Translation system

Abstract

Provided is a translation system capable of efficiently translating source text, including expressions, the translation control server 72 of the translation system including: an auxiliary information extraction unit 134 for extracting, from the source text, auxiliary information for translation; and a translation execution unit 138 for obtaining, as an output of a large language model, a translated text of the source text into a target language by providing the large language model with the source text, the auxiliary information for translation, and a prompt instructing translation of the source text into the target language.

Description

Translation System

[0001] This invention relates to a translation system using a large language model, and particularly to a technology that enables the translation of literary works such as novels into natural sentences in other languages, which has been difficult to perform appropriately in the past. This application claims priority based on Japanese Application No. 2025-011791 filed on January 28, 2025, and the entire description of the Japanese application is incorporated herein by reference.

[0002] The development of technologies related to recent large language models (LLMs) (hereinafter referred to as "LLMs") is remarkable. LLMs have been confirmed to be effective not only in natural language processing such as summarization and dialogue systems, but also in fields that were thought to have nothing to do with natural language, such as protein structure prediction and function prediction, so-called drug discovery. Furthermore, although most of the training data used for training LLMs is data in a single language and the proportion of data in other languages used is low, it is known that translation using LLMs is possible.

[0003] In order to use such LLMs, it is common for users to interact with LLMs in a so-called chat format. When a user inputs text consisting of natural language such as instructions, questions, examples, etc. into an LLM, an output of the LLM is generated from that text.

[0004] The text that a user gives to an LLM to obtain a desired output is called a prompt. It is known that the output of an LLM varies depending on the prompt. While looking at the output from the LLM, the user devises and changes the prompt to interact with the LLM in order to obtain the desired output.

[0005] On the other hand, literary works such as novels are a field with a large demand for translation. Generally, it is said that it is difficult to translate literary works by machine translation. By using an LLM for the translation of such literary works, there is a possibility of improving the accuracy of the translation of literary works.

[0006] However, translating literary works presents different challenges than translating technical documents. For example, in literary works, it's not simply a matter of translating the meaning of the original text into the target language. How to express the language related to the characters' words, actions, and emotions in the target language is of great importance. Such issues exist not only for characters but also for the objects that appear in the literary work. Furthermore, it is necessary to select appropriate expressions throughout the entire translation that correspond to the expressions of the original text. This is because the translation itself needs to become an independent work in itself.

[0007] Currently, even if one were to use LLM to translate literary works, in the case of novels, simply obtaining an LLM translation would not guarantee an accurate translation. If such a translation were to be manually proofread, the effort involved would be immeasurable, and obtaining an accurate translation overall would be virtually impossible.

[0008] One technology that may help solve such problems is the translation device disclosed in Patent Document 1 below. The translation device disclosed in Patent Document 1 obtains the context corresponding to the sentence to be translated from the source text and obtains specific contextual information contained in that context that is consistent with the sentence to be translated. Examples of specific contextual information include, for example, the topic or subject of the sentence contained in the context in the case of Japanese. Other examples of specific contextual information include various types of information, including grammatical information such as dependency relationships.

[0009] Japanese Patent Publication No. 2022-84305

[0010] According to the technology disclosed in Patent Document 1, it may be possible to estimate omitted words in context and reflect them in the translation. However, the technology disclosed in Patent Document 1 is merely a technique for reflecting the content of the original text in the translation. For translating literary works, including their expression, the technology disclosed in Patent Document 1 is not useful.

[0011] Therefore, the objective of this invention is to provide a translation system that can efficiently translate source texts, including their expression.

[0012] A translation system according to the first aspect of this invention includes an auxiliary information extraction means for extracting auxiliary information for translation from a source text, and a translation execution means for obtaining a translated text of the source text into the target language as an output of the first large-scale language model by providing the source text, the auxiliary information, and a prompt instructing the first large-scale language model to translate the source text into the target language.

[0013] Preferably, the auxiliary information extraction means includes a person information acquisition means for obtaining person information as the output of a second large-scale language model by providing the source text and a prompt instructing the second large-scale language model to extract person information, which is information about people appearing in the source text; a glossary acquisition means for obtaining a glossary as the output of a third large-scale language model by providing the source text and a prompt instructing the third large-scale language model to extract a glossary of terms appearing in the source text; or a summary acquisition means for obtaining a summary as the output of a fourth large-scale language model by providing the source text and a prompt instructing the fourth large-scale language model to create a summary of the source text; or any combination of the person information acquisition means, glossary acquisition means, and summary acquisition means.

[0014] More preferably, the prompt used by the person information acquisition means is a prompt that instructs the creation of person information including the name of a person appearing in the source text, the translation of that name into the target language, gender, and personality.

[0015] More preferably, the prompt used by the glossary acquisition means is a prompt that instructs the creation of a glossary that includes terms appearing in the source text, translations of those terms into the target language, and annotations relating to those terms.

[0016] Preferably, the source text for translation includes a plurality of sequentially arranged source text units, and the translation execution means includes a sentence translation execution means that sequentially provides each of the plurality of source text units to a first large-scale language model a prompt instructing to translate into the target language, and the translation of the immediately preceding source text unit into the target language.

[0017] More preferably, the translation system further includes proofreading prompt generation means for generating prompts for proofreading the translation result by translation execution means through interaction with a user, and translation proofreading means for inputting the prompts generated by the proofreading prompt generation means and the translation result by translation execution means into a large-scale language model for proofreading, thereby obtaining a proofread translation result as the output of the large-scale language model for proofreading.

[0018] The above and other objects, features, aspects and advantages of this invention will be understood in conjunction with the accompanying drawings. They will also become apparent from the following detailed description of this invention.

[0019] Figure 1 is a block diagram showing the overall configuration of a translation system according to the first embodiment of this invention. Figure 2 is a block diagram showing the functional configuration of the translation service front end provided by the translation system shown in Figure 1. Figure 3 is a block diagram showing the functional configuration of the translation control server shown in Figure 1. Figure 4 is a schematic diagram of the screen displayed by the translation control information acquisition unit shown in Figure 2. Figure 5 is a schematic diagram showing the screen for translation displayed by the translation control server shown in Figure 1. Figure 6 is a schematic diagram showing the screen for translation displayed by the translation control server. Figure 7 is a schematic diagram showing the screen for translation displayed by the translation control server. Figure 8 is a schematic diagram showing the screen for proofreading the translated text displayed by the proofreading control server shown in Figure 1. Figure 9 is a block diagram showing the functional configuration of the translation assistance information acquisition unit of the translation control server shown in Figure 3. Figure 10 is a flowchart showing a part of the control structure of the program that implements the person information extraction unit shown in Figure 9. Figure 11 is a flowchart showing another part of the control structure of the program that implements the person information extraction unit shown in Figure 9. Figure 12 is a flowchart showing a part of the control structure of the program that implements the glossary extraction unit shown in Figure 9. Figure 13 is a flowchart showing another part of the control structure of the program that implements the glossary extraction unit shown in Figure 9. Figure 14 is a flowchart showing a part of the control structure of the program that implements the summary extraction unit shown in Figure 9. Figure 15 is a flowchart showing another part of the control structure of the program that implements the summary extraction unit shown in Figure 9. Figure 16 is a flowchart showing the control structure of the program that implements the translation control server shown in Figure 1. Figure 17 is a block diagram showing the functional configuration of the proofreading control server shown in Figure 1. Figure 18 is a flowchart showing the workflow of the proofreading process using the proofreading prompt generation helper shown in Figure 17. Figure 19 is a flowchart showing the control structure of the program that implements the function of generating prompts from an entered prompt option template (Prompt Option Template (POT)) and inputting them into the LLM in the proofreading prompt generation helper shown in Figure 17. Figure 20 is a schematic diagram showing the definitions of prompts and constants used in an embodiment of this invention. Figure 21 is a schematic diagram showing the configuration of the POT used in an embodiment of this invention.Figure 22 is a schematic diagram illustrating the definition options. Figure 23 is a schematic diagram illustrating the semantic options for terms. Figure 24 is a diagram showing the external appearance of the server computer on which the translation system shown in Figure 1 operates. Figure 25 is a block diagram showing the configuration of the server computer shown in Figure 24.

[0020] In the following descriptions and drawings, identical parts are assigned the same reference number. Therefore, detailed descriptions of them will not be repeated.

[0021] First Embodiment 1. Configuration (A) Translation Processing System 50 Figure 1 shows the configuration of the translation processing system 50 according to the first embodiment of the present invention in block diagram form. Referring to Figure 1, the translation processing system 50 includes a translation system 60 that provides machine translation services and machine translation proofreading services via the Internet, an LLM server 64 that provides a natural language chat service using LLM, and a client 62 that uses the machine translation service provided by the translation system 60 to perform the process of machine translating original texts such as literary works and the proofreading process. The LLM server 64 may be a service provided by a different operator than the translation system 60, or it may be a service provided by the same operator as the translation system 60. In this embodiment, machine translation and proofreading of the translation result are performed using the same LLM server 64. However, the present invention is not limited to such embodiments. The LLM used during machine translation and the LLM used during proofreading may be different.

[0022] The translation system 60 includes a translation control server 72 that provides machine translation services using an LLM server 64 to clients 62 that access it via the internet, and a proofreading control server 74 that provides proofreading services for the translated text obtained by machine translation to clients 62. Note that only one client is shown in Figure 1. However, in reality, multiple clients are expected to access the translation system 60.

[0023] The translation system 60 further includes a source / translation storage device 76 connected to the translation control server 72 and the proofreading control server 74, for storing the source text for translation and a translation file containing the source text and the translated text in association with each other, and a translation service front end 70 connected to the translation control server 72, the proofreading control server 74, and a network (Internet) not shown, which functions as a window for providing translation and proofreading services to clients such as client 62 that access via the Internet.

[0024] (B) Translation Service Frontend 70 Referring to Figure 2, the translation service frontend 70 includes a translation control information acquisition unit 100 for acquiring translation control information (hereinafter referred to as "translation control information") such as the document to be translated, the source language, and the target language through dialogue processing with the client 62, and transmitting it to the translation control server 72. The translation service frontend 70 further includes a dialogue translation unit 102 for translating the source text specified by the client 62 into the target language using the translation control server 72 based on the translation control information, and transmitting it to the client 62. The translation service frontend 70 further includes a dialogue proofreading unit 104 for causing the proofreading control server 74 to perform proofreading of the translated text translated by the dialogue translation unit 102 through dialogue processing with the client 62.

[0025] (C) Translation control server 72 Figure 3 shows the functional configuration of the translation control server 72 shown in Figure 1 in block diagram form. Referring to Figure 3, the translation control server 72 includes a control information storage unit 130 for receiving and storing translation control information from the translation control information acquisition unit 100 shown in Figure 2, and a source text receiving unit 132 for receiving the source text file specified by the information from the client 62 via the translation control information acquisition unit 100 of the translation service front end 70 shown in Figure 2, in response to the information about the source text being stored in the control information storage unit 130, and saving it to the source text / translation storage device 76.

[0026] In this embodiment, the translation by the translation control server 72 is performed automatically and continuously, sentence by sentence, based on translation control information, as will be described later. On the other hand, proofreading by the proofreading control server 74 is performed paragraph by paragraph after the translation is completed. Of course, the processing units for translation and proofreading are not limited to this embodiment. Translation may be performed in units other than sentences, such as paragraphs. Similarly, proofreading is not limited to paragraphs, but may be performed in units of sentences or multiple paragraphs.

[0027] The translation control server 72 further accesses the source / target text storage device 76 based on the control information stored in the control information storage unit 130, and transmits the source text and a predetermined prompt to the LLM server 64. The output of the LLM server 64 includes a translation assistance information acquisition unit 134 for acquiring translation assistance information, and a translation assistance information storage unit 136 for storing the translation assistance information acquired by the translation assistance information acquisition unit 134. The contents of the translation assistance information will be described later with reference to Figure 9.

[0028] The translation control server 72 further includes a translation execution unit 138 that uses the translation control information stored in the control information storage unit 130, the translation assistance information stored in the translation assistance information storage unit 136, and the source text and translated text stored in the source text / translation storage device 76 to translate into the target language using the LLM server 64 for predetermined units (in this embodiment, one sentence of the translated text) and stores the translation results in the source text / translation storage device 76, and a translation result transmission unit 140 that pairs the translation result of one sentence by the translation execution unit 138 with the source text input to the translation execution unit 138 and transmits it to the dialogue translation unit 102 as a translation result. The dialogue translation unit 102 transmits this translation result together with the source text to the client 62, as will be described later with reference to Figures 5 to 7, and displays it on the client 62's monitor.

[0029] (D) Control Information Acquisition Screen 160 Figure 4 schematically shows the control information acquisition screen 160 that the translation control information acquisition unit 100 shown in Figure 2 displays on the monitor of the client 62. Referring to Figure 4, the control information acquisition screen 160 includes a target file selection button 180 that the user operates to select a target file, and a text display area 182 that displays the text of the beginning of the file selected by the user. If the file selected here is a source file for translation, the processing requested from the translation system 60 will be to generate the translation. If the selected file is a translation file in which the source and its translation are stored as a pair, the processing requested from the translation system 60 will be proofreading processing of the contents of the translation file.

[0030] The control information acquisition screen 160 further includes a source language selection button 184, which consists of a pull-down menu for selecting the source language, a target language selection button 186, which consists of a pull-down menu for selecting the target language, and a language change button 188 for switching between the specified source and target languages. These buttons are all enabled when the specified file is a source text file. In this embodiment, when the selected file is a translated text file, all of these buttons are grayed out and cannot be operated.

[0031] The control information acquisition screen 160 further includes a "close" button 192 that the user operates when ending use of this service, a save button 194 that the user operates when saving the translated or proofread translated file, and a start processing button 196 that is displayed as "start translation" during translation execution and "start proofreading" during proofreading execution. The "close" button 192 is always operable. The save button 194 cannot be operated until the file to be processed has been specified. The start processing button 196 cannot be operated during translation until control information for translation is obtained. The start processing button 196 cannot be operated during proofreading until the proofreading file has been opened.

[0032] (E) Dialogue screen 220 Figure 5 shows the dialogue screen 220, which is used when translation is performed by the translation control server 72 shown in Figure 1, and also when proofreading is performed by the proofreading control server 74. This screen is created by the dialogue translation unit 102 of the translation service frontend 70 shown in Figure 2 and displayed by the client 62. Referring to Figure 5, the dialogue screen 220 includes a source text display field 230 and a translated text display field 232.

[0033] The source text display area 230 includes a translation target text display unit 252 that displays a sentence from the source text to be translated, and a translated source text display unit 250 that displays the translated source text. Similarly, the translated text display area 232 includes a translation in progress display unit 256 that displays the translated text for the source text displayed in the translation target text display unit 252, and a translated text display unit 254 that displays the translated text. During translation, the translation target text display unit 252 displays a sentence to be translated, and the translated text is displayed in the translated text display unit 254. However, as will be described later, the translation of each sentence is performed automatically during translation. Therefore, the translation target text is first displayed in the translation target text display unit 252, then its translation is briefly displayed in the translation in progress display unit 256, and then both are moved to the translated source text display unit 250 and the translated text display unit 254, respectively, and this process is repeated. Of course, it is not necessary to display the sentence being processed separately from others in this way; the source text and the translated text can be displayed by continuously scrolling along with the other sentences as the translation progresses.

[0034] In this embodiment, the dialogue translation unit 102 in Figure 1 does not immediately translate the first sentence of the text to be translated when translation has started and the first original sentence is displayed on the text to be translated display unit 252. Instead, the dialogue translation unit 102 accepts the user's input of the translation. However, when the second or subsequent sentences are displayed on the text to be translated display unit 252, the dialogue translation unit 102 repeatedly generates the translation and updates the display. This is because, as will be described later, in this embodiment, when translating one sentence at a time, the translation of the immediately preceding original sentence is added to the text to be translated as context and input to the LLM server 64. For the first sentence to be translated, there is no such translated sentence. Therefore, it is possible to input a translation for the first sentence. Of course, the method of adding a translation to the first sentence is not limited to this embodiment. A method of adding a translation to the first sentence of the text to be translated using a general editor may also be used.

[0035] In the example shown in Figure 5, when the first sentence of the source text (indicated by "S-0") is displayed, the user can input the translated text into the translation in progress display unit 256. However, after that, the user cannot input anything further into the translation in progress display unit 256.

[0036] The dialogue screen 220 further includes a translation language pair display 234 that shows the translation language pair and the translation direction. The dialogue screen 220 further includes a "Next" button 236 that is displayed only during proofreading and instructs the system to move on to the next paragraph to be proofread when proofreading of the paragraph to be proofread displayed in the translated text display unit 254 is complete. When the "Next" button 236 is pressed, the paragraph of the source text to be translated and the paragraph of the translated text corresponding to that paragraph, displayed in the translated text display unit 252 and the translated text display unit 256, respectively, move to the translated source text display unit 250 and the translated text display unit 254, and the next paragraph of the source text to be translated is displayed in the translated text display unit 252. At the same time, the translated text corresponding to that paragraph is displayed in the translated text display unit 256 as a target for proofreading.

[0037] The dialogue screen 220 further includes a "Back" button 238, which is displayed only during proofreading and is used to cancel the process and return to the previous state after proofreading of the target translated paragraph is completed and the "Next" button 236 is pressed, and a "Start Proofreading" button 240, which the user operates to start proofreading processing using the proofreading control server 74 shown in Figure 1 for the target translated paragraph displayed on the translation display unit 256 during processing. In this embodiment, the LLM server 64 is used for proofreading the translation. Therefore, the "Start Proofreading" button 240 displays "Proofreading by LLM".

[0038] Figure 5 shows the state of the dialogue screen 220 before starting the translation of the sentence (S-0) shown in the translation target sentence display unit 252 during translation. In Figure 5, the user has entered the translation "twenty years later" for the first source text.

[0039] When a process indicating the input of a translation is performed in this state (for example, by pressing the "input key"), the display on the dialogue screen 220 changes to the state shown in Figure 6. In Figure 6, the first sentence to be translated, "S-0", moves to the translated source text display unit 250, and the next sentence to be translated is displayed in the translation target sentence display unit 252. Similarly, the sentence entered as the translation of the first sentence "S-0" ("twenty years later") moves to the translation display unit 254, and the translation in progress display unit 256 becomes blank. However, in this state, translation candidates for the sentence at the beginning of the translation target sentence display unit 252 are already being generated by the dialogue translation unit 102 and the translation control server 72. When the translation by the translation control server 72 is completed, the translation is displayed in the translation in progress display unit 256, as shown in Figure 7. However, in this embodiment, as described above, the translation proceeds automatically. Therefore, the display shown in Figure 7 is only shown for a brief moment, and after the translation is complete, the original text and its translation are immediately moved to the translated original text display unit 250 and the translated text display unit 254, respectively. This process is repeated until the translation of all sentences is complete.

[0040] During proofreading, the source paragraph to be processed is displayed in the text to be translated display unit 252, and the translated paragraph is displayed in the translated text in processing display unit 256. If the translated paragraph displayed in the translated text in processing display unit 256 is appropriate, the user presses the "Next" button 236. The source paragraph displayed in the text to be translated display unit 252 then moves to the translated source text display unit 250. The next source paragraph is displayed in the text to be translated display unit 252. Similarly, the translated paragraph displayed in the translated text in processing display unit 256 moves to the translated text display unit 254. The next translated paragraph is displayed in the translated text in processing display unit 256. If the translation candidate is deemed inappropriate, the user presses the proofreading start button 240. The proofreading window 280 shown in Figure 8 then pops up.

[0041] (F) Proofreading Window 280 The proofreading window 280 is used to create prompts to be given to the LLM server 64 along with the translation. The proofreading window 280 is generated by the dialogue proofreading unit 104 of the translation service frontend 70 shown in Figure 2, similar to the dialogue screen 220. By sending the prompts created using the proofreading window 280 along with the translated paragraph to be proofread to the LLM server 64, output corresponding to the prompts is obtained from the LLM server 64. By providing the LLM server 64 with appropriate prompts to improve the translated paragraph, the LLM server 64 sends back the improved translated paragraph. The user can proofread each translated paragraph sequentially by repeating this proofreading process. In this embodiment, as will be described later, the proofreading prompts include the source paragraph corresponding to the translated paragraph.

[0042] Referring to Figure 8, the proofreading window 280 includes an option specification field 290 for the user to specify options to be included in the prompt based on the POT, and a prompt display / edit field 292 that displays a prompt in the format specified by the POT, including the options specified in the option specification field 290. In this embodiment, the prompt displayed in the prompt display / edit field 292 may include a free-input item. For the free-input item, the user can freely enter text in the prompt display / edit field 292. In the following figures, the position where the user can enter text is indicated by the string "[Free Input]".

[0043] The proofreading window 280 further includes a confirmation button 296 that the user clicks when sending the prompt displayed in the prompt display / edit field 292 to the LLM server 64. When the confirmation button 296 is pressed, the contents of the prompt display / edit field 292 are checked to see if all the necessary inputs have been made. If there are no errors, the prompt displayed in the prompt display / edit field 292 is sent to the LLM server 64 along with the translated paragraph to be proofread. An example of a prompt in this embodiment will be described later.

[0044] The correction window 280 further includes an LLM output field 294 where the corrected translation paragraph, which is the response of the LLM server 64 to the prompt, is displayed, and a translation copy button 298 that the user clicks to copy the translation paragraph to the clipboard among the outputs of the LLM server 64 displayed in the LLM output field 294. Among the outputs of the LLM server 64, for the translation paragraph, its position can be known by marking the front and back with specific tags. When the translation copy button 298 is clicked, the translation part among the outputs of the LLM server 64 is copied to the clipboard. At the same time, the correction window 280 is erased, and the display returns to the dialogue screen 220 as shown in FIG. 7. When returning to the dialogue screen 220, the content of the clipboard is pasted into the in-processing translation display section 256. As a result, in the in-processing translation display section 256, the corrected translation paragraph is displayed instead of the translation paragraph before correction.

[0045] Referring to FIG. 9, the translation assistance information acquisition unit 134 shown in FIG. 3 includes a person information extraction unit 320 for extracting person information, which is information about the characters appearing in the translation original text, a glossary extraction unit 322 for extracting a glossary, which is information about the keywords appearing in the translation original text, from the translation original text, and a summary extraction unit 324 for extracting the summary from the translation original text. The person information extraction unit 320, the glossary extraction unit 322, and the summary extraction unit 324 all use the LLM server 64 for their processing. The details will be described later.

[0046] The translation assistance information storage unit 136 includes a person information storage unit 340 for storing the person information extracted by the person information extraction unit 320, a glossary storage unit 342 for storing the glossary extracted by the glossary extraction unit 322, and a summary storage unit 344 for storing the summary extracted by the summary extraction unit 324.

[0047] (G) Character Information Extraction Unit 320 FIG. 10 is a flowchart showing a partial control structure of a program that implements the character information extraction unit 320 shown in FIG. 9. Referring to FIG. 10, this program includes a step 380 of generating a prompt for causing the LLM server 64 to create an appearance character, a step 382 of transmitting the prompt generated in step 380 to the LLM server 64, and a step 384 of specifying a program to be activated when there is a reply from the LLM server 64 to the prompt and ending the execution of this program.

[0048] FIG. 11 shows the control structure of the program specified in step 384 of FIG. 10 as a program to be executed when there is a reply to the prompt generated by the process shown in FIG. 10. Referring to FIG. 11, this program includes a step 400 of selecting the text corresponding to the character information from the reply text from the LLM server 64 and storing it in the character information storage unit 340 (see FIG. 9), and ending the process.

[0049] As shown in FIGS. 10 and 11, the processing of the character information extraction unit 320 (see FIG. 9) is performed asynchronously. Also, as will be described later, the processing of the glossary extraction unit 322 and the summary extraction unit 324 is also performed asynchronously. As a result, there is an effect that the character information extraction unit 320, the glossary extraction unit 322, and the summary extraction unit 324 can execute processing in parallel.

[0050] An example of the prompt generated in step 480 of FIG. 10 is as follows. In the following example, the source language for translation is English and the target language for translation is Japanese. Also, as the prompt for the LLM server 64, mainly English is used in this embodiment.

[0051] [Prompt Example 1] Extract person names in the text below. The output format is: <person>Person's name ||| Translation of person's name ||| Person's gender ||| Person's personality< / person> The text is: (Insert the text to be translated here) [End of Prompt Example 1]

[0052] The following is an example of the output from the LLM server 64 in response to prompt example 1, designated as output example 1. [LLM Server 64 Output Example 1] <person> Bob ||| Male ||| Confident, optimistic, and slightly arrogant Jimmy Wels ||| Male ||| Honest, loyal, and slightly sentimental< / person> [End of LLM Server 64 Output Example 1] <person> and< / person> The section enclosed by the symbols contains personal information.

[0053] (H) Glossary Extraction Unit 322 Figure 12 shows the control structure of the program for realizing the glossary extraction unit 322 shown in Figure 9 in flowchart form. Referring to Figure 12, this program includes a step 420 which generates a prompt for the LLM server 64 to create a list of glossaries (keyword lists) with explanations; a step 422 which sends the prompt generated in step 420 to the LLM server 64; and a step 424 which specifies a program to be started when the LLM server 64 responds to this prompt and terminates the execution of this program.

[0054] Figure 13 shows the control structure of the program specified in step 424 of Figure 12 as the program to be started when a reply is received from the LLM server 64. Referring to Figure 13, this program includes step 440, which selects the portion of the reply from the LLM server 64 that matches a predetermined format described later, stores it as a glossary in the glossary storage unit 342 (see Figure 9), and terminates the processing.

[0055] An example of the prompt generated in step 420 of Figure 10 is shown below as Prompt Example 2. [Prompt Example 2] Extract keywords in the text below. The output format is: <keyword> Keywords ||| Keyword Translation ||| Keyword Annotations< / keyword> The text is: (Insert the text to be translated here) [End of prompt example 2]

[0056] The following is an example of the output from the LLM server 64 in response to prompt example 2, which is shown as output example 2. [LLM Server 64 Output Example 2] <keyword> Avenue ||| Street ||| A wide street in a city or town Appointment ||| A meeting or engagement arranged for a specific time and place Cigar ||| Cigar ||| A type of tobacco product< / keyword> The above <keyword> and< / keyword> The section enclosed by the tags is a glossary.

[0057] (I) Summary Extraction Unit 324 Figure 14 shows the control structure of a program for realizing the functions of the summary extraction unit 324 shown in Figure 9. Referring to Figure 14, this program includes a step 480 of generating a prompt for the LLM server 64 to create a summary, a step 482 of sending the prompt generated in step 480 to the LLM server 64, and a step 484 of specifying a program to be started when a reply is received from the LLM server 64 and ending the process.

[0058] Figure 15 shows the control structure of the program specified in step 484 of Figure 14. Referring to Figure 15, this program includes step 490, in which it selects a predetermined portion of the reply text from the LLM server 64 that represents a summary, stores it in the summary storage unit 344 (see Figure 9) as a summary of the document to be translated, and terminates the process.

[0059] An example of the prompt generated in step 480 of Figure 14 is shown below as Prompt Example 3. [Prompt Example 3] Summarize the text below in English. Use the format below. <summary> … < / summary> Here is the text. (Insert the text to be translated here) [End of Prompt Example 3]

[0060] The following is an example of the output from the LLM server 64 in response to this prompt example 3, which is shown as output example 3. [LLM Server 64 Output Example 3] <summary>The story "After Twenty Years" is about a man named Bob who has traveled from the West to New York to fulfill an appointment made twenty years ago with his friend Jimmy Wells. Bob waits in the doorway of a hardware store, which was once the location of a restaurant where he and Jimmy had made the appointment. A policeman on his beat stops to talk to Bob, who explains the situation and shows him a watch and a diamond scarfpin. After the policeman leaves, Jimmy arrives, but it is revealed that he is actually a patrolman who has been tasked with arresting Bob, who is wanted in Chicago. Jimmy had recognized Bob when he lit his cigar and had gone to get a plainclothesman to make the arrest. The story ends with Bob reading a note from Jimmy explaining the situation.< / summary> [End of LLM Server 64 Output Example 3]

[0061] (J) Single-Sentence Machine Translation In this embodiment, machine translation by the translation control server 72 is performed on a sentence-by-sentence basis. The LLM server 64 is also used in the machine translation of each sentence. Figure 16 shows the control structure of the program that realizes single-sentence machine translation by the translation control server 72. Referring to Figure 16, this program includes step 500, which reads the contents of the file to be translated selected by the selection button 180 shown in Figure 3 and displays the beginning portion thereof on the dialogue screen 220 as shown in Figure 5. In Figure 5, the translated text is shown in the translation display unit 256 during processing, but the translated text is not yet displayed at the time of step 500.

[0062] The program further includes, following step 500, step 502, which accepts input of a translation for the beginning of the source text to be translated, and step 504, which generates a string by combining the beginning of the source text and the translation entered in step 502.

[0063] This program further includes step 506, which involves repeatedly executing the sentence translation process 508 while increasing the value of a variable N, which indicates the number of the source text being processed, by 1 from 0 until N > number of source texts - 1.

[0064] The sentence translation process 508 includes a step 520 that generates a prompt for the LLM server 64 to translate the N+1th source text. An example of the prompt generated in step 520 is shown below as Prompt Example 4. This prompt is for translating the second sentence (N=1) of the source text (indicated by "S-1"). The first sentence of the source text (N=0) is indicated by "S-0".

[0065] [Prompt Example 4] I want to translate a book into Japanese. Here are the persons in the book in the format: <person> (Insert extracted person information here)< / person> Here are keywords: <keyword> (Insert extracted glossary information here)< / keyword> Here is a summary: <summary> (Insert extracted summary information here)< / summary>Translate S-1 to a faithful and natural Japanese text with the same meaning of S-1. The output format is: <translation> Your Japanese translation of S-1.< / translation> Here is the part of the original book in <context> <context>S-0: AFTER TWENTY YEARS ||| Twenty Years Later S-1: The policeman on the beat moved up the avenue impressively. S-2: The impressiveness was habitual and not for show, for spectators were few. (Insert original text below) ...< / context> [End of Prompt Example 4]

[0066] In other words, this prompt is for person information ( <person> and< / person> (The part enclosed by) Glossary information ( <keyword> and< / keyword> (The portion enclosed by the symbols), and summary information ( <summary> and< / summary> A fixed part including the part enclosed by , a formatting part specifying the format of the text to be translated, and a context part including the entire source text to be translated ( <context> and< / context> The context section includes the part enclosed by and . The context section includes the sentence to be translated (indicated by the label "S-1" in the example above), the already translated entire text (indicated by the label "S-0" in the example above), and all subsequent sentences to be translated (indicated by labels "S-2" and below). Each translated sentence includes the original text and its translation. In this embodiment, translation is performed sentence by sentence, advancing through the sentences to be translated one by one. During this process, the translated original text is accompanied by its translation.

[0067] Referring again to Figure 16, the sentence translation process 508 further includes step 522 of sending the prompt generated in step 520 to the LLM server 64, and step 524 of receiving a reply from the LLM server 64 to the prompt. The sentence translation process 508 further includes step 526 of combining the N+1th source text with the translation received in step 524 for the prompt when translating the next sentence, and step 528 of updating the text displayed in the source text display field 230 and the translation text display field 232 in the dialogue screen 220 shown in Figure 5.

[0068] The sentence translation process 508 further includes a step 530 in which the value of variable N is increased by 1 and the process of the sentence translation process 508 is terminated.

[0069] The program further includes the steps of displaying a translation completion message on the interactive screen 220 after the completion of step 506, and saving a translation result file containing each sentence of the source text paired with each sentence of the translated text, and terminating the program execution, upon receiving a termination confirmation instruction from the user.

[0070] (K) Proofreading Control Server 74 Figure 17 is a block diagram showing a more detailed functional configuration of the proofreading control server 74 shown in Figure 1. Referring to Figure 17, the proofreading control server 74 includes a proofreading prompt generation helper 550 that assists in generating proofreading prompts (hereinafter referred to as "proofreading prompts") to be given to the LLM server 64 for the translated text acquired by the translation control server 72 (see Figure 1) and stored in the source / target text storage device 76, and a proofreading prompt generation unit 552 that generates proofreading prompts in cooperation with the proofreading prompt generation helper 550. The translation service frontend 70 and the proofreading prompt generation helper 550 are realized by a browser that operates on a client 62 that communicates with the translation control server 72 and the proofreading control server 74, and a script that operates on the browser of the client 62, which is sent from the proofreading control server 74 to the browser of the client 62 via the proofreading control server 74 translation service frontend 70 and executed by the browser.

[0071] The calibration control server 74 further includes a prompt storage unit 560 for temporarily storing calibration prompts generated by the calibration prompt generation helper 550 and the calibration prompt generation unit 552, and an LLM access unit 562 that transmits the prompts stored in the prompt storage unit 560 to the LLM server 64 and provides the text of the response from the LLM server 64 to the calibration prompt generation helper 550.

[0072] The calibration control server 74 further includes a machine-readable template storage unit 554 connected to the calibration prompt generation unit 552 and storing prompt option templates (POTs) that pre-store options for generating calibration prompts as templates; a program storage unit 556 connected to the calibration control server 74 and storing programs required by the calibration prompt generation unit 552 and the calibration prompt generation helper 550 to provide calibration services, browser-executable scripts and CSS (Cascading Style Sheets) used by the browser, etc., required by the client 62 to use the calibration services; and a library storage unit 558 connected to the calibration prompt generation unit 552 and storing program libraries required when executing programs and scripts stored in the program storage unit 556.

[0073] In this embodiment, the template storage unit 554, the program storage unit 556, and the library storage unit 558 are located in the same translation system 60 as the proofreading control server 74. However, the invention is not limited to such embodiments. The template storage unit 554, the program storage unit 556, and the library storage unit 558 may be located on any node on the internet, or on a remote system connected to the proofreading control server 74 by a private communication channel, as long as they can communicate with the proofreading control server 74.

[0074] The template storage unit 554 stores prompt templates (POTs) that include various options that can be specified when creating a proofreading prompt. The general configuration of these POTs will be described later. Note that the same POT may be provided for multiple languages. In this embodiment, a single POT is used for translation between Japanese and English, but this invention is not limited to such embodiments. For example, language-specific POTs may be prepared, and a POT suitable for a combination of translation languages ​​may be created by combining the language POTs according to the combination of translation languages.

[0075] (L) Prompt Creation Process Figure 18 shows the flow of the prompt creation process using the proofreading window 280 shown in Figure 8 in flowchart format. Referring to Figure 18, this process includes step 580, which displays the proofreading window 280 shown in Figure 8 at the start of the process. The LLM output field 294 of the proofreading window 280 shown in Figure 8 displays the paragraph of the translated text to be processed. The option specification field 290 displays the option string of the prompt according to the POT sent from the proofreading control server 74. The option string will be described later. Initially, the prompt display / edit field 292 displays the prompt string according to the POT, which has been edited according to the default value of the options. In the initial stage, the user is dissatisfied with the translated text displayed in the LLM output field 294, so the LLM proofreading process is started.

[0076] There are several possible actions the user can take after this point. First, there is the option of trying to perform LLM calibration using the initial value of the prompt displayed in the prompt display / edit field 292. In this case, the determination in step 582 of Figure 18 will be positive ("Creation complete," i.e., "Use the displayed prompt"), and the process will proceed from step 582 to step 584.

[0077] In step 584, the user presses the confirmation button 296 shown in Figure 8. The process proceeds to step 586. In step 586, the calibration prompt generation helper 550 shown in Figure 17 outputs the prompt displayed in the prompt display / edit field 292 shown in Figure 8 to the prompt storage unit 560. Normally, the prompt storage unit 560 is memory. The LLM access unit 562 reads the contents of the prompt storage unit 560 and sends it to the LLM server 64. In response to this prompt, the LLM server 64 generates output text configured according to the prompt's instructions and sends it back to the LLM access unit 562 of the calibration prompt generation helper 550. The LLM access unit 562 sends this text back to the calibration prompt generation helper 550.

[0078] In step 588 of Figure 18, the proofreading prompt generation helper 550 displays this text in the LLM output field 294 shown in Figure 8. Typically, the output text of the LLM server 64 includes the translated paragraph, the source paragraph, and comments generated by LLM that are included in the prompt. Of these, the translated paragraph and the source paragraph are each enclosed by predetermined tags. The user examines the source paragraph, translated paragraph, and LLM comments displayed in the LLM output field 294. Based on the results, in step 590, the user determines whether proofreading of the translated paragraph is complete. If the user believes that proofreading of the translated paragraph is complete (the determination in step 590 is positive), the user clicks the translated copy button 298 displayed in the LLM output field 294 in Figure 8 in step 590 of Figure 18. The translated paragraph can be identified from the position of the tag indicating the translated text in the output of the LLM server 64. Therefore, only the text of the corrected translated paragraph is copied to the clipboard. This text is used to replace the uncorrected translated paragraph when proofreading of the translated paragraph being processed is complete.

[0079] If the judgment in step 590 is negative (the user is not satisfied with the content of the translated paragraph), the user operates on the options displayed in the option specification field 290 in step 596 without clicking the translated copy button 298 in Figure 8. When an option is operated on and changed, the content is immediately reflected in the prompt. That is, the prompt is edited according to the changed value of the option, and the content of the prompt displayed in the prompt display / edit field 292 changes according to the editing result. If the user attempts proofreading using this prompt, the judgment in step 582 becomes positive, and the user clicks the confirmation button 296 in step 584. As a result, the proofreading result by the LLM server 64 for the new prompt is displayed in the LLM output field 294 in Figure 8 after steps 586 and 588.

[0080] If the user is satisfied with the corrected translated paragraph displayed in the LLM output field 294, the judgment in step 590 becomes positive. As a result, in step 592, the user presses the copy translation button 298. The proofreading window 280 is closed, and the dialogue screen 220 shown in Figures 5 to 7 is displayed. The translated paragraph to be proofread is replaced with the corrected translated paragraph and displayed. If the user wishes to proceed to proofreading the next translated paragraph, they should click the "Next" button 236.

[0081] On the other hand, in step 590 of Figure 18, if the user wishes to further edit the prompt, the determination in step 590 becomes negative. The user specifies further options in step 596, and the prompt is edited according to the result. The process returns to step 590, and the edited prompt is displayed. The same process is repeated until the determination in step 590 becomes positive.

[0082] Figure 19 is a flowchart showing the control structure of the program that implements step 586 in Figure 18. This program is activated in response to the confirmation button 296 shown in Figure 8 being pressed.

[0083] This program includes the steps of: generating a single prompt by concatenating the text of individual prompts displayed in the prompt display / editing field 292; sending the prompt generated in step 620, along with the translated paragraph to be proofread and the corresponding source paragraph, to the LLM server 64 in step 622; and terminating the execution of the program by extracting the proofread translated paragraph included in the response output by the LLM server 64 in response to this prompt and displaying it in the LLM output field 294 in Figure 8.

[0084] (M) Prompt Definition and POT Figure 20 shows an example of a prompt definition 640. Referring to Figure 10, the prompt definition 640 includes a prompt structure 650 and a definition 652 of variables used within the prompt.

[0085] According to the prompt structure 650, the prompt includes an option 660 for defining terms within the prompt, an option 662 for monitoring the internal state of the language model, an option 664 for ensuring comprehensive proofreading, an option 666 for facilitating understanding of the source text, and an option 668 for specifying various translation conditions. The prompt structure 650 further includes an option 670 for specifying various formats and a target text 672 consisting of the source text to be translated and the target text to be proofread. The prompt is generated by concatenating texts that indicate these options.

[0086] Variable definition 652 indicates how variables used internally in the POT are replaced with values ​​according to conditions when the actual prompt is generated. In each item of variable definition 652, the description within the POT indicates that the string to the left of the equals sign will be replaced by the string or value to the right of the equals sign. In this way, prompts can be created using the same POT even for combinations of different languages.

[0087] Generally, creating effective prompts for proofreading translations requires numerous trials and errors. Moreover, it is difficult for users to judge the quality of the translations obtained through such work. Therefore, in this embodiment, a researcher with both linguistic and LLM knowledge prepares various options to be included in the prompts as a template (POT) based on their extensive experience. By using this template, users can easily select and combine options consisting of appropriate wording to bring out the performance of the LLM, or easily modify options to suit the target language, thereby creating effective prompts with minimal trial and error. Note that what kind of prompt is effective varies depending on the LLM and the language used. Therefore, it is desirable to prepare a POT individually according to the LLM being used.

[0088] It is desirable that the optional languages ​​that make up the prompts and the language used for the user interface be independent of each other. These can be switched, for example, on the dialogue screen 220 or the proofreading window 280. Also, depending on the type of LLM server 64, a prompt in a specific language (e.g., English) may perform better. In such cases, even if a language other than that specific language (e.g., Japanese) is used for the user interface, it is conceivable that the prompts to the LLM server 64 should use that specific language. Therefore, it is desirable to store information in both the language for the user interface and the language for the prompts in the POT.

[0089] The following is a general overview of each option in the prompt structure 650, but details of each option will be described later with reference to Figures 22 and 23.

[0090] (a) Option 660 for Defining Terms As shown in Figure 20, option 660 for defining terms includes DEFINITION. DEFINITION is not specified by the user and is always used at the beginning of the prompt. Note that DEFINITION may differ depending on the type of LLM server 64.

[0091] (b) Option 662 for monitoring the internal state of the language model The option 662 for monitoring the internal state of the language model includes CHECK and TONE. CHECK is an option to query the internal state of the language model (LLM server 64) (for example, whether there are any words whose meaning the LLM server 64 is unsure of). In the next proofreading round, the user will be able to provide the language model with appropriate translations. As a result, the success rate of corrections over multiple turns will increase. TONE instructs the language model to understand the entire source text before correcting the translation. Specifying TONE will increase the success rate of corrections.

[0092] (c) Option 664 for ensuring comprehensive proofreading Option 664 for ensuring comprehensive proofreading includes AMBIGUITY. AMBIGUITY is an option necessary to extract all answers from the language model when there is ambiguity in the source text. Option 666 for facilitating understanding of the source text includes CONDITION. CONDITION specifies paraphrasing of expressions. CONDITION likely helps to deepen the language model's understanding of the source text.

[0093] (d) Options 668 for specifying various translation conditions In this embodiment, the options 668 for specifying various translation conditions include CONTEXT, STYLE, ATTENTION, GENRE, SENSE, and DICT. These are options that indicate the usual information to be considered in translation. Other options may also be included in the options 668 for specifying various translation conditions as needed. As will be described later, some items can be freely entered.

[0094] (e) Options 670 for specifying various formats Options 670 for specifying various formats are options for specifying the output format or output language.

[0095] (f) Target text 672 The target text 672 includes the source text for translation and the translated text to be proofread. Each of these is labeled with a predetermined label.

[0096] These options are desirable elements for use in high-precision calibration. As described below, the POT includes the information necessary to generate these options, except for the target text 672, and generates prompts to be given to the LLM server 64 as specified by the user.

[0097] (N) POT700 Referring to Figure 21, in this embodiment, POT700 is stored in the template storage unit 554 shown in Figure 17 in HTML (HyperText Markup Language) format and supplied from the template storage unit 554 to the proofreading prompt generation unit 552 and further to the proofreading prompt generation helper 550. The template storage unit 554 may store the POT itself, or it may store only the information necessary for creating the POT, and a program stored in the program storage unit 556 may read this information according to the combination of languages ​​to be translated and create a POT in the appropriate format.

[0098] This POT HTML file contains embedded information for reading scripts and CSS stored in the program storage unit 556, as well as information for reading predetermined libraries stored in the library storage unit 558. The calibration prompt generation helper 550 reads this information, renders and displays this HTML file in the option specification field 290 shown in Figure 8, and edits the prompt by executing the script associated with the element of the HTML file that the user manipulated.

[0099] When processing HTML using scripts, it is common to represent the structure of the HTML file using a DOM (Document Object Model). User Interfaces (UIs) designed to receive user input include checkboxes, radio buttons, free text input fields, and add buttons for copying (adding) specific elements. Each of these UI elements is assigned a unique identifier.

[0100] Referring to Figure 21, POT 700 includes DEFINITION generation information 710, CHECK generation information 712, TONE generation information 714, AMBIGUITY generation information 716, CONTEXT generation information 718, CONDITION generation information 720, STYLE generation information 722, ATTENTION generation information 724, GENRE generation information 726, SENSE generation information 728, DICT generation information 730, NOTE generation information 732, FORMAT generation information 734, and RESPONSELANG generation information 736. These will be described in detail below. In the following explanation, options for translation between Japanese and English will be described.

[0101] (a) DEFINITION generation information 710 Referring to Figure 22, the DEFINITION generation information 710 includes an English string 760 and a Japanese string 762 that constitute the beginning of the prompt. When the prompt is in English, the English string 760 is selected as the beginning of the prompt, and when it is in Japanese, the Japanese string 762 is selected, and these are displayed in the prompt display / edit field 292 shown in Figure 8. Also, when the UI is in English, the English string 760 is displayed in the option specification field 290, and when it is in Japanese, the Japanese string 762 is displayed. The user does not intervene in this selection. As mentioned above, this string may be changed depending on the type of LLM server 64.

[0102] This string is added to the prompt by substituting it into the position (variable) of the term definition option 660 in the prompt structure 650 shown in Figure 20.

[0103] In this example, both the English string 760 and the Japanese string 762 are displayed in the option specification field 290 shown in Figure 8. However, the invention is not limited to such embodiments. When the language of the user interface is English, only the English string 760 may be displayed in the option specification field 290, and when it is Japanese, only the Japanese string 762 may be displayed in the option specification field 290. The generated prompts are selected from either English or Japanese depending on the type of LLM server 64. The language of the user interface for creating these prompts and the language of the prompts themselves are the same as for the other options described below.

[0104] (b) CHECK generation information 712 The CHECK generation information 712 includes options that include English prompts and Japanese prompts with checkboxes. The Japanese prompts are in natural language, for example, "Please check whether you are confident in the meaning of all terms. If not, please tell me the meaning or translation of the terms so that I can specify them in the next proofreading. In either case, please correct 'machine translation' based on your understanding."

[0105] This option is displayed in the option specification field 290 in Figure 8. In the display of the option specification field 290, a checkbox is displayed at the beginning of each option in the CHECK generation information 712. If the checkbox is checked, this option is added to the prompt. When the prompt is in English, the English string of the option is added to the prompt. When the prompt is in Japanese, the Japanese string of the option is added to the prompt. This addition is performed by assigning it to the variable indicated as "CHECK" in the monitoring option 662 for the internal state of the language model of the prompt structure 650 shown in Figure 20.

[0106] The CHECK generation information 712 may also include a cautionary note displayed in the option specification field 290, accompanying the display portion of the option. If a cautionary note exists, it is displayed in the option specification field 290 separately from the prompt itself. The cautionary note includes an explanation of the possible effects of adding the option to the prompt. These effects are based on the results of various experiments conducted using the target LLM server 64. The same applies to cautionary notes accompanying other options.

[0107] As mentioned above, the cautionary notes may include specific cautionary notes regarding the target option, or they may be left blank. This also applies to the cautionary notes for the other options described below.

[0108] (c) TONE generation information 714 The TONE generation information 714 in Figure 21 includes English and Japanese TONE option strings. In the case of Japanese, the option string is a sentence such as, for example, "Please estimate the tone of the 'original text'." A checkbox is provided at the beginning of the TONE option string displayed in the option specification field 290. The TONE generation information 714 may also include a note that is displayed in the option specification field 290 in conjunction with the display of the TONE option string. When this option is selected, either the English or Japanese string from the TONE option string is assigned to the variable indicated by "TONE" in the prompt structure 650 shown in Figure 10.

[0109] (d) AMBIGUITY generation information 716 The AMBIGUITY generation information 716 includes an English and Japanese AMBIGUITY option string. This string is, for example, "Please determine if the "original text" is ambiguous. If the "original text" is ambiguous, please provide the following answers for all meanings." In the option specification field 290 shown in Figure 8, a checkbox is displayed at the beginning of the AMBIGUITY option string. When this option is selected, the English or Japanese string from the AMBIGUITY option string is assigned to the variable indicated by "AMBIGUITY" in the prompt structure 650 shown in Figure 10. Note that in the above string, "below" refers to the instructions given by the series of options following this option.

[0110] (e) CONTEXT generation information 718 The CONTEXT generation information 718 includes a CONTEXT option string in Japanese and English. In the option specification field 290 shown in Figure 8, either the Japanese or English version of this CONTEXT option string is displayed. A check box is displayed at the beginning of this string. If this check box is checked, the Japanese or English CONTEXT option string is added to the prompt.

[0111] The CONTEXT option string can include a "free input field" where the user can freely enter the translation context. When the checkbox at the beginning of the CONTEXT option string is checked, the user can enter text in the prompt display / editing field 292 in Figure 8, specifically in the position indicating free input (for example, the position where "[Free Input]" is displayed). In the prompt structure 650 in Figure 20, the CONTEXT option string containing the text entered by the user is assigned to the variable indicated by "CONTEXT".

[0112] However, this invention is not limited to such embodiments. For example, in the option specification field 290, a window for free input may be popped up in response to the checkbox for this option being checked. When the user enters some string in that window, the string obtained by replacing the "[Free Input]" portion of the CONTEXT option string with that string is added to the prompt in the prompt display / edit field 292, and is also added to the prompt.

[0113] (f) CONDITION generation information 720 The CONDITION generation information 720 includes a CONDITION option string consisting of a plurality (e.g., three) mutually exclusive choices.

[0114] The multiple options contained in the CONDITION option string are mutually exclusive. Therefore, a radio button is displayed at the beginning of each option in the CONDITION option string. This radio button allows the user to select only one of the multiple options.

[0115] Note that each of these options may contain one of the following strings: "|Source|", "{XXX}", or "Original Text". These strings are defined by variable definition 652 shown in Figure 20, and are replaced with predetermined values ​​when the prompt is created (or when the prompt options are displayed).

[0116] The selected option string is assigned to the variable represented by "CONDITION" in Figure 20. However, an option to not use CONDITION generation information may be provided among the choices. If such an option (NONE / not specified) is selected, the variable represented by "CONDITION" is set to empty.

[0117] (g) STYLE generation information 722 The STYLE generation information 722 includes a set of options for specifying the style of the final translation to be obtained. Each of these options includes wording determined by the researcher based on past experience in order to obtain translations of different styles. These options include those that include the variables shown in Figure 20, and those that can be freely entered by the user (indicated by "[Free Input]"). These options are mutually exclusive. Therefore, each option is displayed as a radio button. In this embodiment, the first option is selected as the default. This style can be not only a style in the ordinary sense, but also a style that specifies the format of the sentence, such as the markup of the translation.

[0118] In the prompt structure 650 shown in Figure 20, the variable indicated by "STYLE" is assigned a string corresponding to the option selected by the user. However, even with this option, an option to leave the user's value blank is available. If such an option (None / not specified) is selected, the variable indicated by STYLE is set to empty.

[0119] (h) ATTENTION generation information 724 The ATTENTION generation information 724 includes several ATTENTION option candidates. These options specify to the LLM server 64 what to pay attention to when translating. The options are not mutually exclusive. Possible ATTENTION option candidates include, for example, "Translate so that proper nouns are recognizable as proper nouns," and "Preserve the format of the 'source text' as much as possible even in 'machine translation'." For these options, a free input field may be provided, or the variables shown in the variable definition 652 of Figure 20 may be used. Note that option candidates that include a free input field may be made addable. For example, a "+" (plus) button may be provided at the end of an option that includes a free input field, and when this button is clicked, the option may be duplicated and its contents made editable.

[0120] A checkbox is displayed at the beginning of each ATTENTION option candidate. The checkboxes for certain option candidates may be checked by default, while the checkboxes for other option candidates may be unchecked by default. For example, the checkbox for the first candidate may be checked by default, while all others may be unchecked by default. It is desirable to allow the deletion of options added in this way, in case they become unnecessary.

[0121] The option selected by the user is added to the prompt by assigning the string of the selected option to the variable indicated by "ATTENTION" in the prompt structure 650 of Figure 20. When multiple options are selected, the strings of each option are concatenated and added to the prompt. If no options are selected, the variable indicated by "ATTENTION" is set to empty.

[0122] (i) GENRE generation information 726 The GENRE generation information 726 is used to specify the genre to which the text to be proofread belongs. The GENRE generation information 726 includes multiple genre option candidates.

[0123] The genre option candidates are mutually exclusive. In the option specification field 290 shown in Figure 8, a radio button is displayed at the beginning of each genre option candidate. By default, the first option, "None / Not specified," is selected.

[0124] For this option, the variable indicated by "GENRE" in the prompt structure 650 shown in Figure 20 will have the text 'The general is "{GGG}"' added to it. Furthermore, the 'GGG' part of this text will be replaced with a string containing the text of the specified option (for example, "Newspaper Article"). However, if the first option (None) is selected, this variable will be set to empty.

[0125] (j) SENSE generation information 728 Referring to Figure 23, the SENSE generation information 728 includes a SENSE option 780 that specifies the meaning of the words in the source text and a SENSE option 782 that specifies the meaning of the words in the translated text.

[0126] In SENSE Option 780 and SENSE Option 782 shown in Figure 23, [Original Word], [Original Meaning], [Translated Word], and [Translated Meaning] are all entered by the user.

[0127] In the option specification field 290 shown in Figure 8, a checkbox is displayed at the beginning of each option. When a checkbox is checked, the string of the corresponding option is assigned to the variable indicated by 'SENSE' in the prompt structure 650 shown in Figure 20. However, if no option is specified, this variable is set to empty. Also, when multiple options are selected, the strings corresponding to each option are concatenated and assigned to the variable indicated by 'SENSE'.

[0128] This option also allows for the duplication of each option candidate. Therefore, the add buttons 784 and 786 shown in Figure 23 are displayed after the text of each option candidate shown in the option specification field 290 in Figure 8. When an option candidate is added, an add button is also displayed after the text of that option candidate.

[0129] For example, when the SENSE option 780 is added to the prompt, the LLM server 64 will replace any string containing "[original word]" in the source text with "[original meaning]". When the SENSE option 782 is added, the LLM server 64 will replace any string containing "[translated word]" in the machine-translated text with "[translated meaning]".

[0130] (k) DICT generation information 730 The DICT generation information 730 includes a translation specification option for specifying the translation term. The translation specification option includes the strings "[source word]" and "[translated word]". Both of these can be freely entered by the user. In the option specification field 290 shown in Figure 8, a checkbox is displayed at the beginning of the translation specification option. When this checkbox is checked, the DICT option string is generated according to the content entered by the user and inserted into the prompt.

[0131] Specifically, when a translation option is specified, the string corresponding to that option is assigned to the variable indicated by "DICT" in the prompt structure 650 shown in Figure 20. In other words, that string is inserted into the prompt. However, if no option is selected at all, the variable indicated by "DICT" is set to empty.

[0132] Similar options can be added to the term selection options. To this end, an add button appears immediately after the text of the term selection option is displayed in the option selection field 290 shown in Figure 8. Added options can also be edited in the same way as existing term selection options.

[0133] When two or more options are specified, the strings obtained from those multiple options are concatenated and inserted into the prompt. This DICT option allows you to specify a glossary for the source language and target language. In the source text, source words found in the glossary are translated in the target text with the target words specified by the option.

[0134] (l) NOTE generation information 732 NOTE generation information 732 is an option for specifying the LLM server 64 to generate annotations regarding translation proofreading. For example, NOTE generation information 732 may include a NOTE option group consisting of four choices.

[0135] The NOTE option group includes four choices, for example: no specification regarding annotations, specify annotation output, specify detailed annotation output, and do not output annotations. These choices are mutually exclusive. Therefore, in the option specification field 290 shown in Figure 3, a radio button is displayed at the beginning of each of these choices. In this example, the first choice, "no specification," is selected by default.

[0136] For this option, the string assigned to the user's selected option is assigned to the variable indicated by "NOTE" in the prompt structure 650 of Figure 20. However, if the user selects "None", the variable indicated by "NOTE" is set to empty.

[0137] (m) FORMAT generation information 734 The FORMAT generation information 734 includes several (e.g., three) options that specify the output format to the LLM server 64.

[0138] This option allows various output formats to be specified for the LLM server 64. If the options are not mutually exclusive, a checkbox will appear before each option in the option specification field 290 shown in Figure 8. If these options are mutually exclusive, radio buttons will be used instead of checkboxes.

[0139] For this FORMAT generation information 734, the string corresponding to the selected option is assigned to the variable indicated by "FORMAT" in the prompt structure 650 of Figure 20. When multiple options are selected, the string formed by concatenating the texts of all those options is assigned to the variable indicated by "FORMAT". However, when none of the checkboxes are checked (no options are specified), the variable indicated by "FORMAT" is set to empty.

[0140] (n) RESPONSELANG generation information 736 The RESPONSELANG generation information 736 includes an option group consisting of multiple options that specify the language that the LLM server 64 will use for the response.

[0141] For example, if an option group contains three choices—"Do not specify a response language," "Specify {XXX} as the response language," and "Specify {YYY} as the response language"—these choices are mutually exclusive. Therefore, in the option specification field 290 shown in the figure, radio buttons are displayed for each of these three choices. As a result, only one of these choices is selected. Note that {XXX} and {YYY} are variables defined by the variable definition 652 shown in Figure 20, and are replaced with corresponding words when the option specification field 290 is displayed or when the prompt is generated.

[0142] For this option as well, the string of the option selected by the user is assigned to the variable indicated by "RESPONSELANG" in the prompt structure 650 of Figure 20.

[0143] 2. Operation The operation of the translation processing system 50 shown in Figure 1 will be described below. The operation of the translation processing system 50 has two phases: translation and proofreading, which is performed after translation. These will be explained in order below.

[0144] (A) Translation (a) Acquisition of translation control information When performing a new translation, the user saves the file of the source text to a storage device accessible by the client 62 in advance. To start the translation, the user accesses the translation service front end 70 of the translation system 60 from the client 62 shown in Figure 1. The translation service front end 70, in cooperation with the translation control server 72, sends the control information acquisition screen 160 shown in Figure 4 to the client 62. The control information acquisition screen 160 is displayed on the monitor of the client 62.

[0145] At this point, the source language selection button 184, destination language selection button 186, language change button 188, save button 194, and start processing button 196 in Figure 4 are all disabled and cannot be operated. At this time, the start processing button 196 displays "Start Translation".

[0146] The user opens a new source file to be translated by operating the selection button 180. The program associated with the selection button 180 reads the beginning of that file. The source file differs in format or extension from the file to be proofread. Therefore, the program determines that the user's operation is a translation and enables the source language selection button 184 and the destination language selection button 186. In some cases, the language of the source file can be determined from the contents of the opened file. In that case, the program associated with the selection button 180 selects that language as the source language selection button 184 and displays it as the default value.

[0147] Before performing a translation, the user must specify the source language and the target language. Therefore, the user specifies the source language and the target language using the source language selection button 184 and the target language selection button 186 shown in Figure 4. Once both are specified, the language change button 188, the save button 194, and the start processing button 196 become available.

[0148] To begin translation, the user clicks the Start Processing button 196. The program associated with the Start Processing button 196 sends information specifying the location of the specified file, along with information identifying the source and target languages ​​specified by the user, to the translation service frontend 70 shown in Figure 1.

[0149] Referring to Figure 2, the translation control information acquisition unit 100 of the translation service frontend 70 transmits this translation control information and information indicating the location of the file to be translated to the translation control server 72.

[0150] The control information storage unit 130 of the translation control server 72 shown in Figure 3 receives and stores this translation control information and information indicating the location of the file to be translated. In response to the information indicating the location of the file to be translated being stored in the control information storage unit 130, the source text receiving unit 132 receives the contents of the file to be translated from the location indicated by that information and stores it in the source text / translation storage device 76.

[0151] Referring to Figure 9, the person information extraction unit 320, the glossary extraction unit 322, and the summary extraction unit 324 included in the translation assistance information acquisition unit 134 shown in Figure 3, in response to the translation control information being stored in the control information storage unit 130, read the source text from the source / target text storage device 76, create the aforementioned prompts, and send them to the LLM server 64. The LLM server 64 returns the person information extracted from the source text to the person information extraction unit 320. The person information extraction unit 320 stores this person information in the person information storage unit 340 of the translation assistance information storage unit 136. The LLM server 64 returns the glossary (keyword list) extracted from the source text to the glossary extraction unit 322. The glossary extraction unit 322 stores this glossary in the glossary storage unit 342 of the translation assistance information storage unit 136. The LLM server 64 returns the summary information of the source text to the summary extraction unit 324. The summary extraction unit 324 stores this summary information in the summary storage unit 344 of the translation assistance information storage unit 136. From this point onward, the translation control server 72 starts the sentence translation process.

[0152] (b) Referring to the single-sentence translation process diagram 16, the translation control server 72 shown in Figure 1 first reads the file of the specified text to be translated from the source / target text storage device 76 (step 500), and displays the first sentence on the text to be translated display unit 252 shown in Figure 5. In the case of literary works, this sentence is usually the title.

[0153] In step 502, the translation control server 72 receives input of the translation corresponding to the first sentence to be translated from the user using the translation in progress display unit 256 shown in Figure 5. In the following step 504, the translation control server 72 combines the first sentence to be translated and its translation according to a predetermined format.

[0154] In the subsequent step 506, the translation control server 72 executes the sentence translation process 508 while incrementing the value of the variable N by 1 from 0 to the number of sentences to be translated (source text) minus 1.

[0155] The sentence translation process 508 for the Nth source text is executed as follows: First, a prompt for the (N+1)th source text to be translated by the LLM server 64 is generated using the string obtained in step 504, along with the person information, glossary, and summary information, as shown in the prompt example 4 above. Here, the number of the source text to be translated is "SN", and the number of the next source text to be translated is "S-N+1".

[0156] In step 522, the translation control server 72 sends the prompt generated in step 520 to the LLM server 64, and in step 524, waits for a reply from the LLM server 64. When the proofreading control server 74 receives a reply from the LLM server 64, it combines the N+1th source text and the translation received from the LLM server 64 according to the predetermined format described above. The proofreading control server 74 then uses the translation received from the LLM server 64 to update the source text display field 230 and the translation display field 232 shown in Figure 5, and terminates the sentence translation process 508.

[0157] The translation control server 72 repeatedly executes the processing of the translation processing system 50, moving the source text to be processed one position backward each time. Therefore, once the processing in step 506 is completed, translations for all source texts are obtained.

[0158] In step 510, the translation control server 72 displays a message on the client 62's monitor indicating that the translation is complete. In the following step 512, the translation control server 72 saves the translation result file to a location specified by the user and terminates the process. Here, the translation result file consists of, for example, strings in which each source text is combined with its corresponding translation in a predetermined format. In addition, whether or not the source text is a paragraph break is stored in the source text.

[0159] The translation process is now complete. The screen shown in Figure 4 will be displayed again on client 62. Typically, the user will then begin proofreading while reviewing the translated file.

[0160] (B) Proofreading Process When the proofreading process is started, the user first specifies a saved translation result file by operating the selection button 180 shown in Figure 4. When a translation result file is specified, the source language selection button 184, the destination language selection button 186, and the destination language selection button 186 on the selection button 180 remain disabled. On the other hand, the save button 194 and the start processing button 196 are enabled. In this case, the display on the start processing button 196 becomes "Start Proofreading". When the user clicks the start processing button 196, the proofreading process starts, and the proofreading control server 74 starts executing the program that implements the process shown in Figure 18.

[0161] Referring to Figure 18, in step 580, the proofreading control server 74 first reads the files necessary for prompt creation from the template storage unit 554, program storage unit 556, and library storage unit 558 shown in Figure 17, and then sends the proofreading window 280 shown in Figure 8, which was created using the translation service frontend 70 shown in Figure 1, to the client 62. The proofreading window 280 is displayed as a pop-up window on the monitor of the client 62.

[0162] In the following step 582, the user determines whether the prompt creation is complete, and the processing flow branches according to the result. If the prompt creation is not complete, the user edits the prompt by specifying each option of the POT in step 596 without pressing the confirmation button 296. If the user determines that the prompt creation is complete, the user clicks the confirmation button 296 shown in Figure 8 (step 584). As a result, the calibration control server 74 is notified that the prompt creation is complete. In response to the click of the confirmation button 296, the calibration control server 74 generates a prompt according to the POT edited by the user and sends it to the LLM server 64 (step 586).

[0163] Referring to Figure 1, the LLM server 64 responds to this prompt by taking the translated paragraph received from the proofreading control server 74 as input and generating a corrected translated paragraph that reflects the content specified by the prompt. The LLM server 64 returns this corrected translated paragraph to the proofreading control server 74.

[0164] In step 588, the proofreading control server 74 receives the proofread translated paragraph from the LLM server 64 and displays it in the LLM output field 294 shown in Figure 8. The user reviews this proofread translated paragraph and determines whether further proofreading is necessary. If further proofreading is necessary, the determination result in step 590 in Figure 18 becomes negative. The user further edits the prompt options for proofreading using the option specification field 290 shown in Figure 8 (step 596). The proofreading control server 74 regenerates the prompt according to these edited options and displays it in the prompt display / edit field 292. If the user decides to perform proofreading again based on the displayed prompt, the user clicks the confirmation button 296. As a result, the process proceeds from step 582 to step 594 shown in Figure 18, and the proofreading by the LLM server 64 described above is repeated.

[0165] This process is repeated until the proofreading result of the target translated paragraph is satisfactory. Once this is achieved, the user clicks the translated copy button 298 shown in Figure 6 (step 592 in Figure 18). As a result, the proofreading control server 74 copies the text of the finally obtained proofread translated paragraph to the clipboard and then closes the proofreading window 280 shown in Figure 8. The proofreading control server 74 then pastes the text from the clipboard into the translation in progress display unit 256 of the interactive screen 220. As a result, the proofread translated paragraph is displayed in the translation in progress display unit 256 instead of the translated paragraph before proofreading.

[0166] The user clicks the "Proceed" button 236 when they determine that the corrected translated paragraph is appropriate. The text displayed in the translation in progress display unit 256 moves to the translated text display unit 254. The original text displayed in the translation display unit 254 moves to the translated original text display unit 250. The next paragraph of the original text is displayed in the translated text display unit 254. The paragraph in the translation in progress display unit 256 is displayed in the translated text that corresponds to the original paragraph.

[0167] The above process will be repeated until all proofreading is complete.

[0168] 3. Effects of the First Embodiment As described above, in this embodiment, even with long source documents, prompts for the LLM server 64 to perform translation are generated using character information about the characters, glossary information about terms appearing in the document, summary information of the document, and contextual information consisting of the translation results of each sentence up to the sentence to be translated. Since the LLM server 64 performs translation using the information contained in the prompts, the translation can be performed in a way that reflects not only the information obtained from each sentence, but also the information of the entire document to be translated. As a result, even with long documents such as literary works where expression is important, the translation can be performed in a way that is consistent throughout. Furthermore, appropriate prompts are generated to have the LLM server 64 proofread the translated text using a pre-prepared POT, and the translated text can be proofread. As a result, the final translated text can be proofread to an appropriate expression in accordance with the translator's intentions. As a result, this first embodiment provides a translation system that can efficiently translate while maintaining consistency throughout the entire source text, including expression.

[0169] The prompts used for translation are prepared by researchers with linguistic expertise and a thorough understanding of the operation of the LLM server 64. Therefore, prompts that fully utilize the functions of the LLM server 64 can be created. Furthermore, the Prompts (POTs) used in proofreading can also be prepared by researchers. As a result, the number of proofreading steps required to obtain a satisfactory translation can be significantly reduced compared to conventional methods, while the final translation quality remains sufficiently high. This avoids situations where users unfamiliar with prompt creation spend long hours entering prompts without obtaining the desired results, or become lost and unsure how to create or modify prompts.

[0170] Second Modification In the first embodiment described above, the dialogue screen 220 and the calibration window 280 displayed on the client 62 are implemented using a browser. However, the present invention is not limited to such an embodiment. A program that displays the dialogue screen 220 and the calibration window 280 may be implemented as a dedicated application instead of a browser.

[0171] In the above embodiment, the entire source text to be translated is translated first, and then the translation is proofread paragraph by paragraph. However, this invention is not limited to such embodiments. The source text to be translated may be translated sentence by sentence, proofreading may be performed on each sentence, and then the translation of the next source text may be performed once the proofreading is complete. Alternatively, the source text to be translated may be translated paragraph by paragraph, proofreading may be performed on the translation of that paragraph, and then the translation of the next paragraph of the source text may be performed once the proofreading is complete. Furthermore, the translation unit and proofreading unit are not limited to one sentence or one paragraph. Any number of sentences or paragraphs may be translated and proofread as units, or specific units (chapter units, section units, etc.) may be translated and proofread.

[0172] In the above embodiment, the translation control server 72 and the proofreading control server 74 are both located on nodes separate from the client 62. However, the locations of the translation control server 72 and the proofreading control server 74 are not limited to these locations; they may be on the same node as the client 62, or they may be located inside the client 62 together with the translation service frontend 70, etc. Furthermore, the present invention may be implemented in a way that integrates the functions of the translation service frontend 70, the translation control server 72, and the proofreading control server 74, so that they operate as a single application from an external perspective.

[0173] In the above embodiment, the translation control server 72 and the proofreading control server 74 are located within the same node (translation system 60). However, the present invention is not limited to such an embodiment. The translation control server 72 and the proofreading control server 74 may be located on different nodes. Similarly, the translation service frontend 70 may be located on a node different from either the translation control server 72 or the proofreading control server 74, or both.

[0174] In the above embodiment, the same LLM server 64 is used for both translation and proofreading. However, this invention is not limited to such an embodiment. Different LLM servers may be used for translation and proofreading.

[0175] Furthermore, the above embodiments use translation between Japanese and English as an example. However, this invention is not limited to such embodiments. This invention can also be applied to translation between any languages.

[0176] Furthermore, functions similar to the proofreading prompt generation helper 550 according to the above invention can be used not only for translation but also when using a general LLM such as the LLM server 64. Moreover, this invention can be applied not only to LLMs but also to generative models that process non-linguistic information and output images, music, etc., in response to prompt input. In addition, this invention can be applied to so-called comic book translations in combination with image recognition technology, character recognition technology, and image generation technology.

[0177] The translation prompts and proofreading prompts (POTs) used in this invention can be prepared in advance according to the processing to be performed by the target LLM. Furthermore, as POTs, multiple options known to be effective through experience can be prepared in advance, and an interface can be used in which the user can select any one of the multiple options or modify pre-prepared general options appropriately and easily according to the task. As a result, high-quality output can be obtained from the generative model regardless of the type of generative model. Consequently, the user can obtain the desired output in a shorter time than before, while reducing the burden on the user.

[0178] Figure 3 shows the external appearance of a computer running a computer program that implements the translation system 60 according to the above embodiment, and Figure 25 shows an example of its hardware configuration. Figures 24 and 25 are merely examples, and it goes without saying that the translation system 60 can be implemented using a computer different from the one shown in Figures 24 and 25.

[0179] Referring to Figure 24, this computer system 1030 includes a computer 1050 and a keyboard 1054, a mouse 1056, and a monitor 1052, all connected to the computer 1050, for user interaction. Of course, these are just examples of configurations for when user interaction is required, and any general hardware and software available for user interaction (e.g., touch panels, voice input, pointing devices in general) can be used.

[0180] Referring to Figure 25, the computer 1050 includes a bus 1110, a CPU (Central Processing Unit) 1100, a GPU (Graphics Processing Unit) 1102, a ROM (Read-Only Memory) 1104 that stores the computer 1050's boot-up program, etc., all connected to the bus 1110, a RAM (Random Access Memory) 1106 that stores the instructions, system programs, and work data that make up the program when the program is executed, and an SSD (Solid State Drive) 1108, which is a non-volatile memory connected to the bus 1110. SSD 1108 is for storing programs executed by CPU 1100 and GPU 1102, as well as data used by programs executed by CPU 1100 and GPU 1102. These programs include programs for implementing the translation system 60 described above.

[0181] The computer 1050 further includes a network interface 1118 that provides connectivity to the network 1060, and a USB port 1116 to which a USB memory 1058 can be attached and detached, providing communication between the USB memory 1058 and various parts within the computer 1050. In this embodiment, the network 1060 is the internet.

[0182] The computer 1050 further includes an input / output interface 1114 connected to the bus 1110 for inputting input from the keyboard 1054 and mouse 1056 to the CPU 1100, and a display control unit 1112 connected to the bus 1110 for controlling the display on the monitor 1052 according to the control from the CPU 1100.

[0183] The computer program that causes this computer system 1030 to perform the functions described above is stored in the USB memory 1058. The USB memory 1058 is plugged into the USB port 1116, and the program is transferred to the SSD 1108. Alternatively, the program may be transmitted to the computer 1050 via the network 1060 and stored in the SSD 1108. The program is loaded into the RAM 1106 when it is executed. Part of the program may be stored in the ROM 1104.

[0184] The CPU 1100 reads the program from RAM 1106 according to the address indicated by its internal program counter, interprets the instructions, reads the data necessary for executing the instructions from RAM 1106, SSD 1108, or other devices according to the address specified by the instructions, and executes the processing specified by the instructions. The CPU 1100 stores the execution result data at an address specified by the program, such as RAM 1106, SSD 1108, or registers within the CPU 1100. At this time, the value of the program counter is also updated by the program. The computer program may be loaded directly into RAM 1106 from USB memory 1058 or via a network. Among the programs executed by the CPU 1100, tasks that can be executed in parallel may be found for iterative processing. For these tasks, it is thought that it will be possible to output the results according to the translation system 60 described above by executing them with the GPU 1102 according to the instructions included in the program or according to the analysis results when the CPU 1100 executes the instructions.

[0185] The program that implements the translation system 60 in cooperation with the computer 1050 includes a number of instructions that are written and arranged to cause the computer 1050 to operate in order to perform each of its functions. The method by which the computer 1050 operates to execute these instructions is well known. Therefore, the explanation of the method of operation will not be repeated here.

[0186] The embodiments disclosed herein are illustrative and not limited to those embodiments. The scope of the present invention is defined by the claims, with reference to the detailed description of the invention, and includes all modifications within the meaning and scope equivalent to the wording contained herein.

[0187] 50 Translation Processing System 60 Translation System 62 Client 64 LLM Server 70 Translation Service Frontend 72 Translation Control Server 74 Proofreading Control Server 76 Source / Target Text Storage Device 100 Translation Control Information Acquisition Unit 102 Dialogue Translation Unit 104 Dialogue Proofreading Unit 134 Translation Support Information Acquisition Unit 138 Translation Execution Unit 220 Dialogue Screen 230 Source Text Display Field 232 Target Text Display Field 250 Translated Source Text Display Unit 252 Translation Target Text Display Unit 254 Target Text Display Unit 256 Translated Text in Processing Display Unit 280 Proofreading Window 320 Person Information Extraction Unit 322 Glossary Extraction Unit 324 Summary Extraction Unit 340 Person Information Storage Unit 342 Glossary Storage Unit 344 Summary Storage Unit 508 Sentence Translation Processing 700 POT (Prompt Option Template) 1030 Computer System 1050 Computer< / context>

Claims

1. A translation system comprising: an auxiliary information extraction means for extracting auxiliary information for translation from a source text; and a translation execution means for obtaining a translated text of the source text into the target language as an output of the first large-scale language model by providing the source text, the auxiliary information, and a prompt instructing the first large-scale language model to translate the source text into the target language.

2. The translation system according to claim 1, wherein the auxiliary information extraction means includes: a person information acquisition means for obtaining the person information as an output of the second large-scale language model by providing the source text for translation and a prompt instructing the second large-scale language model to extract person information which is information relating to a person appearing in the source text; a glossary acquisition means for obtaining the glossary as an output of the third large-scale language model by providing the source text for translation and a prompt instructing the third large-scale language model to extract a glossary relating to terms appearing in the source text; or a summary acquisition means for obtaining the summary as an output of the fourth large-scale language model by providing the source text for translation and a prompt instructing the fourth large-scale language model to create a summary of the source text; or any combination of the person information acquisition means, the glossary acquisition means, and the summary acquisition means.

3. The translation system according to claim 2, wherein the prompt used by the person information acquisition means is a prompt that instructs the creation of person information including the name of a person appearing in the source text, the translation of the name of the person into the target language, gender, and personality.

4. The translation system according to claim 2, wherein the prompt used by the glossary acquisition means is a prompt that instructs the creation of a glossary including terms appearing in the source text, translations of those terms into the target language, and annotations relating to those terms.

5. The translation system according to claim 1, wherein the source text to be translated includes a plurality of source text units arranged in order, and the translation execution means includes a sentence translation execution means that sequentially provides each of the plurality of source text units to the first large-scale language model a prompt instructing to translate into the target language, and the translation of the immediately preceding source text unit into the target language.

6. The translation system according to any one of claims 1 to 5, further comprising: proofreading prompt generation means for generating prompts for proofreading the translation result of the translation execution means through dialogue with a user; and translation proofreading means for inputting the prompts generated by the proofreading prompt generation means and the translation result of the translation execution means into a large-scale language model for proofreading, thereby obtaining a proofread translation result as the output of the large-scale language model for proofreading.

Images