.

[0017] A web browser runs on the user terminal 10, allowing access to the writing generation assistance device 20. A user who uses the writing generation system 1 using the confidentiality prompt according to the present invention registers in advance and obtains a user account, user ID (identification number), etc. A method for user registration can be a common method used when using a web service, so a detailed description thereof will be omitted. Also, users can be registered not only as individuals, but also as a company, and members of the company can use the system as a group. In the case of group use, a user ID is assigned to each member of the company, and the system can also be configured to manage the group's usage status.

[0018] The writing generation support device 20 is a computer that exchanges data with the user terminal 10. The writing generation support device 20 and the generation AI server 30 may be configured to be viewed as a single virtual computer from the user terminal 10. The writing generation support device 20 functions as a web server, provides the user terminal 10 with a screen (web page) written in HTML, and allows the user terminal 10 to input and output text related to the writing generation procedure. However, the method of cooperation between the writing generation support device 20 and the user terminal 10 is not limited to this, and processing may be executed via a pre-installed application.

[0019] The generation AI server 30 is a computer equipped with a large-scale language model, which receives prompts as input sentences to the large-scale language model and generates new sentences in response to the received prompts. As an API server, the generation AI server 30 receives prompts specified by the user via the sentence generation support device 20 and generates new sentences (answer sentences) in response to the prompts.

[0020] The writing generation assistance device 20 and the generation AI server 30 may be configured as a general-purpose computer such as a workstation or a personal computer, but are not limited thereto and may be implemented by cloud computing.

[0021] 2 is a block diagram showing an example of the hardware configuration of the writing generation assistance device and the generation AI server according to this embodiment. However, the illustrated configuration is an example and is not limiting. The user terminal 10 may also have a hardware configuration similar to that shown in FIG.

[0022] As shown in FIG. 2, the writing generation assistance device 20 according to this embodiment is configured as a general-purpose computer and includes a CPU (Central Processing Unit) 101, a memory 102, a storage device 103, a communication interface 104, an input device 105, and an output device 106. The storage device 103 stores various data and programs, and is, for example, a hard disk drive, a solid-state drive, or a flash memory. The communication interface 104 is an interface for connecting to a communication network, and is, for example, an adapter for Ethernet (registered trademark), a modem for connecting to a public telephone network, a wireless communication device for wireless communication, or a USB (Universal Serial Bus) connector for serial communication. The input device 105 is a device for inputting data, and is, for example, a keyboard, a mouse, a touch panel, buttons, or a microphone. The output device 106 is a device for outputting data, and is, for example, a display, a printer, or a speaker.

[0023] The generation AI server 30 may also have the hardware configuration shown in Figure 2, but in order to realize higher data processing capabilities, it may be configured as a dedicated computer equipped with multiple GPUs (Graphics Processing Units) in addition to or instead of a CPU.

[0024] The functional configuration of the writing generation support device 20 and the generation AI server 30 according to this embodiment will be described below.

[0025] FIG. 3 is a block diagram showing an example of the functional configuration of the writing assistance device according to this embodiment.

[0026] The writing production assistance device 20 according to this embodiment has the hardware configuration shown in FIG. 3, the writing generation assistance device 20 according to this embodiment includes a control unit 200 configured with a CPU 101 and a memory 102, a storage unit 280 configured with a storage device 103, and a communication unit 290 configured with a communication interface 104. Furthermore, the writing generation assistance device 20 includes an input screen generation unit 210, a screen input reception unit 220, a confidential information conversion unit 230, a prompt generation unit 240, a writing generation instruction unit 250, a confidential information restoration unit 260, and a generated document output unit 270 as functional units realized by the control unit 200 executing a predetermined program.

[0027] In response to an access request from the user terminal 10, the input screen generation unit 210 generates an input screen displaying a fixed form into which an instruction sentence for creating an input sentence (prompt) to the generation AI is input, and transmits the generated input screen to the user terminal 10. Note that the fixed form can include input fields for multiple items.

[0028] The screen input receiving unit 220 receives instructions entered into a standard form on the input screen displayed on the user terminal 10, i.e., character data (text data) for creating a prompt, and transmits the received character data (instructions) for creating a prompt to the prompt generating unit 240 for each input field of the standard form.

[0029] The confidential information conversion unit 230 hashes the character data of the wording designated as confidential information, which is input for each item (field) designated as confidential information from the input screen displayed on the user terminal 10. Hashing can be performed using a general hash function. The hashed character string is associated with the original character data before hashing and stored in the storage unit 280, for example, in the form of a correspondence table. The wording designated as confidential information is input from a predetermined confidential information designation field (details will be described later) displayed on the input screen transmitted to the user terminal 10.

[0030] The prompt generation unit 240 combines (integrates) the character data for creating prompts for each input field of the received standard form to create the original text of the prompt, and replaces the character data that corresponds to words designated as confidential information (proper nouns, addresses, words that constitute confidential matters, etc.) contained in the original text of the created prompt with a hashed string to generate a prompt for inputting into a large-scale language model.

[0031] The sentence generation instruction unit 250 transmits the generated prompt for inputting the large-scale language model to the generation AI server 30 to instruct it to generate sentences.

[0032] The confidential information restoration unit 260 detects hashed character strings contained in the sentence (answer sentence) generated by the generation AI server 30, and restores the detected hashed character strings to character data corresponding to the wording before hashing, based on a correspondence table stored in the memory unit 280.

[0033] The generated document output unit 270 generates an output screen on which the generated text (answer text) is displayed, in which the hashed character string is restored, and transmits the output screen to the user terminal 10.

[0034] The storage unit 280 stores data and programs for image display and input / output control, including standard forms to be displayed on the input screen, designated items for confidential information, etc. Also, a correspondence table between character data of instruction statements entered by users, character data of wording designated as confidential information, and character strings obtained by hashing the character data, is stored for each user (for example, in association with a user account or a user ID).

[0035] The communication unit 290 controls communication with the user terminal 10 and the generation AI server 30.

[0036] FIG. 4 is a block diagram showing an example of the functional configuration of the generation AI server according to this embodiment.

[0037] The generation AI server 30 according to this embodiment is an API server computer equipped with a trained large-scale language model provided by the operator of the generation AI, and has the hardware configuration shown in Fig. 2. It is desirable that the generation AI server 30 has enhanced functionality and performance compatible with large-scale language models.

[0038] As shown in FIG. 4, the generation AI server 30 according to this embodiment includes a control unit 300 including a CPU (GPU) 101 and memory 102, a model storage unit 340 including a storage device 103, and a data communication unit 350 including a communication interface 104. Furthermore, the generation AI server 300 includes a generation request receiving unit 310, a sentence generation unit 320, and a generation result output unit 330 as functional units realized by executing a predetermined program in the control unit 300. Examples of large-scale language models used for sentence generation include ChatGPT-3.5 or ChatGPT-4 provided by OpenAI OPCO, LLC of the United States, and Bard provided by Google LLC of the United States. The generation AI server 30 according to this embodiment can also use API servers provided by these companies.

[0039] The generation request receiving unit 310 receives an API usage request for a large-scale language model. In this embodiment, API usage from the writing generation support device 20 is performed by sending a prompt for inputting a large-scale language model, and the prompt for inputting a large-scale language model sent from the writing generation instruction unit 250 of the writing generation support device 20 is received by the generation request receiving unit 310.

[0040] The sentence generation unit 320 is provided with a predetermined large-scale language model, and generates a response sentence to the sentence included in the prompt by providing the large-scale language model with a prompt for inputting the large-scale language model received by the generation request receiving unit 310.

[0041] The generation result output unit 330 outputs (transmits) the sentence (answer sentence) generated by the sentence generation unit 320 to the request source (transmitter) of the sentence generation assistance device 20.

[0042] The model storage unit 340 stores large-scale language models, parameters, and the like.

[0043] The data communication unit 350 controls communication with the writing assistance device 20 .

[0044] Next, the operation of a sentence generation system using secure prompts according to one embodiment of the present invention will be described.

[0045] 5 is a diagram showing the document generation processing procedure of the writing generation system according to one embodiment of the present invention. Hereinafter, with reference to FIG. 5, the procedure by which the user terminal 10, writing generation support device 20, and generation AI server 30 operate in cooperation will be described.

[0046] When the user terminal 10 sends an access request to the writing generation assistance device 20 (step S100), the control unit 200 of the writing generation assistance device 20 authenticates the user terminal 10, and if the authentication based on the user ID or the like is successful, the control unit 200 causes the input screen generation unit 210 to generate an input screen (see FIG. 6) displaying a template form for creating an input sentence (prompt) to the generation AI, and transmits display data for the generated input screen to the user terminal 10 (step S200). When the input screen is displayed on the user terminal 10 that has received the display data (step S110), the user inputs the genre and confidential information items of the prompt to be created, etc., on the displayed input screen (step S120).

[0047] Fig. 6 is a diagram showing an example of an input screen displaying a fixed form for creating a prompt generated by the writing generation assistance device according to this embodiment. The input screen 400 shown in Fig. 6 displays a fixed form including a title input field 410 for specifying the genre of the prompt to be created, a confidential information item selection field 420, and an additional information specification field 430.

[0048] The title input field 410 may allow the user to enter the genre of the prompt to be created in free text, or may allow the user to select from a pull-down menu of preset genres. The confidential information item selection field 420 may include a list box with preset target items written as field names, allowing the user to select from the list box. The additional information specification field 430 may allow the user to enter text such as a sentence or phrase that contains information to further supplement the instructions for creating the prompt.

[0049] After completing the input into each field 410, 420, 430 on the input screen 400, when the "Execute" button 440 is pressed, prompt creation data consisting of the genre name, confidential information item, and additional information entered respectively into the title input field 410, confidential information item selection field 420, and additional information specification field 430 from the user terminal 10 is sent to the writing generation support device 20.

[0050] The screen input accepting unit 220 of the writing generation assistance device 20 transmits the prompt creation data received from the user terminal 10 to the input screen generating unit 210. The input screen generating unit 210 reads a prompt template (hereinafter simply referred to as a template sentence) associated with the genre specified in the prompt creation data from the storage unit 280 based on the genre name, generates a prompt confirmation screen (see FIG. 7 ) including the read template sentence, and transmits it to the user terminal 10 (step S220). The template sentence is a fixed phrase configured as an instruction sentence containing standard content used in each genre and is stored in the storage unit 280 in advance. Specifiable genres include specifications, inquiries, replies to inquiries, translations, minutes, requests, abstracts, and notifications regarding hiring and sales. The template sentences include shared template sentences with general specifications and user-specific template sentences created by users (including corporate users). The user-specific template sentences can be used only by the user who created them (including group use by corporate users).

[0051] 7 is a diagram showing an example of a prompt confirmation screen generated by the writing generation assistance device according to this embodiment. The prompt confirmation screen 500 includes a title field 510 showing the name of the selected genre, an original text confirmation field 520 showing a template sentence corresponding to the selected genre, an additional information confirmation field 530 showing additional sentences to be combined with the template sentence, and a confidential information specification field 540 for specifying the actual wording of the selected confidential information item.

[0052] The original text confirmation field 520 and the additional information confirmation field 530 are provided with a function that enables text correction such as correction, addition, and deletion of the displayed template sentence, and changes made (input) by the user in these fields are successively transmitted to the screen input acceptance unit 220 of the writing production assistance device 20, and the prompt generation unit 240 dynamically (for each change input) reflects the changes received by the screen input acceptance unit 220 in the template sentence currently being confirmed, creates or updates the original text of the prompt, and redisplays it on the user terminal 10 via the input screen generation unit 210. The prompt generation unit 240 may temporarily store the original text of the created prompt in the storage unit 280.

[0053] The additional information confirmation field 530 displays the context of the additional information corresponding to the information item entered in the additional information specification field 430 of the input screen 400 of FIG. 6. The context includes role settings, conditions, prior information, and knowledge. The confidential information specification field 540 displays input fields 541 and 542 for each field of the target item specified in the confidential information item selection field 420 of the input screen 400 of FIG. 6, and the user may be able to input specific confidential information (e.g., proper nouns of individuals or companies, internal corporate information related to research and development or new business, personal information, etc.) in text form into each input field. The context of the additional information may be stored in advance in the storage unit 280 in association with keywords used to specify the information item.

[0054] After the specific confidential information wording is entered as text in the confidential information specification field 540, when the "Update" button 550 is pressed on the prompt confirmation screen 500, the confidential information wording (character data) entered in the confidential information specification field 540 is sent to the prompt generation unit 240 via the screen input acceptance unit 220 of the text generation support device 20, and the prompt generation unit 240 activates the confidential information conversion unit 230 to hash the character data corresponding to the wording specified as the confidential information, associates the hashed character string (hash value) with the original character data (wording) before hashing, and stores the hashed character string (hash value) in the memory unit 280 for each user.

[0055] Furthermore, the prompt generation unit 240 detects any wording designated as confidential information included in the original text of the prompt, and replaces the detected wording corresponding to the confidential information with a hashed character string. Thereafter, the input screen generation unit 210 generates an input screen (described later) displaying the original text of the prompt after the replacement process, transmits the screen to the user terminal 10, and causes the user terminal 10 to redisplay the screen (step S230). The prompt generation unit 240 also updates the original text of the prompt temporarily stored in the storage unit 280 to the original text of the prompt after the replacement process.

[0056] Fig. 8 is a diagram showing an example of an original prompt generated by the writing generation assistance device according to this embodiment, where (a) in Fig. 8 shows an example of the original prompt before confidential information is hashed (prompt before hashing), and (b) in Fig. 8 is a diagram showing an example of the original prompt in which confidential information has been replaced with a hashed character string (prompt after hashing). Note that the display format of the input screen on which the original prompt shown in Fig. 8 is displayed is not particularly limited, and therefore Fig. 8 shows only a display field for the original prompt.

[0057] An example of the original prompt text shown in Figure 8 illustrates a notification text for a job interview. The pre-hashed prompt shown in Figure 8(a) is created by combining a template text corresponding to the interview notification (a fixed text instructing the user to create an interview notification) with additional text that defines prerequisites. The template text has fields (e.g., 601 to 604) in which any word or phrase can be placed within the text. The user can complete the input text (the original prompt text) simply by rewriting the field parts with the desired phrases, which is highly convenient and greatly improves work efficiency.

[0058] The hashed prompt shown in (b) of Figure 8 shows an example in which the wording of fields (601, 603, 604) designated as confidential information items among any fields placed in the template sentence has been hashed. In this way, confidential information is concealed in the hashed prompt.

[0059] The user checks the original text of the prompt redisplayed on the user terminal 10, and if the user determines that the necessary requirements are satisfied, the user presses the "Confirm" button 560 on the prompt confirmation screen 500 of the user terminal 10. As a result, an instruction to send the confirmed original text of the prompt to the generation AI server 30 is transmitted to the sentence generation instruction unit 250 via the screen input acceptance unit 220 of the sentence generation support device 20 (step S140).

[0060] When the sentence generation instruction unit 250 of the sentence generation support device 20 receives an instruction from the user terminal 10 to send the original text of the prompt to the generation AI server 30, it sends the original text of the prompt after the replacement process, which has been temporarily stored in the memory unit 280, to the generation AI server 30 as a prompt for input to the large-scale language model (step S240). At the same time, the sentence generation instruction unit 250 stores information including the title of the sent prompt, the content (wording) of the confidential information, the user ID, etc., together with the transmission log, as a usage history for each user in the memory unit 280.

[0061] By storing usage history for each user in memory unit 280, users of this system (including corporate users) can centrally manage data so that they can know when, who, what generation AI was used, and what confidential information was included, thereby enabling safe data management.

[0062] When the generation AI server 30 receives a prompt for inputting a large-scale language model from the text generation support device 20 at the generation request receiving unit 310, it transmits the received prompt for inputting a large-scale language model to the text generation unit 320, and the text generation unit 320 generates a new text (called an answer text) in accordance with the received prompt for inputting a large-scale language model.

[0063] The prompts for inputting the large-scale language model received by the generation AI server 30 do not contain confidential information in plain text, so the confidential information is concealed from the generation AI side and can be kept confidential even if it is used as training data. In addition, because the confidential information is replaced with a string of characters consisting of hash values, the large-scale language model recognizes the string of characters consisting of hash values ​​as a single word (proper noun), which has the advantage of being able to output highly accurate generation results (answer sentences) without affecting the understanding of the context.

[0064] When the answer sentence is generated by the sentence generation unit 320, the generation result output unit 330 outputs (transmits) the generated answer sentence to the requesting sentence generation assistance device 20 (step S300).

[0065] When the control unit 200 of the writing generation assistance device 20 receives the generated answer sentence from the generation AI server 30, it activates the confidential information restoration unit 260 to detect the hashed character string included in the generated answer sentence and restore the detected hashed character string to the character data before hashing. A correspondence table stored in the storage unit 280 is used to detect and restore the hashed character string.

[0066] When the restoration of the hashed character string included in the answer sentence is completed, the generated document output unit 270 of the writing generation assistance device 20 generates an output screen displaying the restored answer sentence and transmits it to the user terminal 10 (step S250).

[0067] The user terminal 10 displays the output screen received from the writing production assistance device 20 on the display means of the user terminal (step S150). Note that the user terminal 10 can also acquire the answer sentence as text data from the output screen.

[0068] Figure 9 is a diagram showing an example of a generation result (answer sentence) generated by the writing generation support device according to this embodiment, where (a) of Figure 9 shows an example of a generation result (answer sentence) before the hashed confidential information is restored, and (b) of Figure 9 is a diagram showing an example of a generation result (answer sentence) after the hashed confidential information is restored.

[0069] The generated result (answer sentence) before restoration shown in Figure 9(a) shows that the wording corresponding to the confidential information is displayed as a hash value, and it can be seen that the confidential information was concealed during the sentence generation process and the sentence was generated without affecting the interpretation of the sentence.Furthermore, the generated result (answer sentence) after restoration shown in Figure 9(b) shows that the concealed confidential information has been correctly restored to the original wording.

[0070] As described above, the sentence generation system using confidentiality prompts according to the present invention makes it possible to easily generate prompts for inputting a large-scale language model in which confidential information is concealed, which enables highly accurate sentence generation. Furthermore, since the user does not need to directly send the generated prompts for inputting a large-scale language model to the generation AI and usage history is also saved, the user can easily obtain sentences generated by the generation AI and can safely manage data.

[0071] Although the embodiments of the present invention have been described in detail above with reference to the drawings, the present invention is not limited to the above-described embodiments and can be modified in various ways without departing from the technical scope of the present invention. [Explanation of symbols]

[0072] 1. A sentence generation system using confidential prompts 10 User terminal 20 Sentence generation support device 30 Generative AI Server 101 CPU 102 memory 103 Storage device 104 Communication Interface 105 Input Device 106 Output Device 200, 300 control section 210 Input screen generation unit 220 Screen input reception section 230 Confidential Information Conversion Department 240 Prompt Generation Unit 250 Sentence generation instruction section 260 Confidential Information Recovery Department 270 Generated Document Output Unit 280 Storage section 290 Communications Department 310 Creation request reception unit 320 Sentence generation section 330 Generation result output section 340 Model Memory Unit 350 Data Communications Department

Claims

1. A sentence generation system that generates a prompt for inputting a large-scale language model while concealing confidential information contained in an input instruction sentence, The sentence generation system includes a sentence generation support device connected to a user terminal and a generation AI (Generative Artificial Intelligence) server via a communication network, The sentence generation assistance device an input screen generation unit that generates an input screen displaying a fixed form for inputting the instruction text and transmits the input screen to the user terminal; a screen input receiving unit that receives character data of the instruction sentence input into the fixed form of the input screen from the user terminal; a confidential information conversion unit that converts character data of a phrase designated as confidential information input from the user terminal via the screen input acceptance unit into a hashed character string; a prompt generation unit that generates a prompt for inputting a large-scale language model by replacing character data corresponding to a word designated as confidential information among character data of the instruction sentence input to the template form with the hashed character string; a sentence generation instruction unit that transmits a prompt for inputting the generated large-scale language model to the generation AI server to instruct the server to generate sentences; a confidential information restoration unit that receives the generated text from the generation AI server, detects the hashed character string included in the received text, and restores the detected hashed character string to character data before hashing; a generated document output unit that transmits an output screen to the user terminal, the output screen displaying the generated sentence in which the hashed character string has been restored; a control unit that controls the entire sentence generation assistance device; The template form displayed on the input screen includes a plurality of input fields with different items, the prompt generation unit dynamically generates a prompt for inputting a large-scale language model by combining character data entered into each of the plurality of input fields of the template form, generates an input screen on which the generated prompt for inputting a large-scale language model is displayed, and transmits the input screen to the user terminal; The control unit generates a usage history for each user, including a prompt for inputting the generated large-scale language model and character data corresponding to the wording specified as the confidential information, and stores the usage history in a memory unit.

2. the confidential information conversion unit stores the hashed character string in the storage unit in association with character data of a word specified in the confidential information before hashing; The sentence generation system using the confidentiality protection prompt described in claim 1, characterized in that the confidential information restoration unit obtains and restores character data of the wording specified in the confidential information before hashing that corresponds to the detected hashed string from the memory unit.

3. A sentence generation system using a confidentiality protection prompt as described in claim 1, characterized in that the text to be designated as confidential information is entered from a confidential information designation field displayed on the input screen sent to the user terminal.

4. the template form includes an input field for specifying the genre of the text to be generated; The sentence generation system using a confidentiality protection prompt described in claim 1, characterized in that the input screen generation unit generates a prompt confirmation screen including a prompt template corresponding to the genre name entered in the input field specifying the genre and sends it to the user terminal.

Images