Translation support system and translation support method

Abstract

The present invention supports work for translating an original document in a first language into a translated document in a second language.　The present invention comprises: a parallel translation acquisition unit that acquires parallel translation information including an original sentence in a first language and a translated sentence obtained by translating the original sentence in a second language; and a correction processing unit that gives an input including the parallel translation information and a correction instruction for the translated sentence to a generation model and acquires a response from the generation model. The correction processing unit repeats the input of the correction instruction and the acquisition of the response until a repetition end condition is satisfied.

Description

Translation support system and translation support method

[0001] The present invention relates to a translation support system and a translation support method that support the work of translating an original document in a first language into a translated document in a second language.

[0002] Conventionally, in document translation work, a professional translator reads the original text and manually creates a translated text while appropriately interpreting the context and the meaning of terms. Such work relies heavily on the translator's language skills and specialized knowledge, and requires significant time and effort. In addition, it is difficult to maintain uniform translation quality, and efficiency improvements have been required for large-scale translation operations.

[0003] In recent years, with the advancement of artificial intelligence technology, machine translation models have been put into practical use, and machine translation systems that automatically convert original text into translated text have come into widespread use. This makes it possible to process the initial stage of translation quickly and efficiently, reducing the burden of translation work. In particular, the method of "post-editing", in which translated text generated by machine translation is manually corrected, is becoming common, and there is a demand for both translation quality and efficiency to be achieved.

[0004] The translation apparatus described in Patent Document 1 discloses a configuration for converting a first text written in a first language into a translated text in a second language using a neural machine translation model. In this apparatus, when a plurality of different translated words are output by machine translation for the same original word, a unified third text is generated by replacing the other translated words with one representative translated word. Furthermore, the apparatus is configured to display the representative translated word so that it can be distinguished from other translated words, and measures are provided to support post-translation editors in easily recognizing and correcting inconsistencies in translated words.

[0005] Patent Document 2 discloses a translation support device capable of generating natural-sounding translations by translating specified words with specified translations. The translation support device described in Patent Document 2 acquires parallel text information consisting of source text and translated text obtained by machine translation, and further associates and references specified words and corresponding specified translations registered in a terminology database. The translation support device generates instructions to make the sentence such that the specified words contained in the source text are translated with the specified translations, and performs sentence generation processing using a generation model to obtain a revised translation. This realizes a means of achieving both consistency in terminology and naturalness in the translated text.

[0006] Japanese Patent Publication No. 2024-128094 Japanese Patent Publication No. 2025-005833

[0007] Prior art documents disclose techniques for correcting inconsistencies in terminology in translated texts using machine translation and generative models. However, post-editing requires corrections not only for inconsistencies in terminology but also for avoiding ambiguity, grammatical issues, mistranslations, omissions, and additions. In other words, the challenge lies in the inability to streamline these overall correction processes.

[0008] The problem to be solved by this invention is to provide a technology that assists in the process of translating a source document in a first language into a translated document in a second language.

[0009] [1] A translation support system that assists in the process of translating a source document in a first language into a translated document in a second language, comprising: a parallel text acquisition unit that acquires parallel text information including the source document in the first language and a translated text obtained by translating the source document in the second language; and a correction processing unit that provides input including the parallel text information and correction instructions for the translated text to a generation model and acquires a response from the generation model, wherein the correction processing unit repeatedly inputs the correction instructions and acquires the response until a repeat termination condition is met. [2] The translation support system according to [1], wherein the repeat termination condition includes at least one of the following: the response from the generation model includes a statement that no correction is necessary, and the number of responses to be acquired reaches a predetermined number set in advance. [3] The translation support system according to [1] or [2], wherein the correction processing unit performs input to the generation model and acquires responses for multiple sentences or multiple segments included in the parallel text information, and excludes translated texts that were determined to not require correction based on the previous response from the target of correction instructions input in subsequent times. [4] The translation support system according to any one of [1] to [3], wherein the correction processing unit generates multiple subsets, each containing multiple sentences or segments included in the parallel text information, provides each subset to a generation model, obtains a response from the generation model for each subset, and if it is determined that all sentences or segments included in the subset do not require correction based on the previous responses, excludes the subset from the input target for correction instructions from the next time onward. [5] The translation support system according to any one of [1] to [4], wherein the correction processing unit provides the generation model with a processing request to generate a response containing the source text, the translated text, and a corrected translated text obtained by correcting the translated text, obtains the response from the generation model, and sets the corrected translated text included in the response as the subject of the next processing request. [6] The translation support system according to any one of [1] to [5], wherein the correction instruction includes an instruction to reflect in the correction of the translated text without substantially changing at least one of the following selected from the quantity, elements, order, and relative positional relationship of the tag information included in the source text. [7] The correction processing unit generates an explanation in a specified language for the corrected parts of the translation, as described in any of [1] to [6].

[0010] The invention described in [1] allows the generative model to repeatedly make corrections to the translated text, thereby improving the accuracy of the translation and streamlining post-editing. The invention described in [2] allows the generative model to perform an appropriate number of repetitions, thereby optimizing the accuracy of the translation and the processing time. The invention described in [3] reduces the amount of text processed by the generative model, thereby shortening the translation time. The invention described in [4] reduces the number of requests and responses to the generative model and enables parallel processing of subsets by processing the target text or segment in subset units, thereby shortening the processing time. The invention described in [5] allows the generative model's responses to be appropriately structured, enabling effective iterative corrections. The invention described in [6] allows the generative model to infer whether the tag information of the source text is appropriately reflected in the translated text and to make corrections as necessary. The invention described in [7] allows for obtaining explanations regarding the corrected parts.

[0011] According to the present invention, it is possible to provide a technology that supports the process of translating an original document in a first language into a translated document in a second language.

[0012] System configuration diagram according to this embodiment. Hardware configuration diagram according to this embodiment. Processing sequence diagram according to this embodiment. Data configuration diagram according to this embodiment. Processing flowchart according to this embodiment. Different processing flowchart according to this embodiment. Graph showing the relationship between the number of corrections and accuracy according to this embodiment.

[0013] The following describes a translation support system and a translation support method according to embodiments of the present invention, with reference to the drawings. Note that the embodiments shown below are examples of the present invention, and the present invention is not limited to these embodiments; various configurations can be adopted.

[0014] In this embodiment, the configuration and operation of the translation support system and translation support device will be described, but translation support methods, computer programs, and program recording media on which such programs are recorded will have similar effects. Using a program recording medium, for example, the program can be installed on a computer. The series of processes according to this embodiment, which will be described below, are provided as a program executable on a computer and can be provided via non-transient computer-readable recording media such as CD-ROMs and flexible disks, as well as via communication lines.

[0015] The translation support system consists of a computer device. The computer device includes a processor such as a CPU (Central Processing Unit) and memory. The computer device can function as a translation support device by executing a translation support program stored in memory using its processing unit. The translation support method is implemented through processing by the computer device, including the translation support device.

[0016] In this embodiment, translation refers to converting a sentence in a first language into a sentence in a second language. The first language and the second language include various languages ​​used as natural languages. For example, the first language is Japanese and the second language is English, but are not limited to these.

[0017] In this embodiment, translation assistance refers to assistance with the post-editing process within the translation work. Post-editing means making necessary corrections to the translated text obtained by machine translation of the source text. These corrections include clarifying ambiguous expressions, correcting grammatical errors, correcting mistranslations and omissions, deleting unnecessary words, adding words and phrases, and standardizing terminology.

[0018] In this embodiment, the source document refers to a document file containing sentences in the first language. The document file only needs to contain at least text, and the file format is not limited. The source text refers to the sentences that make up the source document. The translated document refers to a document file containing sentences that are translations of the sentences in the first language contained in the source document into the second language. The translated text refers to the sentences that make up the translated document. In this specification, the terms translated document and translated text are defined to include the translated document and translated text before and after post-editing.

[0019] In this context, the "original translation" refers to, for example, a text translated from the original text using machine translation, and the "revised translation" refers to a text that has undergone one or more post-editing revisions to the original translation. When distinguishing between revised and unrevised translations, the term "n-time revised translation" is used. n is a natural number of 1 or more that indicates the number of revisions. In this embodiment, the n-time revised translation refers to a text that has been revised n-1 times. When n=1, the n-1-time revised translation refers to the original translation. The original translation document, the revised translation document, and the n-time revised translation document are defined in a similar manner.

[0020] In this embodiment, the parallel text information includes the source text and its corresponding translation. The parallel text information manages the source text and the target text as a pair in a single file. In this embodiment, the parallel text information holds tag information consisting of font information and document structure information for the source text and the target text. Specifically, the parallel text information can use an XLIFF (XML Localization Interchange File Format) file, but the format is not limited to this.

[0021] Figure 1 shows a system configuration diagram of the translation support system 1 according to this embodiment. As shown in Figure 1, the translation support system 1 comprises a translation support device 2, a user terminal 3, and a generation device 4. The translation support device 2, the user terminal 3, and the generation device 4 are each connected via a communication network NW.

[0022] Translation support device 2 is a computer device installed by the operator of translation support system 1. Translation support device 2 includes, as functional components, a document acquisition unit 21 that accepts input of a source document or a document pair including a source document and a translated document, a parallel text acquisition unit 22 that acquires parallel text information including the source and translated texts, a correction processing unit 23 that performs corrections on the translated text, and an output unit 24 that outputs the correction results including the corrected translated text.

[0023] User terminal 3 is a terminal device operated by the user. The user can perform operations such as specifying source documents, document pairs, and translation information via user terminal 3, and obtain the results of those modifications. User terminal 3 includes, as functional components, an input unit that accepts operation instructions and various data inputs, and an output unit that outputs screen displays and audio. Although only one user terminal 3 is shown in Figure 1, multiple user terminals 3 may exist.

[0024] The generation device 4 is a computer device installed by the operator of the translation support system 1 or an external operator. The generation device 4 receives a processing request input from the translation support device 2 and generates a response thereto. In this embodiment, the processing request includes parallel text information and correction instructions for the translated text. The generation device 4 sends the generated response back to the translation support device 2. In Figure 1, only one generation device 4 is shown, but multiple devices may be installed.

[0025] The generation device 4 can generate a response by inputting the data included in the processing request into the generation model. A known model that has been trained using machine learning can be used as the generation model. In this embodiment, a Large Language Model (LLM) is used as the generation model. An LLM is a type of natural language processing model that has been trained using large amounts of natural language data. For example, an LLM receives text containing questions or instructions, or various files, as input and outputs text that is a response to them. Note that the LLM may have a multimodal configuration and can handle data input and output in various formats such as images, videos, and audio. Furthermore, in this embodiment, the generation model is not limited to an LLM, and other machine learning models or algorithms capable of generating responses to inputs may be used.

[0026] In this embodiment, an example is described in which the generated model is stored in the generation device 4, but a configuration in which the translation support device 2 is equipped with the generated model can also be adopted. Furthermore, the generation device 4 may be installed inside the translation support device 2, or installed outside the translation support device 2 and connected to the device by wire or wireless.

[0027] In this embodiment, the translation support system 1 is described as comprising a translation support device 2 and a user terminal 3, but is not limited to this configuration. For example, the translation support device 2 may be configured as a terminal device operated by the user, and may be configured to accept input of source documents, document pairs, and parallel translation information. In this case, the translation support system 1 may not include the user terminal 3.

[0028] Figure 2(a) shows the hardware configuration diagram of the translation support device 2. The translation support device 2 comprises a processor 201, a memory 202, and a communication interface 203 as its hardware configuration. In this embodiment, the translation support device 2 can use a computer device such as a server or a personal computer. The translation support device 2 may be composed of multiple computer devices, and is not limited to the configuration shown in Figure 2(a) as long as the above-described functional components (21-23) can be realized as a whole. The generation device 4 has the same hardware configuration as the translation support device 2.

[0029] The processor 201 consists of one or more processors, such as a CPU, and controls the overall processing in the translation support device 2 by executing translation support programs, an OS (Operating System), and other applications. The memory 202 consists of one or more transient and / or non-transient storage devices, such as an HDD (Hard Disk Drive), SSD (Solid State Drive), flash memory, and RAM (Random Access Memory), and stores translation support programs and various data. The communication interface 203 controls communication with the communication network NW and realizes data communication with the user terminal 3, the generation device 4, and the storage unit 5. The processor 201 makes the computer function as the translation support device 2 by executing the translation support program and executes the translation support method.

[0030] Figure 2(b) shows the hardware configuration diagram of the user terminal 3. The user terminal 3 comprises a processor 301, memory 302, communication device 303, input interface 304, and output device 305. In this embodiment, the user terminal 3 can be a smartphone, personal computer, tablet terminal, or the like.

[0031] The processor 301 consists of one or more processors such as a CPU and controls the overall processing on the user terminal 3 by executing the OS and other applications. The memory 302 is an HDD, SSD, flash memory, RAM, etc., and stores browser applications and various data. The communication device 303 controls communication with the communication network NW and realizes data communication with the translation support device 2, etc. The input interface 304 receives operation instructions from the operator and consists of at least one of the following devices: a touch panel, mouse, keyboard, etc. The output device 305 consists of a display that shows the processing results of the processor 301 and a speaker that outputs sound.

[0032] Figure 3 shows the processing sequence in the translation support system 1.

[0033] In step S101, the user terminal 3 transmits the source document or a document pair including the source document and the translated document to the translation support device 2. The translation support device 2 acquires parallel information based on the acquired documents. In step S101, the user terminal 3 may also be configured to transmit parallel information including the source and translated texts to the translation support device 2.

[0034] In step S102, the translation support device 2 sends a processing request to the generation device 4, which includes parallel text information and instructions for modifying the translated text. In step S103, the generation device 4 sends a response generated in response to the acquired processing request to the translation support device 2. This response may include at least a revised translated text in which the original text has been modified.

[0035] In this embodiment, the translation support device 2 repeatedly executes the processes of steps S102 and S103 until the repeat termination conditions, which will be described in detail later, are met. If the response of step S103 includes a revised translation, the revised translation is set as the translation for the next processing request.

[0036] When the repetition termination condition is met, in step S104, the translation support device 2 generates a correction result based on the response from the generation device 4 and transmits it to the user terminal 3. The correction result is based on the corrected translation or corrected translated document.

[0037] Figure 4 shows examples of the structure of various types of information handled by the translation support system 1. Figure 4(a) shows an example of the structure of parallel text information. Figures 4(b) and (c) show examples of the structure of the response (response information) from the generative model specified by the processing request.

[0038] According to Figure 4(a), the parallel text information comprises a segment, source text, and translation text. The parallel text information can be extracted from the source document and the translated document obtained by machine translation or the like. A segment is information used to identify the source text and translation text as a pair of predetermined units. A segment is set up with one or more sentences as units, and each unit is assigned a sequential number or the like. For example, segment "02" has the Japanese source text "It's nice weather." and the English translation "It's nice day." A segment is set up to divide a text into single sentences based on, for example, a sentence-ending marker, but is not limited to this, and may be divided to include multiple sentences based on a predetermined sentence structure or the like. In addition, some documents may contain sentences written without sentence-ending markers or sentences with ambiguous sentence divisions. Therefore, even when a segment is set up to divide into single sentences, it is permitted to include multiple sentences or fragments less than one sentence.

[0039] In this embodiment, the source text and the translated text include tag information consisting of font information and document structure information, in addition to text information. For example, segment "01" has the Japanese source text "<g id= / "1 / ">Hello.< / g>". The text information in this source text is "Hello.", and the tag information is "<g id= / "1 / ">", "< / g>", etc. Tag information adds font information and the like to the text information. Font information includes, but is not limited to, typeface, font size, bold, italics, underline, strikethrough, color, etc. Similarly, the translated text may also include font information and tag information.

[0040] According to Figure 4(b), the response information includes a segment, source text, translated text, revised translated text, and explanation. The response information is information generated by the generative model when it receives input (processing request) that includes parallel text information and instructions for revising the translated text of the parallel text. In the response information, in addition to the segment, source text, and translated text of the parallel text, a revised translated text and explanation are newly generated.

[0041] The revised translation is the revised version of the original translation. If there are no revisions from the original translation, the revised translation may be the same as the original translation or left blank. The explanation is a descriptive text about the revised translation in the specified language (Japanese in the example). The explanation is generated to show what revisions were made between the original translation and the revised translation. The explanation may be generated as a descriptive text for each aspect of the revision, such as mistranslation, omission, addition, ambiguity, or grammatical issues. If there are no revisions from the original translation, the explanation will state that no revisions were necessary, such as "No revisions."

[0042] In the example in Figure 4(b), segment "01" has been modified from the translated text "Hello. <g id= / "1 / "><g>" to the corrected translated text "<g id= / "1 / ">Hello.< / g>". This modification corrects the position of the tag information in the translated text based on the position of the tag information in the original text "<g id= / "1 / ">Hello.< / g>". The response information generates an explanation for this modification: "Tag position corrected. Translation unchanged". It should be noted that modifications may occur not only to tag information but also to text information.

[0043] In the present embodiment, a processing request for a generation model and a response thereto are repeatedly executed. Accordingly, response information is generated a plurality of times. The post-correction translated text of the response information is set as the "translated text" in the next processing request. FIG. 4(c) shows an example of response information from the second correction with respect to the response information from the first correction in FIG. 4(b). In FIG. 4(c), the translated text of the segment "01" is set with a post-first-correction translated text. Further, the post-second-correction translated text has not been corrected from the post-first-correction translated text, and an explanation "no correction" is generated. In the present embodiment, a translated text determined to require no correction can be excluded from input targets of correction instructions from the next time onward. As a result, processing in the generation model can be reduced, and processing time can be shortened.

[0044] FIG. 5 shows a processing flowchart in a translation support apparatus 2 according to one embodiment. In FIG. 5, step S201 denotes a document acquisition step, step S202 denotes a parallel translation acquisition step, steps S203 to S207 denote a correction processing step, and step S208 denotes an output step, respectively.

[0045] In step S201, the document acquisition unit 21 accepts input of an original document in a first language, or a document pair including the original document in the first language and a translated document obtained by translating the original document into a second language. In the present embodiment, the document acquisition unit 21 acquires the original document, and generates an XLIFF file including the original text (source). The document acquisition unit 21 can generate a translated text (target) by applying machine translation to the XLIFF file. Note that in step S201, the document acquisition unit 21 may be configured in any manner as long as it can acquire the original text and its translated text, and a configuration for directly acquiring a document pair or the like can be appropriately adopted.

[0046] In step S202, the parallel translation acquisition unit 22 acquires parallel translation information including an original text in the first language and a translated text obtained by translating the original text into the second language, and stores the parallel translation information in the memory 202. The parallel translation acquisition unit 22 can acquire the parallel translation information by extracting the original text and the translated text based on the information acquired in step S201. Note that the parallel translation acquisition unit 22 may acquire the parallel translation information by accepting input of the parallel translation information via the user terminal 3. In this case, the document acquisition step (S201) can be omitted.

[0047] In step S203, the correction processing unit 23 transmits a processing request including bilingual information and a correction instruction for the translated text to the generation device 4. The correction instruction is configured as a prompt that refers to the bilingual information and generates response information including at least the corrected translated text. In the present embodiment, the correction instruction employs an instruction to generate a corrected translated text for the translated text, focusing on mistranslation, omission, addition, ambiguity, grammatical problems, and the like. Further, it is possible to further employ an instruction to generate the corrected translated text without substantially changing at least any one or more selected from the quantity, elements, order, and relative positional relationship of tag information between the original text and the corrected translated text. The phrase "without substantially changing" means that at least any one or more of the quantity, elements, order, and relative positional relationship of tag information is retained between the original text and the corrected translated text. However, even if the order of tag information or the like is changed due to grammatical differences between languages, as long as such a change does not impair the semantic and structural consistency of the translated content, this is permitted as "not substantially changing" in the present embodiment.

[0048] In the present embodiment, the processing request may include prompts related to an output format, constraint conditions, and a correction instruction. The output format includes, for example, an instruction to output each piece of data of the response information illustrated in FIG. 4 in JSON format, an instruction to specify the first language and the second language, and the like. The constraint conditions are for setting predetermined rules to improve the accuracy of the response.

[0049] In step S203, the bilingual information included in the processing request may be in units of one or a plurality of segments. When the number of segments included in the processing request exceeds a preset predetermined number, the processing request may be transmitted to the generation device 4 in a plurality of separate times.

[0050] In step S204, the correction processing unit 23 acquires a response to the processing request from the generation device 4.

[0051] In step S205, the correction processing unit 23 refers to at least one of the corrected translation and the explanation included in the response and determines whether the corresponding translation does not require correction. In this embodiment, for example, the correction processing unit 23 can determine that the translation does not require correction if the translation included in the response information (corrected translation) matches the previous uncorrected translation. Alternatively, the correction processing unit 23 may determine that the translation does not require correction if, for example, the explanation corresponding to the translation included in the response information (corrected translation) is "no correction".

[0052] If the translation is determined to not require correction (YES in step S205), the correction processing unit 23 excludes the translation from the input of correction instructions for subsequent steps (step S206). The excluded translation is stored in memory 202 in a way that distinguishes it from the translations that were not excluded. If the translation is not determined to not require correction (NO in step S205), the correction processing unit 23 sets the corresponding corrected translation as the target translation for the next parallel information and stores it in memory 202.

[0053] In this embodiment, in steps S205 and S206, the correction processing unit 23 may define exception conditions for determining whether correction is unnecessary. Specifically, the exception conditions include not excluding a translation that is determined to not require correction in the first response from the input target for the next correction instruction. These exception conditions are set for the purpose of causing the generation model to infer at least once whether there are any errors in the content of the corrections included in the first response.

[0054] In step S207, the correction processing unit 23 determines whether the repetition termination condition is met. If it is determined that the repetition termination condition is not met (NO in step S207), the correction processing unit 23 returns to step S203 and repeatedly executes the input of correction instructions and the acquisition of responses in steps S203 to S206. These processes are executed repeatedly until the repetition termination condition is met.

[0055] In this embodiment, the repetition termination condition includes at least one of the following: the response from the generation model indicates that no correction is needed, and the number of responses obtained reaches a predetermined number of times. Here, "no correction is needed" means that no correction is needed for all target translations, for example, a state in which all translations included in the translated document no longer require correction. The predetermined number of responses obtained can be arbitrarily set by the administrator, for example, to 30 or less. In this embodiment, the repetition termination condition may be set as reaching a predetermined number of responses obtained, and may also be set as a condition for terminating the repetition before reaching that number of responses obtained, by having the response from the generation model indicate that no correction is needed.

[0056] If it is determined that the repeated termination condition is met (NO in step S207), the output unit 24 outputs the correction result based on the response information (step S208). The correction result is generated based on the corrected translation and the explanation. In one embodiment, the correction result is output in a format that allows comparison of the difference between the original translation and the final corrected translation. Here, an explanation corresponding to the final corrected translation may also be included. In another embodiment, the correction result is output in a list format of the original translation and a series of corrected translations. Again, an explanation corresponding to the series of corrected translations may also be included. Furthermore, the correction result may include a translated document file that reflects all of the final corrected translations. By sending the correction result to the user terminal 3, the output unit 24 allows the user to obtain a translated document with improved translation accuracy and streamline post-editing.

[0057] Figure 6 shows a processing flowchart in a translation support device 2 according to a different embodiment. In this embodiment, the translation support device 2 generates a subset containing multiple sentences included in the parallel text information, and performs correction instructions and obtains responses for the generation model on a subset-by-subset basis. Processes common to each step in Figure 5 will be omitted from explanation as appropriate.

[0058] In step S303, the correction processing unit 23 generates multiple subsets, each containing multiple segments from the parallel translation information. In this embodiment, a subset is defined as a unit that groups together multiple segments, and is set up for purposes such as appropriately adjusting the number of processing requests to the generation model and parallelizing the processing requests. Note that a subset may be defined as a unit that groups together multiple sentences instead of segments.

[0059] As a concrete example, suppose the parallel text information has 100 segments (or 100 sentences). If the generation model processes and responds to 100 segments in a single processing request, problems arise such as the processing time being long and the number of characters being limited. In this embodiment, by dividing the 100 segments into 10 subsets, each of these subsets can be processed in 10 processing requests. For example, the first subset includes the parallel text information of segments "01" to "10", and the second subset includes the parallel text information of segments "11" to "20". Note that the number of segments to be set in a subset is not limited to 10 and can be set as appropriate considering processing time and processing efficiency.

[0060] In step S304, the modification processing unit 23 sends a processing request to the generation device 4 for each subset. Here, multiple processing requests relating to multiple subsets may be sent to multiple generation devices 4 or multiple generation models and processed in parallel.

[0061] In step S305, the correction processing unit 23 obtains the response for each subset. In step S306, the correction processing unit 23 determines which subsets do not require correction. A subset that does not require correction includes cases where all translations within the subset are determined to not require correction.

[0062] If it is determined that a subset does not require modification (YES in step S306), in step S307, the modification processing unit 23 excludes the subset from subsequent processing requests. If it is determined that a subset does not require modification (NO in step S306), the modification processing unit 23 sets the modified translations included in the subset as the target translations for the parallel information in the next processing request and stores them in memory 202. At this time, translations determined not to require modification within the subset may be excluded from subsequent processing requests. Alternatively, the translations that are excluded from a subset may be replaced with translations from other subsets. In other words, subsets may be optimized to have a set number of segments.

[0063] In step S308, the correction processing unit 23 determines whether the repetition termination condition is met, and repeatedly performs the input of correction instructions and acquisition of responses until the repetition termination condition is met. In step S309, the output unit 24 generates the correction result based on the acquired response and transmits it to the user terminal 3.

[0064] Next, we will explain the results of our analysis of the relationship between the number of iterations of correction by the generative model and the accuracy of the translation, referring to Tables 1 and 2 and Figure 7.

[0065] Table 1 shows the accuracy of the machine translation (MT) pre-correction translation and the generative model n-correction translation for translation documents A through F. Here, the accuracy of the translation document is represented as 100% when no corrections are needed by human post-editing and the document is complete. In Table 1, "na" represents the state where the generative model's response no longer requires correction.

[0066] Table 1 shows the accuracy of the translation after up to 10 revisions by the generative model on the original translation document before correction by MT. For example, Rev:1 represents the translation document after one revision, Rev:2 represents the translation document after two revisions, and Final represents the translation document after 10 revisions. The number of targets indicates the number of segments to be translated. In this embodiment, translation documents A to F each have different file formats and document contents. Specifically, translation document B is a presentation file, and translation document F is a document file containing tabular data. The other translation documents are text-only document files. The document contents include topics related to management, economics, news, etc.

[0067]

[0068] Table 1 shows that the first revision corrects translated documents A to E by more than 70% of their final accuracy. Furthermore, the accuracy of translated documents A to E generally improves by the sixth revision, after which no further revisions are needed or the degree of improvement in accuracy becomes smaller. Based on these results, in this embodiment, the number of responses obtained for a predetermined number of times can be 10 or less, and more preferably 6 or less. Translated document F is a document containing accounting terms, and the accuracy in Table 1 is calculated based on the list of translations published by the Financial Services Agency. Although the translations in translated document F did not match those in the said list, the translations were at a level that is commonly used. If these translations were considered correct, the accuracy of translated document F would improve by 2-3% on average.

[0069] Figure 7 is a graph showing the relationship between the number of revisions and the accuracy of the translation for each translated document in Table 1. In Figure 7, the vertical axis represents accuracy, and the horizontal axis represents the number of revisions (number of repetitions). According to Figure 7, it can be seen that the accuracy of each document improves as the number of revision repetitions increases. Thus, according to the invention of this embodiment, it is possible to provide a technology that improves the efficiency of post-editing by repeatedly performing revisions using a generative model.

[0070] Table 2 shows the accuracy of the uncorrected machine translation (MT) version and the n-corrected version using the generative model for translated documents A through F. While Table 1 shows results for segments only, Table 2 shows results for the sentences included in each segment. The number in Table 2 indicates the number of sentences to be translated. Note that Table 2 only evaluates the accuracy of the final corrected translation.

[0071] Table 2 shows that the accuracy of translated documents A through F has improved compared to the uncorrected and corrected translated documents in Table 1. This result can be attributed to the fact that when a segment requiring correction contains multiple sentences, not all of those sentences necessarily need correction. Therefore, the results in Table 2 can be said to more accurately represent the reduction in actual post-editing work.

[0072]

[0073] 1 Translation support system 2 Translation support device 21 Document acquisition unit 22 Bilingual acquisition unit 23 Correction processing unit 24 Output unit 3 User terminal 4 Generation device

Claims

1. A translation support system that assists in the process of translating a source document in a first language into a translated document in a second language, comprising: a parallel text acquisition unit that acquires parallel text information including the source document in the first language and a translated text obtained by translating the source document into the second language; and a correction processing unit that provides a processing request including the parallel text information and correction instructions for the translated text to a generation model and acquires a response from the generation model, wherein the correction processing unit sets the corrected translated text included in at least the first response as the translated text for the next processing request, and in the second and subsequent responses, repeatedly inputs the correction instructions and acquires the responses until a repeat termination condition is met, wherein the repeat termination condition includes a condition indicating that the response from the generation model does not require correction.

2. The translation support system according to claim 1, wherein the repetition termination condition includes reaching a predetermined number of response acquisitions.

3. The translation support system according to claim 1 or 2, wherein the correction processing unit inputs and obtains responses to the generation model for multiple sentences or multiple segments included in the parallel information, and excludes translated texts that were determined not to require correction based on the previous response from the input of correction instructions for subsequent times.

4. The translation support system according to claim 1 or 2, wherein the correction processing unit generates multiple subsets, each containing multiple sentences or segments included in the parallel translation information, provides each subset to a generation model, obtains a response from the generation model for each subset, and, if it is determined that all sentences or segments included in the subset do not require correction based on the previous responses, excludes the subset from the input target for correction instructions in subsequent operations.

5. The translation support system according to claim 1 or 2, wherein the correction processing unit provides the generation model with a processing request to generate a response that includes the source text, the translated text, and a corrected translated text obtained by correcting the translated text, retrieves the response from the generation model, and sets the corrected translated text included in the response as the translation for the next processing request.

6. The translation support system according to claim 1 or 2, wherein the correction instructions include instructions to reflect in the correction of the translation at least one of the following selected from the quantity, elements, order, and relative positional relationships of the tag information contained in the source text without substantially changing it.

7. The translation support system according to claim 1 or 2, wherein the correction processing unit generates an explanation in a specified language for the corrected parts of the translation.

8. A translation support method that assists in the process of translating a source document in a first language into a translated document in a second language, comprising: a parallel acquisition step of acquiring parallel information including the source document in the first language and a translated text obtained by translating the source document in the second language; a correction processing step of providing a processing request to a generation model including the parallel information and correction instructions for the translated text, and obtaining a response from the generation model, wherein the correction processing step sets the corrected translated text included in at least the first response as the translated text for the next processing request, and in the second and subsequent responses, repeats the input of the correction instructions and the acquisition of the response until a repeat termination condition is met, wherein the repeat termination condition includes a condition indicating that the response from the generation model does not require correction.

Images