Translation support apparatus and program

Abstract

To provide a translation support device and a program for improving the efficiency of translation using machine translation software.SOLUTION: A translation support device divides a source text to be translated that is written in a first language into predetermined segments, generates a translated text in a second language that is different from the first language for each of the segments, associates the source text for each of the segments with the corresponding translated text and outputs the associated source text and translated text as parallel translation information, outputs, to a large scale language model, the parallel translation information and revision instruction information that instructs revision of the parallel translation information, acquires an output from the large scale language model, and generates revised parallel translation information.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a translation support device and a program. [Background technology]

[0002] In recent years, machine translation software technology using neural networks and the like has become widely used to assist translation work. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Chinese Patent Application Publication No. 108984540 Summary of the Invention [Problem to be solved by the invention]

[0004] While machine translation software that uses neural networks and other technologies can improve translation accuracy to a certain extent through training, various problems remain unresolved, such as inconsistent translations of certain terms, an inability to handle unknown words, and the inability to output a translation that corresponds to the original text, or so-called missing translations, or the appearance of translations that do not correspond to the original text, or so-called "ghosting."

[0005] For this reason, even when machine translation software was used, a post-editing process was required in which a translator checked and corrected the translation results, which ultimately led to the problem that this did not lead to improved efficiency.

[0006] The present invention has been made in view of the above circumstances, and one of its objects is to provide a translation support device and program that improve the efficiency of translation using machine translation software. [Means for solving the problem]

[0007] One aspect of the present invention that solves the problems of the above-mentioned conventional examples is a translation support device that includes: machine translation means that divides an original text to be translated, which is written in a first language, into predetermined segments, generates a translation in a second language different from the first language for each segment, and associates the original text for each segment with the corresponding translation, and outputs them as bilingual information; corrected output acquisition means that outputs to a large-scale language model the bilingual information output by the machine translation means, along with correction instruction information that instructs correction of the bilingual information, and acquires output from the large-scale language model; and corrected bilingual generation means that extracts a corrected translation from the output acquired by the corrected output acquisition means, divides the translation into parts corresponding to segments included in the bilingual information, and associates the divided parts of the corrected translation that correspond to the segments with the original text of the corresponding segment, thereby generating corrected bilingual information. [Effects of the Invention]

[0008] According to the present invention, the translation results produced by machine translation software are corrected by a machine learning system that utilizes a large-scale language model, thereby reducing the amount of manual post-editing required and improving the efficiency of translation using machine translation software. [Brief explanation of the drawings]

[0009] [Figure 1] 1 is a block diagram illustrating an example of the configuration of a translation support device according to an embodiment of the present invention. [Figure 2] 1 is a functional block diagram illustrating an example of a translation support device according to an embodiment of the present invention. [Figure 3] FIG. 2 is a flowchart illustrating an example of the operation of the translation support device according to the embodiment of the present invention. [Figure 4] FIG. 2 is an explanatory diagram illustrating an example of processing performed by the translation support device according to the embodiment of the present invention. [Figure 5] FIG. 10 is an explanatory diagram illustrating another example of processing performed by the translation support device according to the embodiment of the present invention. [Figure 6]1 is a configuration block diagram illustrating an example of a translation support device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0010] An embodiment of the present invention will be described with reference to the drawings. As shown in Fig. 1, a translation support device 1 according to the embodiment of the present invention is realized by a general computer system, and includes, for example, a control unit 11, a storage unit 12, an operation unit 13, a display unit 14, and a communication unit 15.

[0011] The control unit 11 is a program-controlled device such as a CPU, and operates according to a program stored in the storage unit 12. In one example of the present embodiment, the control unit 11 divides an original text to be translated, written in a first language, into predetermined segments, generates a translation for each segment in a second language different from the first language, associates the original text for each segment with the corresponding translation, and outputs it as bilingual information (for example, so-called XLIFF (XML Localization Interchange File Format) data).

[0012] The control unit 11 also outputs the bilingual information and correction instruction information that instructs correction of the bilingual information to the large-scale language model, and acquires output from the large-scale language model. Here, the large-scale language model may be realized by a server device R communicably connected to the translation support device 1, and is a generative AI system that receives instructions written in a natural language as input and generates sentences based on the instructions. This machine learning system may be, for example, a well-known system such as GPT-4o, Llama, or Gemini (each a registered trademark or trademark). For the purpose of this embodiment, it is also preferable to select a machine learning system with an MMLU (Massive Multitask Language Understanding) score of 75 or higher.

[0013] The control unit 11 extracts the corrected translation from the output of the large-scale language model, divides the translation into parts corresponding to the segments included in the bilingual information, associates the corrected translation parts corresponding to the segments obtained by the division with the original text of the corresponding segments, and generates and outputs the corrected bilingual information. The detailed operation of the control unit 11 will be described later.

[0014] In this embodiment, the control unit 11 executes the programs stored in the storage unit 12, thereby functioning as a machine translation unit, an edited output acquisition unit, and an edited translation generation unit.

[0015] The storage unit 12 is a memory device, a disk device, or the like, and stores a program executed by the control unit 11. This program may be provided by being stored in a computer-readable, non-transitory recording medium and copied to the storage unit 12. The storage unit 12 also operates as a work memory for the control unit 11.

[0016] The operation unit 13 includes a mouse, a keyboard, etc. The operation unit 13 accepts instructions from the user of the translation support device 1 and outputs information indicating the content of the accepted instructions to the control unit 11. The display unit 14 is a display device or the like, and displays information according to instructions input from the control unit 11.

[0017] The communication unit 15 is a network interface or the like, and sends and receives information to and from an external server R (for example, a server that provides the above-mentioned GPT-4o service) in accordance with instructions input from the control unit 11.

[0018] Next, an example of the operation of the control unit 11 will be described. In this example of the present embodiment, the control unit 11 executes a program stored in the storage unit 12, thereby realizing a configuration that functionally includes an original text acceptance unit 20, a machine translation unit 30, a correction acquisition unit 40, a corrected parallel translation generation unit 50, and an output unit 60, as illustrated in Fig. 2. The machine translation unit 30 here is configured to include a segmentation processing unit 31, a preprocessing unit 32, a machine translation processing unit 33, a parallel translation information output unit 34, and a postprocessing unit 35.

[0019] The original text receiving unit 20 receives an original text to be translated, which is written in a first language, in response to a user instruction.

[0020] The segmentation processing unit 31 of the machine translation unit 30 divides into predetermined segments the original text to be translated that has been accepted by the original text acceptance unit 20. This process of dividing into segments is, for example, a process of dividing into sentences or paragraphs, and is a process that is widely known in the field of machine translation, so a detailed explanation of it will be omitted here.

[0021] The preprocessing unit 32 executes predetermined preprocessing upon receiving input of the original text to be translated accepted by the original text accepting unit 20 or the original text divided into segments by the segmentation processing unit 31. In one example of the present embodiment, the preprocessing unit 32 detects inline tags by pattern matching or the like, replaces each different inline tag with a unique replacement string corresponding to the inline tag, and stores the replacement string in association with the string of the corresponding inline tag as an inline tag database.

[0022] The preprocessing unit 32 also sends the source text to be translated to the server R on which the large-scale language model is running, extracts terms from the source text, and outputs term extraction instruction information requesting that the extracted terms be associated with their expressions (translations) in a second language and output them. In response to this instruction, the preprocessing unit 32 acquires a term database containing the terms output by the large-scale language model and their translations, and stores the acquired term database as a term database. This term database can be used later for processing by the correction acquisition unit 40 or to support manual corrections.

[0023] Specifically, the term extraction instruction information here is: "Extract key terms from the following sentence and match them with their translations in the second language." The preprocessing unit 32 sends the original text to be translated together with this term extraction instruction information to the server R, and obtains from the server R a list of terms extracted from the original text to be translated and their corresponding translations.

[0024] One of the features of this embodiment is that the sentences provided to the large-scale language model are the entire original sentences to be translated, and not the results of segmentation.

[0025] The machine translation processor 33 sequentially performs machine translation processing on the original text G1, G2... divided into segments into the second language to sequentially obtain translated text T1, T2... for each segment corresponding to the original text of each segment. In one example of this embodiment, the machine translation processor 33 performs so-called neural network translation (NMT), and when a translated text Ti corresponding to an original text Gi is obtained, the original text Gi and the corresponding translated text Ti are output as a pair to the bilingual information output unit 34.

[0026] The bilingual information output unit 34 generates bilingual information that associates the original text Gi with the corresponding translation Ti for each segment. As already explained, this bilingual information may be in the form of data in the so-called XLIFF (XML Localization Interchange File Format). Data in this format is widely known, so a detailed explanation will be omitted here.

[0027] The post-processing unit 35 refers to the inline tag database generated by the pre-processing unit 32, and replaces the replacement strings of inline tags contained in the original text Gi and the translated text Ti (i = 1, 2 ...) in the bilingual information output by the bilingual information output unit 34 with the corresponding inline tags, thereby updating the bilingual information (XLIFF data).

[0028] The correction acquisition unit 40 outputs the bilingual information output by the machine translation unit 30 (bilingual information updated by the post-processing unit 35) along with correction instruction information that instructs correction of the bilingual information to the large-scale language model, and acquires the output from the large-scale language model.

[0029] Specifically, the correction acquisition unit 40 according to this embodiment outputs instruction information including an instruction to correct missing translations or translations that do not correspond to the original text. The correction instruction information in this example is, for example, as follows: "If there are any missing parts in the translation of the following sentence in the first language in the second language, please complete them. If there are any missing parts, please delete the missing parts and adjust the context. However, do not change the wording of the translation in the second language. Please output only the corrected translation in the second language."

[0030] The correction acquisition unit 40 sequentially sends the bilingual translations (pairs of original text Gi and translated text Ti) for each segment contained in the bilingual information together with such correction instruction information to the server R on which the large-scale language model is running, and obtains corrected output (output including the corrected translation) based on the above correction instruction information by the large-scale language model.

[0031] However, what is shown here is merely an example, and the correction acquisition unit 40 may send to the server R correction instruction information that is not limited to instructions to correct missing translations or translations that have no corresponding parts in the original text, as explained here, but also includes instructions to replace unknown words, correct terminology, correct inline tags, etc. Examples of these will be explained later.

[0032] The corrected parallel translation generation unit 50 extracts corrected translations from the output acquired by the correction acquisition unit 40. Specifically, when the output of the large-scale language model is data in JSON (JavaScript Object Notation) format, the corrected parallel translation generation unit 50 extracts corrected translations generated by the machine learning system from this data.

[0033] The revised translation generation unit 50 divides the extracted revised translation sentence into parts corresponding to the segments included in the translation information, and associates the divided parts of the revised translation sentence corresponding to the segments with the original sentence of the corresponding segment to generate revised translation information.

[0034] In this example of the present embodiment, the correction acquisition unit 40 sends out a translation for each segment and obtains a corrected translation for each segment, so the corrected translation generation unit 50 replaces the corrected translation with the translation of the corresponding segment of the original text, updates the translation information, and creates the corrected translation information.

[0035] When the output unit 60 obtains modified parallel translation information in which the translations for all segments of the parallel translation information output by the machine translation unit 30 have been replaced with modified translations through processing by the modification acquisition unit 40 and the modified parallel translation generation unit 50, it outputs this modified parallel translation information as the result of the translation.

[0036] [Operation] The translation support device 1 of this embodiment basically has the above configuration and operates as follows: In the following explanation, an example is taken in which the first language is English and the second language is Japanese.

[0037] 3, the translation support device 1 receives an original text to be translated written in English (S11) and divides the original text to be translated into segments for each paragraph (S12). This English sentence to be translated may contain an inline tag indicating that, for example, an f and an i are to be displayed as a ligature in a portion where these are arranged side by side.

[0038] As preprocessing, translation support device 1 detects inline tags by pattern matching or the like, and replaces each different inline tag with a unique replacement string corresponding to the inline tag. Furthermore, translation support device 1 associates the replacement string with the string of the corresponding inline tag and stores it as an inline tag database (S13: preprocessing).

[0039] The translation support device 1 sequentially processes the original texts G1, G2, etc. for each segment obtained in step S12, and performs machine translation processing into the second language, Japanese, using neural network translation (NMT), thereby sequentially obtaining corresponding translated texts T1, T2, etc. (S14).

[0040] The translation support device 1 generates bilingual information that pairs an original text with a corresponding translation, and also refers to the inline tag database generated in step S13 to replace the replacement strings of inline tags contained in the original text and the translation in the bilingual information with the corresponding inline tags (S15: Generation of bilingual information).

[0041] The translation support device 1 receives the following correction instruction information together with the pairs of original sentences Gi and translation sentences Ti (i=1, 2, . . . ) included in the bilingual information: "Regarding the Japanese translation of the following English sentence, please complete any missing translations. If there are any missing translations, please delete the missing expressions and adjust the context. However, please do not change the Japanese translation. Please output only the corrected Japanese translation." This sentence is sent to server R on which a large-scale language model is running (S16: Instruction to generation AI).

[0042] The translation support device 1 receives a corrected translation T'i of the translation Ti from the server R (S17), and generates corrected translation information by replacing the translation Ti associated with the corresponding original text Gi included in the bilingual information with this corrected translation T'i (S18).

[0043] The translation support device 1 performs the processes of steps S14 to S18 for each segment, generates corrected bilingual information, and then outputs this corrected bilingual information as the translation result (S19).

[0044] For example, the translation support device 1 imports the XLIFF file, which is the corrected bilingual information, into a translation support tool (such as a so-called translation memory) and completes the process.

[0045] According to this embodiment, post-editing of translations using machine translation software is performed using a large-scale language model (for example, generative AI), thereby improving the efficiency of translations using machine translation software.

[0046] [Other processing examples of the correction acquisition part] As explained above, the processing of the correction acquisition unit 40 of the control unit 11 is not limited to instructions to correct missing translations or translations that have no corresponding parts in the original text, but may also output instructions to the large-scale language model, such as instructions to replace unknown words, correct terminology, or correct inline tags. These instructions may be executed separately or in combination.

[0047] [Term correction] In the example of term correction, the correction acquisition unit 40 also instructs processing of the bilingual translations for each segment contained in the bilingual information (pairs of original text Gi and translated text Ti), and extracts words contained in the terminology database generated by the pre-processing unit 32 from the words contained in this original text Gi. The correction acquisition unit 40 then generates correction instruction information including an instruction to correct the extracted words by replacing them with translations associated with them in the terminology database. The correction instruction information in this example is, for example, as follows: "In the following sentence in the first language, please translate the word XX into the second language with the translated word YY. Please output only the corrected translation in the second language."

[0048] Although the example described here is one in which words contained in both the original text and the terminology database are extracted and indicated, the correction acquisition unit 40 may also indicate the terminology database and generate correction instruction information that includes an instruction to correct the translated text while referring to the terminology database.

[0049] The correction acquisition unit 40 then sends the generated correction instruction information, along with the pair of the original text Gi and the translation text Ti, to the server R on which the large-scale language model is running, and obtains a corrected output (output including the corrected translation) based on the correction instruction information by the large-scale language model.

[0050] In this embodiment, a term database is first generated from the entire original text (the entire text before segmentation) in the preprocessing unit 32. This prevents variations in translations of terms between segments, and ensures consistency of terminology.

[0051] [Unknown word] Furthermore, the correction acquisition unit 40 may detect parts that have not been translated as unknown words as a result of the processing by the machine translation unit 30, and instruct the large-scale language model to perform processing corresponding to the unknown words.

[0052] The correction acquisition unit 40 in this example identifies, as an unknown word, a portion of a character string included in a translation Ti included in a bilingual translation for each segment (a pair of an original text Gi and a translation text Ti) that matches a predetermined pattern. For example, if the original text is in English, and there is a portion of a character string included in the translation text Ti that contains a predetermined number of consecutive alphanumeric characters, the character string in that portion is identified as an unknown word.

[0053] The correction acquisition unit 40 then generates correction instruction information including an instruction to replace the identified unknown words with translations. In this example, the correction instruction information is, for example, as follows: "In the second language translation of the following sentence in the first language, please replace the word XX with the translated sentence. Please output only the corrected second language translation."

[0054] The correction acquisition unit 40 sends the generated correction instruction information together with the pair of original text Gi and translation text Ti to the server R on which the large-scale language model is running, and obtains a corrected output (output including the corrected translation text) based on the correction instruction information by the large-scale language model.

[0055] In general, large-scale language models have machine-learned a huge number of documents, including specialized documents, and are known to have knowledge of translations for words that cannot be translated by the neural network used in NMT. ​​This example of the present embodiment focuses on this point, and makes it possible to replace words that are identified as unknown words by the NMT of the machine translation unit 30 with translations.

[0056] [Inline tag correction] Furthermore, when an inline tag portion is detected from the original text in the preprocessing unit 32, the correction acquisition unit 40 generates correction instruction information including an instruction to make the inline tag portion correspond to the original text in the corresponding translated text.

[0057] The correction acquisition unit 40 detects inline tags (e.g., <tag0001>For example, if the source text Gi has an inline tag <tag0001>, the correction instructions for this example would be: "For the following sentence in the first language translated in the second language, <tag0001>Please leave it as is, do not replace or delete it. Please output only the corrected second language translation."

[0058] The correction acquisition unit 40 sends the generated correction instruction information together with the pair of original text Gi and translation text Ti to the server R on which the large-scale language model is running, and obtains a corrected output (output including the corrected translation text) based on the correction instruction information by the large-scale language model.

[0059] [combination] As explained above, the correction instruction information may include a plurality of types of correction instructions. For example, when an instruction is given to deal with inline tags in addition to dealing with missing translations and overflows, the correction acquisition unit 40 "If there is a missing translation in the second language for the following sentence in the first language, please complete it. If there is a missing translation, please delete the missing expression and adjust the context. However, please do not change the Japanese translation. Also, please use the tag <tag0001>Please leave it as is, do not replace or delete it. Please output only the corrected second language translation." Such correction instruction information is generated and sent to a server R running a large-scale language model together with a pair of the original text Gi and the translation text Ti.

[0060] [Large-scale language models running locally] In the explanation so far, the large-scale language model has been assumed to operate on an external server R separate from the translation support device 1, but the present embodiment is not limited to this. For example, the control unit 11 of the translation support device 1 may operate as a large-scale language model. In this example, when performing the process of step S16 in FIG. 3 (the process of the correction acquisition unit 40), the control unit 11 of the translation support device 1 uses the generated correction instruction information and a pair of an original text Gi to be corrected by the instruction and a translation text Ti as input to the large-scale language model it is operating, and also uses the output of the large-scale language model as input to the corrected parallel translation generation unit 50 (to be used as the corrected translation text T'i in the process of step S17).

[0061] [Speed ​​up] The machine translation process (NMT process) in the machine translation unit 30 is generally faster than the process in the correction acquisition unit 40 (process of making corrections in a large-scale language model).

[0062] Therefore, in one example of the present embodiment, the translation support device 1 performs processing in the machine translation unit 30 and processing in the correction acquisition unit 40 in parallel. For example, assume that the processing time (average time) of the machine translation unit 30 for one segment is T, and the processing time (average time) of the correction acquisition unit 40 for the same segment is α·T (α is generally a number greater than or equal to 1). In this case, if the original text is divided into 10 segments, when the processing in the machine translation unit 30 and the processing in the correction acquisition unit 40 are performed consecutively, a time of (1+α)·T is required per segment, and if α=3, for example, a total time of 40T is required.

[0063] 4, if the machine translation unit 30 finishes processing one segment and then processes the next segment while the correction acquisition unit 40 processes the segment that the machine translation unit 30 has already processed, processing of n segments of original text will be completed in time T+nαT. Here, if α=3 and n=10, the total processing time will be 31T, which enables faster processing.

[0064] Such processing can be realized by the control unit 11 executing the processing of the machine translation unit 30 and the processing of the correction acquisition unit 40 (and the processing of the corresponding corrected translation generation unit 50) in parallel as separate threads. There are various widely known methods for such parallel processing, including this thread processing, and any of them may be adopted.

[0065] Furthermore, in one example of this embodiment, the translation support device 1 may perform multiple processes of the correction acquisition unit 40 (and the corresponding processes of the corrected parallel translation generation unit 50) in parallel. In this process, too, the control unit 11 may execute the processes of the correction acquisition unit 40 (and the corresponding processes of the corrected parallel translation generation unit 50) in multiple threads.

[0066] The translation support device 1 in this example prepares multiple (for example, k, where k>1 and preferably k≧α) threads that execute the processing of the correction acquisition unit 40 and the corrected parallel translation generation unit 50, and each time the machine translation unit 30 finishes processing one segment and outputs parallel translation information, one of the threads that has not started processing at that time processes the parallel translation information output by the machine translation unit 30. At this time, the machine translation unit 30 may be processing another segment, and other threads that execute processing of the correction acquisition unit 40, etc. may also be processing other segments in parallel.

[0067] In this example, as illustrated in Figure 5, when the machine translation unit 30 completes processing for the first segment, a first thread related to the correction acquisition unit 40, etc., starts processing, and even if this processing is not completed, when the machine translation unit 30 completes processing for the second segment, a second thread related to the correction acquisition unit 40, etc., starts processing, and so on.

[0068] In this example, when α=3, k=3, and n=10, the first thread processes the 1st, 4th, 7th, and 10th segments, the second thread processes the 2nd, 5th, and 8th segments, and the third thread processes the 3rd, 6th, and 9th segments. The processing of the machine translation unit 30 ends when n·T=10T, and immediately thereafter the first thread, which is free, processes the 10th segment. As a result, the entire processing ends when 10T+3T=13T, enabling faster processing. [Example]

[0069] As an example, we will explain an example in which the translation support device 1 of this embodiment is used to convert the detailed description of an actual patent document (European patent publication EP3042922B1 is used as an example) up to paragraph

[0022] into a Word document and process it.

[0070] The text was first segmented by the translation support device 1, resulting in a total of 44 segments. Next, each segment was translated into Japanese using the general-purpose NT, a neural network machine translation engine provided by the National Institute of Information and Communications Technology (NICT). For comparison, the bilingual results were manually post-edited in advance to extract errors.

[0071] Next, the translation support device 1 automatically generated query sentences (prompt sentences) that referenced the original text information for each segment and the results of the general-purpose NT translation, and then generated corrected translations using a large-scale language model that runs locally.These results were compared with the error extraction results obtained by manual post-editing, and the following summarizes how corrections were made by the translation support device 1.

[0072] [Fixed status] Number of segments post-edited: 10 segments Percentage of modified segments to total segments: 23% Total number of post-edited parts: 13 The breakdown is shown in Table 1, and the details are shown in Tables 2 to 4.

[0073] [Table 1]

[0074] [Table 2] [Table 3] [Table 4]

[0075] [Comparative Example] The results of translating the English sentence in segment No. 44 in Tables 2 to 4 using the local large-scale language model used in the embodiment are shown below. Original text: <tag0001>

[0022] <tag0002> <tag0003> <tag0004>The phrase "substituted alkyl", "substituted alkenyl", and "substituted alkynyl" refers to alkyl, alkenyl and alkynyl groups as defined above in which one or more bonds to a carbon(s) or hydrogen(s) are replaced by a bond to non-hydrogen or non-carbon atoms such as, but not limited to, a halogen atom in halides such as F, Cl, Br, and I; and oxygen atom in groups such as carbonyl, carboxyl, hydroxyl groups, alkoxy groups, aryloxy groups, and ester groups; a sulfur atom in groups such as thiol groups, alkyl and aryl sulfide groups, sulfone groups, sulfonyl groups, and sulfoxide groups; a nitrogen atom in groups such as amines, amides, alkylamines, dialkylamines, arylamines, alkylarylamines, diar- ylamines, N-oxides, imides, enamines imines, oximes, hydrazones, and nitriles; a silicon atom in groups such as in trialkylsilyl groups, dialkylarylsilyl groups, alkyldiarylsilyl groups, and triarylsilyl groups; and other heteroatoms in various other groups. Translation result: <tag0001>

[0022] <tag0002> <tag0003> <tag0004>The phrases "substituted alkyl," "substituted alkenyl," and "substituted alkynyl" refer to alkyl, alkenyl, and alkynyl groups, as defined above, in which one or more carbon or hydrogen bonds have been replaced with a bond to a non-hydrogen atom or a non-carbon atom. This includes halide groups containing halogen atoms (e.g., F, Cl, Br, I); oxygen atoms in carbonyl, carboxyl, hydroxyl, alkoxy, aryloxy, and ester groups; sulfur atoms in thiol, alkyl and aryl sulfide groups, sulfone groups, sulfonyl groups, and sulfoxide groups; nitrogen atoms in amines, anamides, alkylamines, diallylamines, allylamines, alkylarylamines, diallylamines, N-oxides, imides, enamine imines, oximes, hydrazones, and nitriles; silicon atoms in trialkylsilyl groups, dialkylarylsilyl groups, alkyldiallylsilyl groups, and triallylsilyl groups; and various groups containing other heteroatoms.

[0076] As shown in this comparison example, when translation processing was performed solely using generative AI with a large-scale language model, the translated sentence was split into two sentences and the syntax was inappropriate. Furthermore, all instances of "group = group" in words such as "alkynyl group" were mistranslated as "group" in the translated sentence, and "group" at the end of the sentence was also mistranslated.

[0077] [Effects of the Example] On the other hand, with the translation support device 1 (FIG. 6) of this embodiment, as described above, 13 translation defects in the NICT general-purpose NT, a typical neural network machine translation engine, are resolved by automatic post-editing using the large-scale language model (LLM) of this embodiment, and results similar to those obtained when post-editing is performed manually can be obtained.

[0078] In other words, the effect of this embodiment is not due to the translation capabilities of the large-scale language model, but rather is the result of correct corrections being made using the inference capabilities of the large-scale language model based on the translation produced by NMT and the prompt instructions.

[0079] International standardization for translation is defined by ISO 17100. This standard specifies requirements for ensuring the quality and reliability of the translation process at translation service providers. ISO 17100 focuses on translator qualifications, project management, the translation process, and other aspects of quality assurance. The ISO 17100 workflow includes specific steps to ensure quality throughout the translation process. The main steps are listed below:

[0080] 1. Project Preparation Customer requirements review: A detailed review of customer requirements, objectives, and expected outcomes. Establish project specifications: Establish specific specifications for the project, such as language pairs, deadlines, formatting, style guides, glossaries, etc. 2. Selection of translators Translator qualification verification: Check the translator's education, experience, and subject matter knowledge to select the right translator. 3. Translation Process Translation: The translator translates the source text into the target language, adhering to terminology and style guides to ensure accurate translation. Self-review: The translator reviews the translation himself / herself to check for errors or inappropriate expressions. 4. Proofreading and Review Proofreading: A proofreader separate from the translator checks the translation and points out errors and areas for improvement. Review: If necessary, reviews are conducted by experts and native speakers. 5. Final confirmation and delivery Final Check: The project manager performs a final quality check to ensure all requirements are met. Delivery: Deliver the completed translation to the client, and if necessary, receive feedback from the client and incorporate any improvements.

[0081] Of these steps, step 3, the translation process, is the most time-consuming, and the advent of NMT was expected to significantly improve productivity in this process. However, due to the challenges shown in [Table 1], translators are currently forced to perform post-editing to correct NMT errors.

[0082] The automatic post-editing using the large-scale language model in this embodiment effectively eliminates the need for translators in step 3, and the workflow can begin with the proofreading process by a checker in step 4, making it possible to significantly reduce translation costs and delivery times. [Explanation of symbols]

[0083] 1 Translation support device, 11 Control unit, 12 Memory unit, 13 Operation unit, 14 Display unit, 15 Communication unit, 20 Original text acceptance unit, 30 Machine translation unit, 31 Segmentation processing unit, 32 Preprocessing unit, 33 Machine translation processing unit, 34 Bilingual information output unit, 35 Postprocessing unit, 40 Correction acquisition unit, 50 Corrected bilingual generation unit, 60 Output unit. < / tag0003> < / tag0002> < / tag0001> < / tag0003> < / tag0002> < / tag0001>

Claims

1. a machine translation means for dividing an original text to be translated, which is written in a first language, into predetermined segments, generating a translation for each segment in a second language different from the first language, and outputting the original text for each segment and the corresponding translation as bilingual information; a corrected output acquisition means for outputting, to a large-scale language model, bilingual information output by the machine translation means, together with correction instruction information for instructing correction of the bilingual information, and acquiring an output from the large-scale language model; a corrected translation generating means for extracting a corrected translation from the output acquired by the corrected output acquiring means, dividing the translation into parts corresponding to segments included in the bilingual information, and associating the divided parts of the corrected translation corresponding to the segments with the original text of the corresponding segments to generate corrected bilingual information; A translation support device comprising:

2. 2. The translation support device according to claim 1, The acquisition means acquires, as the instruction information, A translation support device that outputs instruction information including an instruction to correct missing translations or translations that have no corresponding parts in the original text.

3. 2. The translation support device according to claim 1, The translation system further includes a term database acquisition means for outputting term extraction instruction information to a large-scale language model, which extracts terms from an original text to be translated, and requests the large-scale language model to output the extracted terms in association with expressions of the terms in the second language, and for acquiring a term database including terms and their corresponding translations; The correction output acquisition means outputs the bilingual information output by the machine translation means and the terminology database acquired by the terminology database acquisition means to a large-scale language model, and also outputs correction instruction information that instructs correction of the bilingual information using the terminology database.

4. 2. The translation support device according to claim 1, Further, the system has an unknown word extraction means for extracting words that satisfy a predetermined condition as unknown words from the translations included in the bilingual information, The correction output acquisition means of the translation support device includes, in the correction instruction information, an instruction to correct the bilingual information by replacing the extracted word with a translation when the unknown word is extracted.

5. 2. The translation support device according to claim 1, Further, the apparatus has an inline tag extraction means for extracting an inline tag portion from the bilingual information, The correction output acquisition means of the translation support device includes, in the correction instruction information, an instruction to correct the bilingual information by making the inline tag portion correspond to the original text when the extracted inline tag portion is present.

6. 2. The translation support device according to claim 1, the machine translation means sequentially processes the segments obtained by dividing the bilingual original text information to generate translations in a second language, and outputs bilingual information that associates the original text for each segment with the corresponding translation; A translation support device in which, when the machine translation means outputs a segment, the corrected output acquisition means and the corrected parallel translation generation means process the parallel translation information related to the output segment in parallel with the processing by the machine translation means.

7. 7. A translation support device according to claim 6, A translation support device comprising a plurality of said corrected output acquisition means and corrected parallel translation generation means, wherein when said machine translation means outputs a segment, if there is a corrected output acquisition means or a corrected parallel translation generation means that has not started processing, said corrected output acquisition means or corrected parallel translation generation means that has not started processing processes the parallel translation information relating to said output segment in parallel with the processing by said machine translation means.

8. Computer, a machine translation means for dividing an original text to be translated, which is written in a first language, into predetermined segments, generating a translation for each segment in a second language different from the first language, and outputting the original text for each segment and the corresponding translation as bilingual information; a corrected output acquisition means for outputting, to a large-scale language model, bilingual information output by the machine translation means, together with correction instruction information for instructing correction of the bilingual information, and acquiring an output from the large-scale language model; a corrected translation generating means for extracting a corrected translation from the output acquired by the corrected output acquiring means, dividing the translation into parts corresponding to segments included in the bilingual information, and associating the divided parts of the corrected translation corresponding to the segments with the original text of the corresponding segments to generate corrected bilingual information; A program that functions as a

Images