Translation method and device

Abstract

The embodiment of the invention provides a translation method and device, computer equipment, a computer readable storage medium and a computer program product, and belongs to the field of data processing. The translation method comprises the steps that a to-be-translated text is received, and a translation first draft is generated based on the to-be-translated text through a large language model; the translation first draft is used as a translation draft of the current round when iteration starts to execute translation iteration, a translation result is obtained, and the translation iteration process comprises the steps that an evaluation problem set is constructed based on the translation draft of the current round, and structured feedback information is generated based on the evaluation problem set and the translation draft of the current round; revising the translation manuscript of the current round based on the structured feedback information, and continuing iteration by taking the revised translation manuscript as the translation manuscript of the next round under the condition that the revised translation manuscript does not meet the preset quality requirement and does not reach the preset number of iterations; and outputting a translated text based on the translation result. According to the technical scheme, automation of translation process optimization can be achieved, and the accuracy and consistency of translation are improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the technical field of data processing, and in particular, to a translation method and device, computer equipment, computer readable storage medium, and computer program product. BACKGROUND

[0002] Currently, when machine translation is performed, a large language model is usually used to directly generate a translation text according to a single prompt engineering. However, since there is a lack of quality verification mechanism, the translation text obtained in this way usually has low accuracy and consistency, and the quality of the translation text is low.

[0003] It should be noted that the above content is not necessarily prior art, and is not used to limit the patent protection scope of the present application. SUMMARY

[0004] Embodiments of the present application provide a translation method and device, computer equipment, computer readable storage medium, and computer program product to solve or alleviate one or more technical problems raised above.

[0005] One aspect of embodiments of the present application provides a translation method, which includes: receiving a text to be translated, and generating a translation draft based on the text to be translated using a large language model; performing a translation iteration with the translation draft as the current round of iteration at the beginning of the iteration, and obtaining a translation result in a case where the translation draft after the iteration meets a preset quality requirement or reaches a preset iteration number, the translation iteration process including constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and continuing the iteration with the revised translation draft as the translation draft of the next round in a case where the revised translation draft does not meet the preset quality requirement and the preset iteration number has not been reached; outputting a translation text based on the translation result.

[0006] Optionally, the evaluation question set covers at least two dimensions of term consistency, style matching degree, logical coherence, sentence specification, and semantic fidelity.

[0007] Optionally, the structured feedback information includes at least an error position interval, an error type, and a revision suggestion field.

[0008] Optionally, the revising the translation draft of the current round based on the structured feedback information includes: determining a to-be-revised segment in the translation draft based on the error position interval; rewriting the to-be-revised segment locally based on the error type and the revision suggestion field.

[0009] Optionally, the preset quality requirement includes one or more of term consistency, style adherence, semantic fidelity, and sentence pattern matching up to a preset threshold.

[0010] Optionally, the performing of the translation iteration includes: generating, by a feedback agent, structured feedback information based on the evaluation question set and the translation draft of the current round; revising, by a revision agent, the translation draft of the current round based on the structured feedback information.

[0011] Optionally, the feedback agent is a combination of multiple models that generate the structured feedback information through a voting mechanism.

[0012] Optionally, the revision agent is composed of a neural machine translation and a large language model.

[0013] Optionally, the method further includes: during the translation iteration, adjusting a construction direction of the evaluation question set of the next round based on difference information between the revised translation draft of the current round and the translation draft of the previous round.

[0014] Another aspect of the embodiments of the present application provides a translation device, the device comprising: a generation module configured to receive a text to be translated, and generate a translation draft based on the text to be translated by using a large language model; an iteration module configured to perform a translation iteration by taking the translation draft as a translation draft of a current round at the beginning of the iteration, and obtain a translation result in a case where an iteration result after the iteration meets a preset quality requirement or reaches a preset number of iterations, the translation iteration process including constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and continuing the iteration by taking a revised translation draft as a translation draft of a next round in a case where the revised translation draft does not meet the preset quality requirement and the preset number of iterations has not been reached. an output module configured to output a translated text based on the translation result.

[0015] Another aspect of the embodiments of the present application provides a computer device, comprising: at least one processor; and a memory in communication connection with the at least one processor; Wherein: the memory stores instructions that can be executed by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described above.

[0016] Another aspect of this application provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method described above.

[0017] Another aspect of this application provides a computer program product including a computer program that, when executed by a processor, implements the method described above.

[0018] The embodiments of this application employing the above-described technical solution may have the following advantages: By leveraging a large language model to generate initial translation drafts, multiple iterations are performed on these drafts. During each iteration, an evaluation question set is constructed based on the current draft. Structured feedback information is generated based on this evaluation question set and the current draft. The draft is then revised based on this feedback. If the revised draft does not meet the preset quality requirements, it is used as the next draft for further iteration. The final translation is obtained when the revised draft meets the preset quality requirements or reaches the preset number of iterations. Finally, the translated text is output. This multi-round iteration process—question set construction, structured feedback, and revision—ensures translation quality and improves the accuracy and consistency of the translation. Furthermore, the structured feedback intuitively drives the automated revision process and enables optimized automation, meeting the needs of large-scale applications. Attached Figure Description

[0019] The accompanying drawings exemplify embodiments and form part of the specification, serving together with the textual description to explain exemplary implementations of the embodiments. The illustrated embodiments are for illustrative purposes only and do not limit the scope of the claims. Throughout the drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0020] Figure 1 The diagram schematically illustrates the operating environment of the translation method according to Embodiment 1 of this application; Figure 2 A flowchart illustrating a translation method according to Embodiment 1 of this application is shown schematically; Figure 3 Schematic illustration Figure 2 Flowchart of the sub-steps in step S202; Figure 4 Schematic illustration Figure 2 Flowchart of another sub-step in step S202; Figure 5A flowchart illustrating a translation method according to Embodiment 1 of this application is shown schematically. Figure 6 A block diagram of a translation apparatus according to Embodiment 2 of this application is schematically shown; and Figure 7 A schematic diagram of the hardware architecture of a computer device according to Embodiment 3 of this application is shown. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that the descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0023] It should be noted that, in any stage of this application involving the collection, storage, use, transmission, and processing of data, each stage strictly adheres to the laws, regulations, industry standards, and regulatory requirements of the data source, usage location, and relevant countries and regions to ensure the legality and compliance of data activities. In the collection stage, the purpose, method, and scope of collection are clearly communicated to the data subject in a prominent manner. Collection is conducted only after obtaining the data subject's legal authorization, ensuring that the collection process follows the "minimum necessary" principle and does not exceed the scope of data collection. In the storage stage, storage periods are limited, and data is promptly deleted or anonymized / encrypted after the storage purpose is achieved. In the usage stage, a strict data security protection mechanism is implemented, using field-level desensitization technology and processing the original data according to preset desensitization rules. For different types of data, multiple desensitization strategies, such as data generalization, data anonymization, and data encryption, are employed to effectively mitigate the risk of sensitive information leakage and ensure that all data used is securely processed and desensitized, comprehensively protecting the rights and interests of data subjects and data security. In the transmission and processing stages, the confidentiality and security of data are ensured during transmission and processing.

[0024] In the description of this application, it should be understood that the numerical labels before the steps do not indicate the order of the steps, but are only used to facilitate the description of this application and to distinguish each step, and therefore should not be construed as a limitation of this application.

[0025] First, a definition of the terminology used in this application is provided: An intelligent agent is an agent that can perceive its environment and take actions to achieve specific goals. In the field of artificial intelligence, it is usually based on a large language model and has the ability to autonomously perceive, understand, plan, make decisions, remember, act, and use tools. It can automatically perform complex tasks with less human intervention and exhibits characteristics such as autonomy, interactivity, responsiveness, and adaptability.

[0026] Semantic fidelity is a core evaluation indicator in the field of translation, referring to the degree to which the translation matches the original text in terms of core meaning, logical relationships, emotional tone, and implicit information.

[0027] Neural Machine Translation (NMT) is a machine translation method that uses artificial neural networks to probabilistically model and generate translations from source language sentences to target language sentences. It typically models entire sentences end-to-end in a single ensemble model, employing an encoder-decoder structure combined with autoregressive decoding to generate translations word by word.

[0028] Secondly, to facilitate understanding of the technical solutions provided in the embodiments of this application by those skilled in the art, the relevant technologies are described below: In machine translation, large language models are typically used to directly generate translations based on a single prompt. However, due to the lack of quality verification mechanisms, the accuracy and consistency of the translations are low, resulting in low quality. Other methods rely on large language models for quality assessment, scoring or labeling translations using these models. However, the assessment results are open-ended texts lacking structured representation, making it difficult to intuitively drive automated revision processes. Still other methods are human-led prompt optimization approaches, depending on the translator's experience to iteratively optimize the translation prompt templates. However, this reliance on human experience limits its scalability.

[0029] Therefore, this application provides a translation technology solution. In this solution, a preliminary translation draft is generated using a large language model. This draft is then iterated through multiple rounds. During each iteration, an evaluation question set is constructed based on the current draft, structured feedback information is generated, and the draft is revised based on this feedback. Finally, the translated text is output based on the iterative translation results. This multi-round iteration—question set construction, structured feedback, and revision—ensures translation quality and improves the accuracy and consistency of the translation. Furthermore, structured feedback can intuitively drive an automated revision process and achieve optimized automation, meeting the needs of large-scale applications. Details are provided below.

[0030] Finally, for ease of understanding, an exemplary operating environment is provided below.

[0031] like Figure 1 As shown in the diagram, the runtime environment includes service platform 2, network 4, and client 6, where: Service platform 2 can consist of one or more computing devices. These computing devices may include virtualized computing instances. Virtualized computing instances may include virtual machines, such as emulations of computer systems, operating systems, servers, etc. The computing devices can load virtual machines based on virtual images and / or other data that define specific software used for emulation (e.g., operating systems, dedicated applications, servers). As the demand for different types of processing services changes, different virtual machines can be loaded and / or terminated on one or more computing devices. A hypervisor can be implemented to manage the use of different virtual machines on the same computing device.

[0032] Service platform 2 can be configured to communicate with clients 6, etc., via network 4. Network 4 includes various network devices, such as routers, switches, multiplexers, hubs, modems, bridges, repeaters, firewalls, proxy devices, and / or similar devices. Network 4 may include physical links, such as coaxial cable links, twisted pair cable links, fiber optic links, or combinations thereof, or wireless links, such as cellular links, satellite links, Wi-Fi links, etc.

[0033] Service platform 2 can provide services such as storage, reading, writing, querying, and deleting, such as providing translation services to clients.

[0034] Client 6 can be an electronic device running operating systems such as Windows, Android™, or iOS, including smartphones, tablets, laptops, virtual reality devices, gaming devices, set-top boxes, in-vehicle terminals, and smart TVs. Based on these operating systems, various applications can run, such as translation applications.

[0035] Client 6 can provide / configure user access pages, which can be used to control service platform 2 or upload objects, etc.

[0036] It should be noted that the above-mentioned equipment is exemplary, and the number and type of equipment can be adjusted in different scenarios or according to different needs.

[0037] The technical solutions of this application are described below through multiple embodiments, using a service platform as the implementing entity. It should be understood that these embodiments can be implemented in many different forms and should not be construed as being limited to the embodiments described herein.

[0038] Example 1 Figure 2 A flowchart illustrating a translation method according to Embodiment 1 of this application is shown schematically.

[0039] like Figure 2 As shown, the translation method may include steps S200~S204, wherein: Step S200: Receive the text to be translated and generate a first draft of the translation based on the text using a large language model.

[0040] Step S202: The initial translation draft is used as the translation draft for the current round at the start of the iteration to perform translation iteration. If the translated draft after iteration meets the preset quality requirements or reaches the preset number of iterations, the translation result is obtained. The translation iteration process includes constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and if the revised translation draft does not meet the preset quality requirements, the revised translation draft is used as the translation draft for the next round to continue iteration.

[0041] Step S204: Output the translated text based on the translation results.

[0042] The translation method provided in this embodiment generates an initial translation draft using a large language model, iterates through the draft multiple times, constructs an evaluation question set based on the current draft, generates structured feedback information based on the evaluation question set and the current draft, and revises the draft based on the structured feedback information. If the revised draft does not meet the preset quality requirements, it is used as the next draft for iteration. The translation result is obtained when the iterated draft meets the preset quality requirements or reaches the preset number of iterations. Finally, the translated text is output. This multi-round iteration of question set construction, structured feedback, and revision ensures translation quality and improves the accuracy and consistency of the translation. Furthermore, the structured feedback intuitively drives the automated revision process and enables optimized automation, meeting the needs of large-scale applications.

[0043] The following combination Figure 2 The steps in steps S200 to S204, as well as other optional steps, are described in detail.

[0044] Step S200 It receives the text to be translated and uses a large language model to generate a first draft of the translation based on the text.

[0045] Specifically, multiple knowledge bases can be used to search based on the text to be translated. Based on the search results, a prompting project can be generated. Then, the text to be translated and the prompting project are input into a large language model, which generates the first draft of the translation. The knowledge base can be, for example, a knowledge base for terminology, sentence structure, or style. When the text to be translated is received, key terms such as nouns, verb phrases, or technical terms can be extracted. Relevant entries are then retrieved from the knowledge base, such as reference translations of similar sentence structures and terminology definitions. The retrieved knowledge is organized into a "guidance text" as the prompting project, ready to be injected into the large language model.

[0046] Step S202 The initial translation draft is used as the translation draft for the current round at the start of the iteration. Translation iteration is performed. If the translated draft meets the preset quality requirements or reaches the preset number of iterations, the translation result is obtained. The translation iteration process includes constructing an evaluation question set based on the current round of translation draft, generating structured feedback information based on the evaluation question set and the current round of translation draft, revising the current round of translation draft based on the structured feedback information, and if the revised translation draft does not meet the preset quality requirements and has not reached the preset number of iterations, the revised translation draft is used as the translation draft for the next round of iteration.

[0047] When constructing an evaluation question set based on the current round of translations, the questions in the question set may include, for example, whether there are mistranslations, omissions, whether the terminology is consistent with the terminology knowledge base or the terminology is consistent with the context, whether the translation is fluent, whether the role style is appropriate, and whether the format is standardized.

[0048] In an optional embodiment, the evaluation question set covers at least two dimensions among term consistency, style matching, logical coherence, sentence pattern standardization, and semantic faithfulness. Preferably, the evaluation question set can include these five dimensions: term consistency, style matching, logical coherence, sentence pattern standardization, and semantic faithfulness. Questions corresponding to term consistency are, for example: Are there any term inconsistencies? Please point out the specific problems. Questions corresponding to style matching are, for example: Does the translation match the style of the previous character? Questions corresponding to logical coherence are, for example: Is there any logical contradiction between this sentence and the previous one? Questions corresponding to sentence pattern standardization are, for example: Are there more natural sentence patterns that conform to the target language expression? Questions corresponding to semantic faithfulness are, for example: Are there any possible omissions or additional additions? It can be understood that the above are only exemplary explanations, and the actual questions can be other similar or related questions. When constructing the evaluation question set, it can be constructed according to some predefined question templates; optionally, a lightweight large language model can also be used, and some prompt words can be used to construct the question set. Specific prompt words can be, for example: Please act as an expert in constructing translation questions and generate a list of questions in five dimensions: term consistency, style matching, logical coherence, sentence pattern standardization, and semantic faithfulness, for guiding subsequent precise inspections...

[0049] After constructing the evaluation question set, each question in the evaluation question set can be identified for the current round of translation draft. In the case of identifying specific problems, the specific problems can be marked, and specific revision suggestions can also be given. On this basis, structured feedback information can be generated based on the identified problems, marked problems, and specific revision suggestions, so that the current round of translation draft can be precisely revised directly according to the structured feedback information in the future. For example, the following problems can be identified and revision suggestions can be given for a certain translation fragment according to the evaluation question set: A1: The style is weak, and the character should be more arrogant. It is recommended to change "I think it's okay" to "Humph, I don't just think... I'm not... not at all"; A2: The term "accommodation house" should be uniformly translated as "training camp", and there is an inconsistency here; A3: There is a sudden change in perspective between the upper and lower sentences, and it should be unified to the first person. Optionally, a lightweight large language model can be used, and some prompt words can be used to generate structured feedback information. Specific prompt words are, for example: Please answer according to the evaluation question set, mark the problems where the problems are located, give specific revision suggestions, and output these contents in JSON format. The content in JSON format is as follows...

[0050] In an optional embodiment, the structured feedback information includes at least the error location range, error type, and revision suggestion fields, and the structured feedback information is in JSON format. The JSON-formatted structured feedback information allows for direct parsing, and the current round of translation revisions can be based on the parsing results.

[0051] The following is a specific example of the output of a JSON structured feedback message: {"term_inconsistency":[{"span":[0,10],"detected":"training camp","correct":"training camp","reason":"inconsistent with terminology database"}], "style_issues":[{"span":[12,25],"detected":"I think it's okay","suggest":"Hmph, I don't think...I, I don't...","reason":"It should reflect a tsundere style"}], "coherence_errors":[{"span":[30,45],"issue":"sudden change in perspective","suggest":"consistently use first-person perspective"}], "fidelity_issues":[{"type":"omission","span":[50,55],"missing_content":"exclamatory tone in the original text"}]} Among them, the top-level keys are term_inconsistency (term inconsistency), style_issues (style issues), coherence_errors (coherence errors), and fidelity_issues (fidelity issues), which are the four dimensions of the evaluation problem set. This classification provides a routing basis for subsequent revisions. The values corresponding to the top-level keys are arrays. In the content corresponding to the top-level key of term_inconsistency: "span:[0,10]" indicates the position from the 0th character to the 10th character in the current round of translation draft; "detected: 合宿所" indicates the incorrect term "合宿所" in the current translation; "correct: 训练营地" indicates that the correct translation should be "training camp". "reason: inconsistent with the term library" indicates that the reason for the error is "inconsistent with the term library", enhancing interpretability. In the content corresponding to the top-level key of style_issues: "span:[12,25]" indicates the position from the 12th character to the 25th character in the current round of translation draft; "detected: 我觉得可以" indicates that the original text with inconsistent style is "我觉得可以", "suggest: 哼,我才不是觉得...才、才不是..." indicates that the revision suggestion is to change it to "哼,我才不是觉得...才、才不是...", "reason: should reflect the arrogant style" indicates that the reason for the error is "should reflect the arrogant style". In the content corresponding to the top-level key of coherence_errors: "span:[30,45]" indicates the position from the 30th character to the 45th character in the current round of translation draft; "issue: perspective mutation" indicates the nature of the problem is perspective mutation; "suggest: unified to the first person" indicates that the revision suggestion is "unified to the first person". In the content corresponding to the top-level key of fidelity_issues: "type: omission" indicates that the type of the sub-issue is "omission (missing translation)"; "span:[50,55]" indicates the position from the 50th character to the 55th character in the current round of translation draft; "missing_content: the exclamatory tone in the original text" indicates that the missing content is the exclamatory tone in the original text.

[0052] Since the structured feedback information includes fields such as the error location interval, error type, and revision suggestion, when revising the current round of translation draft based on the structured feedback information, it is possible to locate according to the error location interval and then make specific revisions according to the error type and revision suggestion.

[0053] In an optional embodiment, in step S202, when revising the current round of translation draft based on the structured feedback information, as Figure 3 shown, it may further include: Step S300, determining the revised segment in the initial translation draft based on the error location interval.

[0054] Step S302: Perform a partial rewrite of the revised fragment based on the error type and revision suggestion fields.

[0055] Specifically, the revised segments in the initial translation draft can be located based on the error location range in the structured feedback information, the segments to be revised in the initial translation draft can be determined, and then the revised segments can be rewritten locally according to the error type and revision suggestions.

[0056] Taking the above example of structured feedback output, the revised segments in the initial translation draft can be determined based on the error location range corresponding to the "span" field. For partial rewriting, for example, replacing "shukusho" with "training camp" and changing "I think it's okay" to "Hmph, I don't think... I, I don't...".

[0057] Optionally, partial rewriting can be performed directly based on the suggested revision fields, or partial rewriting can be performed in combination with the error type. For example, if the error type is terminology inconsistency, terminology substitution can be performed; if the error type is style problem, partial rewriting can be performed on the error range; if the error type is sentence error, template rewriting can be performed; if the error type is semantic error, global or multi-sentence rewriting can be performed.

[0058] In the embodiments corresponding to steps S300-S302, by determining the segments to be revised in the translated manuscript based on the error location range, and by locally rewriting the segments to be revised based on the error type and revision suggestion fields, targeted local rewriting can be performed based on structured feedback information, avoiding full-text retranslation, improving the efficiency of translation correction, and maintaining the stability of the already correct parts.

[0059] In an optional embodiment, in step S202, the initial translation draft is used as the translation draft for the current round at the start of the iteration to perform translation iterations, such as... Figure 4 As shown, it may further include: Step S400: The feedback agent generates structured feedback information based on the evaluation question set and the translation of the current round.

[0060] Step S402: Use the revision agent to revise the translation of the current round based on structured feedback information.

[0061] Specifically, a feedback agent can be used to identify problems, pinpoint their locations, analyze them, and provide specific revision suggestions for the current round of translation based on the problems in the evaluation problem set. This process can be carried out according to pre-configured strategies, such as "first perform terminology scanning, then analyze style sentence by sentence, then check logical coherence, etc." When locating problems, string matching or location tools can be used to pinpoint their exact location. Problem identification can follow predetermined strategies; for example, for terminology consistency issues, terms in the current round of translation can be compared with those in a terminology knowledge base; for style fit issues, tone, attitude, and other style-related words and phrases in the current round of translation can be compared with corresponding characters in a style knowledge base, and so on. Problem analysis and specific revision suggestions can be based on the problem identification process. For example, for terminology consistency issues, if a term is found to be inconsistent with a term in the terminology knowledge base during the comparison process, the revision suggestion can be to replace that term with one from the terminology knowledge base, with the error reason being "inconsistent with the terminology knowledge base." After all the questions in the evaluation question set have been processed, structured feedback information is generated based on the processing results.

[0062] After generating structured feedback information, the feedback agent can send it to the revision agent. The revision agent parses the structured feedback information, locates the specific problem, and revises the corresponding problem according to the revision suggestions in the structured feedback information. After all structured feedback information has been processed, the revised translation can be judged. If the revised translation does not meet the preset quality requirements and has not reached the preset number of iterations, the revised translation will be used as the next round of translation for iteration. The preset quality requirements can be a comprehensive score of terminology consistency, style fit, semantic fidelity, sentence structure matching, and logical coherence, or a threshold can be set for each dimension, and whether the preset quality requirements are met is determined based on whether each dimension reaches the threshold. In an optional embodiment, the preset quality requirements include one or more of terminology consistency, style fit, semantic fidelity, and sentence structure matching reaching a preset threshold. For example, the threshold for terminology consistency can be 100%, and a terminology consistency of 100% is considered to meet the requirements; the threshold for sentence structure matching can be 80%, and a sentence structure matching greater than or equal to 80% is considered to meet the requirements; the threshold for style fit can be 70%, and a style fit greater than or equal to 70% is considered to meet the requirements; and the threshold for semantic fidelity can be 95%, and a semantic fidelity greater than or equal to 95% is considered to meet the requirements.

[0063] In some possible embodiments, the construction of the evaluation question set based on the translation of the current round in step S202 can also be implemented by an independent agent, for example, by a questioning agent constructing the evaluation question set based on the translation of the current round.

[0064] In some possible embodiments, the feedback agent can be a combination of multiple models, which generate structured feedback information through a voting mechanism. These models can be of the same architecture or different architectures; preferably, different architectures can be used to achieve model diversity and reduce systematic bias. Multiple models simultaneously analyze the translation in the current round and then vote on each question in the evaluation question set. The majority vote determines whether to generate the corresponding structured feedback information. For example, the feedback agent might consist of three models, where two models consider a term inconsistent, and one model considers it consistent. In this case, the majority vote determines that the term is inconsistent, and the corresponding structured feedback information is generated. It is understood that implementing the feedback agent through a combination of multiple models, with multiple models generating structured feedback information through a voting mechanism, can improve the reliability and accuracy of the generated structured feedback information and avoid the limitations and biases of a single model.

[0065] In some possible embodiments, the revision agent consists of neural machine translation and a large language model. Neural machine translation and the large language model can each have different functions. For example, when the structured feedback information points to local, explicitly instructed problems (such as terminology substitution or specific sentence style adjustments), the revision agent invokes the large language model for precise revision of the relevant segments. When the structured feedback information points to global, deep semantic biases or large-scale quality problems, the revision agent invokes neural machine translation to perform large-scale or overall retranslation of the corresponding source text. It is understood that the revision agent, composed of neural machine translation and a large language model, can simultaneously leverage the advantages of neural machine translation in translation fluency and stability, and the advantages of the large language model in local rewriting, achieving an optimal balance between ensuring the quality of the final translation and the revision effect.

[0066] In the embodiments corresponding to steps S400-S402, by using a feedback agent to generate structured feedback information based on the evaluation question set and the current round of translation, and using a revision agent to revise the current round of translation based on the structured feedback information, the collaboration of multiple agents can help achieve translation iteration, avoid confirmation bias that may be caused by a single model, and improve the accuracy of translation.

[0067] In optional embodiments, the translation method of this application may further include: during the translation iteration process, adjusting the construction direction of the evaluation question set for the next round based on the difference information between the revised current-round translation and the previous-round revised translation. Specifically, the revised current-round translation can be compared with the previous-round revised translation to determine the differences between the two, including the positions and contents of additions, deletions, and modifications; then, the successfully executed parts and potentially affected surrounding contexts can be analyzed, and the next round of question set can be constructed based on the analysis results. For example, for the successfully executed parts, verification questions can be added to ensure that the revision is error-free and without side effects; for potentially affected surrounding contexts, context review questions can be added to check logical coherence; for some unaffected parts, the check density can be appropriately reduced. It is understood that by adjusting the construction direction of the evaluation question set for the next round based on the difference information between the revised current-round translation and the previous-round revised translation during the translation iteration process, the construction of the evaluation question set can be made more targeted, thereby improving the quality of the iteration and the quality of the final translation.

[0068] Step S204 The translated text is output based on the translation results.

[0069] After obtaining the translation results, if the results meet the preset quality requirements, they can be directly output as the final translated text. Alternatively, further processing can be performed, such as checking the engineering format, before finally outputting the translated text. If the translation results do not meet the preset quality requirements but were obtained after reaching the preset number of iterations, manual revision or other revision methods can be used before outputting the final translated text. Alternatively, the translation results can be directly output as the translated text, marked as low confidence for user reference, and potential problem annotations can be provided so that users can revise and use the text accordingly.

[0070] To make this application easier to understand, the following is combined with... Figure 5 An example application is provided.

[0071] Figure 5 This is a flowchart illustrating a translation method according to an embodiment of this application, which may generally include the following: 1. The input layer receives hierarchical text and knowledge points (terminology, sentence structure, style) constraints to generate the initial translation draft V0.

[0072] 2. The initial draft enters the review and iteration cycle, which includes the following steps: ① Automatically construct a set of review questions Q (corresponding to the evaluation question set) for multi-dimensional quality (such as style, terminology, coherence, etc.); ② The LLM reviewer analyzes the current translation Vn based on the review question set Q and outputs a response that includes error location and improvement suggestions; ③ The feedback extractor transforms the above answers into structured, machine-executable feedback information (such as JSON format). ④ The revising agent rewrites the translation locally or globally based on the structured feedback, generating a revised translation V_{n+1}.

[0073] The revised translation is sent to the convergence determination module and compared with preset multi-dimensional quality thresholds (terminology consistency, sentence structure matching, style fit, and semantic fidelity). If all dimensions meet the criteria, the loop is exited, and a high-quality translation V_final, along with complete proofreading records and version differences, is output. If the criteria are not met but the number of iterations has not exceeded the limit, the loop returns to S1 to start the next iteration. If the number of iterations exceeds the limit, the current optimal version is forcibly output.

[0074] In this exemplary application, the quality of the translation is ensured and the accuracy and consistency of the translation are improved through multiple iterations of question set construction, structured feedback, and revision. At the same time, the structured feedback can intuitively drive the automated revision process and also achieve optimized automation to meet the needs of large-scale applications.

[0075] Example 2 Figure 6 The diagram schematically illustrates a translation apparatus according to Embodiment 2 of this application. This apparatus can be divided into one or more program modules. One or more program modules are stored in a storage medium and executed by one or more processors to complete the embodiments of this application. The program module referred to in the embodiments of this application refers to a series of computer program instruction segments capable of performing a specific function. The following description will specifically introduce the function of each program module in this embodiment. For example... Figure 6 As shown, the device 500 may include a generation module 510, an iteration module 520, and an output module 530, wherein: The generation module 510 is used to receive the text to be translated and generate a preliminary translation based on the text using a large language model. The iteration module 520 is used to perform translation iteration with the initial translation draft as the translation draft of the current round at the start of the iteration. If the translated draft after iteration meets the preset quality requirements or reaches the preset number of iterations, the translation result is obtained. The translation iteration process includes constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and if the revised translation draft does not meet the preset quality requirements and has not reached the preset number of iterations, the revised translation draft is used as the translation draft of the next round for continued iteration. The output module 530 is used to output the translated text based on the translation result.

[0076] In an optional embodiment, the evaluation question set covers at least two dimensions of terminology consistency, style matching, logical coherence, sentence structure normativity, and semantic fidelity.

[0077] In an optional embodiment, the structured feedback information includes at least the error location range, error type, and revision suggestion fields, and the structured feedback information is in JSON format.

[0078] In an optional embodiment, the iteration module 520 is further configured to: Based on the error location range, determine the segments in the translated manuscript that need revision; The fragment to be revised is partially rewritten based on the error type and the revision suggestion field.

[0079] In an optional embodiment, the preset quality requirements include one or more of the following reaching a preset threshold: terminology consistency, style fit, semantic fidelity, and sentence structure matching.

[0080] In an optional embodiment, the iteration module 520 is further configured to: The feedback agent generates structured feedback information based on the evaluation question set and the translation of the current round; The revision agent revises the translation of the current round based on the structured feedback information.

[0081] In an optional embodiment, the feedback agent is a combination of multiple models, which generate the structured feedback information through a voting mechanism.

[0082] In an optional embodiment, the revised agent consists of neural machine translation and a large language model.

[0083] In an optional embodiment, the iteration module 520 is further configured to: During the translation iteration process, the direction of constructing the evaluation question set for the next round is adjusted based on the differences between the revised translation of the current round and the revised translation of the previous round.

[0084] Example 3 Figure 7This illustration schematically shows a hardware architecture diagram of a computer device 10000 suitable for implementing a translation method according to Embodiment 3 of this application. In some embodiments, the computer device 10000 may be a terminal device such as a smartphone, wearable device, tablet computer, personal computer, in-vehicle terminal, game console, virtual device, workbench, digital assistant, set-top box, robot, etc. In other embodiments, the computer device 10000 may be a rack server, blade server, tower server, or cabinet server (including independent servers or server clusters composed of multiple servers), etc. Figure 7 As shown, the computer device 10000 includes, but is not limited to: a memory 10010, a processor 10020, and a network interface 10030 that can communicate and be linked with each other via a system bus. Wherein: The memory 10010 includes at least one type of computer-readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 10010 may be an internal storage module of a computer device 10000, such as the hard disk or memory of the computer device 10000. In other embodiments, the memory 10010 may also be an external storage device of the computer device 10000, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 10000. Of course, the memory 10010 may also include both the internal storage module and the external storage device of the computer device 10000. In this embodiment, the memory 10010 is typically used to store the operating system and various application software installed on the computer device 10000, such as the program code for translation methods. In addition, the memory 10010 can also be used to temporarily store various types of data that have been output or will be output.

[0085] In some embodiments, processor 10020 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other chip. Processor 10020 is typically used to control the overall operation of computer device 10000, such as performing control and processing related to data interaction or communication with computer device 10000. In this embodiment, processor 10020 is used to run program code stored in memory 10010 or process data.

[0086] Network interface 10030 may include a wireless network interface or a wired network interface, which is typically used to establish a communication link between computer device 10000 and other computer devices. For example, network interface 10030 is used to connect computer device 10000 to an external terminal via a network, establishing a data transmission channel and communication link between computer device 10000 and the external terminal. The network may be an intranet, the Internet, Global System for Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G network, Bluetooth, Wi-Fi, or other wireless or wired networks.

[0087] It should be pointed out that, Figure 7 Only computer devices with components 10010-10030 are shown; however, it should be understood that it is not required to implement all of the shown components, and more or fewer components may be implemented instead.

[0088] In this embodiment, the translation method stored in memory 10010 can also be divided into one or more program modules and executed by one or more processors (such as processor 10020) to complete the embodiment of this application.

[0089] Example 4 This application also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the translation method in the embodiments.

[0090] In this embodiment, the computer-readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the computer-readable storage medium can be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the computer-readable storage medium can also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device. Of course, the computer-readable storage medium can also include both the internal storage unit and the external storage device of the computer device. In this embodiment, the computer-readable storage medium is typically used to store the operating system and various application software installed on the computer device, such as the program code of the translation method in the embodiment. In addition, the computer-readable storage medium can also be used to temporarily store various types of data that have been output or will be output.

[0091] Example 5 This application also provides a computer program product, including a computer program that, when executed by a processor, implements the methods described in the above embodiments.

[0092] Obviously, those skilled in the art should understand that the modules or steps of the embodiments of this application described above can be implemented using general-purpose computer devices. They can be centralized on a single computer device or distributed across a network of multiple computer devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computer device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the embodiments of this application are not limited to any particular combination of hardware and software.

[0093] It should be noted that the above are merely preferred embodiments of this application and do not limit the scope of patent protection of this application. Any equivalent structural or procedural changes made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this application.

Claims

1. A translation method, characterized in that, The method includes: Receive the text to be translated and generate a preliminary translation based on the text using a large language model; The initial translation draft is used as the translation draft for the current round at the start of the iteration. Translation iteration is performed. If the translated draft after iteration meets the preset quality requirements or reaches the preset number of iterations, the translation result is obtained. The translation iteration process includes constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and if the revised translation draft does not meet the preset quality requirements and has not reached the preset number of iterations, the revised translation draft is used as the translation draft for the next round of iteration. The translated text is output based on the translation results.

2. The method according to claim 1, characterized in that, The evaluation question set covers at least two dimensions of terminology consistency, style matching, logical coherence, sentence structure standardization, and semantic fidelity.

3. The method according to claim 1, characterized in that, The structured feedback information includes at least the error location range, error type, and revision suggestion fields.

4. The method according to claim 3, characterized in that, The revision of the current round of translation based on the structured feedback information includes: Based on the error location range, determine the segments in the translated manuscript that need revision; The fragment to be revised is partially rewritten based on the error type and the revision suggestion field.

5. The method according to claim 1, characterized in that, The preset quality requirements include one or more of the following: terminology consistency, style fit, semantic fidelity, and sentence structure matching, all of which must reach a preset threshold.

6. The method according to claim 1, characterized in that, The step of performing translation iterations using the initial translation draft as the translation draft for the current round at the start of the iteration includes: The feedback agent generates structured feedback information based on the evaluation question set and the translation of the current round; The revision agent revises the translation of the current round based on the structured feedback information.

7. The method according to claim 6, characterized in that, The feedback agent is a combination of multiple models, which generate the structured feedback information through a voting mechanism.

8. The method according to claim 6, characterized in that, The revised intelligent agent consists of neural machine translation and a large language model.

9. The method according to claim 1, characterized in that, The method further includes: During the translation iteration process, the direction of constructing the evaluation question set for the next round is adjusted based on the differences between the revised translation of the current round and the revised translation of the previous round.

10. A translation device, characterized in that, The device includes: The generation module is used to receive the text to be translated and generate a preliminary translation based on the text using a large language model. An iteration module is used to perform translation iterations with the initial translation draft as the translation draft for the current round at the start of the iteration. If the translated draft after iteration meets the preset quality requirements or reaches the preset number of iterations, a translation result is obtained. The translation iteration process includes constructing an evaluation question set based on the translation draft of the current round, generating structured feedback information based on the evaluation question set and the translation draft of the current round, revising the translation draft of the current round based on the structured feedback information, and if the revised translation draft does not meet the preset quality requirements and has not reached the preset number of iterations, the revised translation draft is used as the translation draft for the next round of iterations. The output module is used to output the translated text based on the translation result.

11. A computer device, characterized in that, include: At least one processor; and A memory communicatively connected to the at least one processor; wherein: The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 9.

Images