Low-resource language translation method and system based on inference and retrieval fusion large model

Abstract

The application provides a large model low-resource language translation method and system based on reasoning and retrieval fusion, which comprises the following steps: obtaining enhancement information, which contains retrieved similar parallel sentence pairs and extracted auxiliary knowledge; performing first-stage implicit correction, constructing a comparative demonstration context containing retrieved example source sentences, auxiliary knowledge, initial translation and standard reference translation, inputting the large model with the to-be-translated source sentence and initial translation, and generating first-stage corrected translation; performing second-stage explicit correction, performing error detection and labeling on the first-stage corrected translation based on the large model according to multi-dimensional quality measurement standards, generating structured feedback, and inputting the large model again to generate second-stage corrected translation. The application effectively improves the accuracy and robustness of low-resource language translation.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically providing a large-model, low-resource language translation method and system based on the fusion of reasoning and retrieval. Background Technology

[0002] With increasingly frequent global exchanges, machine translation, as a fundamental task in the field of Natural Language Processing (NLP), has received widespread attention. Traditional neural machine translation (NMT) systems heavily rely on large-scale bilingual parallel corpora and task-specific supervised training. However, in real-world applications, the vast majority of languages ​​worldwide are low-resource languages, making it difficult to obtain sufficient high-quality parallel data. This leads to traditional methods often failing to be fully trained in low-resource scenarios due to data hunger, resulting in limitations such as missing word meanings and difficulties in semantic alignment caused by data scarcity, significantly reducing translation quality.

[0003] In recent years, the emergence of large language models (LLMs) has brought about a new paradigm shift in machine translation. Through context learning (ICL) capabilities, large language models can directly perform translation tasks in zero-shot or few-shot settings without the need for costly parameter updates. While this data-driven generation approach demonstrates strong multilingual generalization potential, in extremely low-resource scenarios, although large language models possess multilingual understanding capabilities, the proportion of extremely low-resource languages ​​in their pre-training corpora is minimal. During zero-shot translation, due to the lack of sufficient domain knowledge and specific linguistic constraints in static prompts, the model-generated translations often suffer from severe illusions and semantic biases. This means generating fluent but meaningless text unrelated to the original meaning of the source sentence, or, when faced with proper nouns or domain-specific terms, producing severe disambiguation errors and local semantic deviations due to a lack of background knowledge.

[0004] To overcome the limitations of static context and compensate for the insufficient knowledge coverage of large models, the current mainstream optimization approach is to introduce Retrieval Augmentation (RAG) technology. This type of method dynamically retrieves semantically relevant parallel sentence pairs, terminology dictionaries, or error examples from external corpora and inputs them into the model as contextual examples. With the help of this explicit external knowledge, the model can more effectively handle low-frequency vocabulary and domain-specific translation tasks. Its advantage lies in significantly improving the adaptability to the translation of specialized terminology and low-resource languages. However, this type of method still faces several challenges. First, due to the inherent scarcity of data and complex linguistic structures in low-resource languages, even with the help of retrieved examples, the initial translation generated by the model in a single iteration may still contain various errors, which are difficult to completely eliminate in a single decoding process. Secondly, the current common "piecing-together" example retrieval method lacks fine-grained control: on the one hand, the retrieved sentences may deviate from the source sentences in terms of domain background or proper nouns, and simple splicing can easily introduce irrelevant noise, interfering with the model's translation judgment; on the other hand, only providing examples as contextual input lacks explicit extraction and constraints on the core information of the source sentences (such as topic summary and key entities), which may cause the model to exhibit "attention shift" when generating long texts, affecting translation quality and consistency.

[0005] To address the limitations of single-output translation, a number of improved methods based on self-correction and feedback mechanisms have recently emerged. For example, by introducing fine-grained feedback or multidimensional quality metrics (MQM), large models are prompted to evaluate and rewrite their initial output. These methods treat translation as an iterative process, improving the fluency and accuracy of the final translation to some extent. However, in practice, the corrective capabilities of large models are often not fully explored: on the one hand, the complexity of error sources makes it difficult for a single round of explicit correction to completely resolve hidden semantic conflicts; on the other hand, existing retrieval mechanisms only treat external parallel sentence pairs as static references, failing to fully utilize their potential "contrastive correction signals" (i.e., the difference between the model's own translation and the actual reference translation) in the error correction process. Therefore, in current few-shot or retrieval-enhanced translation prompts, simply providing the model with correct or standard translations as positive examples—a one-way guidance approach—cannot make the model aware of its potential errors. Consequently, large models often exhibit inherent error patterns when dealing with low-resource languages, such as fixed mistranslations of specific words. Without a comparative evolutionary path from error to correctness, the model struggles to spontaneously cross these implicit error boundaries, thus limiting the effectiveness of self-correction.

[0006] While the aforementioned methods have promoted the development of low-resource machine translation to some extent, it is evident that numerous technical bottlenecks remain to be addressed in this field. First, high-quality supervisory signals are extremely scarce in low-resource scenarios, and existing methods still only utilize the retrieved parallel data superficially, failing to effectively extract implicit error correction patterns. Second, traditional correction strategies mostly rely on single-dimensional feedback, lacking joint constraints on structured knowledge (such as topics and keywords) and grammatical context. Third, single correction steps are prone to overlooking some aspects, failing to simultaneously address both major semantic errors and fine-grained fluency refinement. Finally, when faced with complex long sentences or high-density technical terms, existing augmented translation models are prone to cognitive overload, leading to deviations in the final emotional and semantic expression from the original. Therefore, overcoming the data scarcity problem in low-resource scenarios, effectively suppressing noise interference encountered by large language models in retrieval-enhanced translation, and establishing a closed-loop mechanism from implicit error avoidance to explicit quality correction to substantially improve the translation fidelity and robustness of extremely low-resource languages ​​are critical technical challenges that urgently need to be addressed in the field of natural language processing and machine translation. Summary of the Invention

[0007] In order to overcome the above-mentioned defects, the present invention is proposed to provide a solution or at least a partial solution to the above-mentioned problems.

[0008] In a first aspect, the present invention provides a large-model low-resource language translation method based on the fusion of reasoning and retrieval, comprising the following steps: obtaining enhanced information associated with the source sentence to be translated, wherein the enhanced information includes at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs; The first phase of implicit modification includes: A comparative demonstration context is constructed, comprising at least one comparative example unit, which consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated for the retrieved example source statement based on a large language model; and a standard reference translation corresponding to the retrieved example source statement. The source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context are input into the large language model to generate the first-stage corrected translation; Perform the second phase of explicit modifications, including: Based on the large language model, and according to the preset multidimensional quality measurement standards, the first-stage revised translation is subjected to error detection and annotation, and a structured quality feedback signal is generated. The source sentence to be translated, the first-stage corrected translation, and the quality feedback signal are input again into the large language model to generate the second-stage corrected translation.

[0009] Preferably, obtaining enhanced information associated with the source statement to be translated includes: A hybrid retrieval strategy is employed to retrieve similar parallel sentence pairs from external corpora, including: Using a dense retrieval method, based on a multilingual sentence embedding model, the source sentence to be translated and the source sentences in the external corpus are mapped into high-dimensional vectors respectively, and the cosine similarity is calculated. Select the two most similar parallel sentence pairs to form a dense search subset; Using a sparse retrieval method, the literal similarity between the source sentence to be translated and the source sentences in the external corpus is calculated, and the parallel sentence pair with the highest similarity is selected to form a sparse retrieval subset; The union of the dense retrieval subset and the sparse retrieval subset is used as the retrieved similar parallel sentence pairs.

[0010] Preferably, the auxiliary knowledge includes the first topic information, the first keyword information, the second topic information, and the second keyword information; Obtaining enhanced information associated with the source statement to be translated also includes: The source sentence to be translated is semantically parsed using a large language model to extract the first topic information and the first keyword information; A large language model is used to perform semantic analysis on the source sentences of each retrieval example in the retrieved similar parallel sentence pairs, and to extract the second topic information and the second keyword information.

[0011] Preferably, the comparison demonstration context : in, The source language sentence to be translated. This represents the set of auxiliary knowledge extracted. and These represent the initial translation and the actual reference translation of the search example, respectively. This is a preliminary translation of the sentence to be translated. For the k-th sentence retrieved, The index of the retrieved sentence.

[0012] Preferably, the explicit correction performed in the second phase further includes: Using the second-stage revised translation as the current translation, the steps of error detection and annotation, as well as generating the second-stage revised translation, are repeated to form an iterative revision process; The output of the (t+1)th round of the iterative correction process Represented as: in, The generation and processing of a large language model. Indicates the first Explicit feedback signals generated in each round The source language sentence to be translated. This is the second stage of the translation correction for the output of round t.

[0013] Preferably, in the dense retrieval method, the cosine similarity between the source sentence to be translated and the source sentences in the external corpus is calculated based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus. This indicates a feature encoding operation.

[0014] Preferably, in the sparse retrieval method, the literal similarity between the source sentence to be translated and the source sentences in the external corpus is calculated based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus.

[0015] Secondly, this invention provides a large-model, low-resource language translation system based on the fusion of reasoning and retrieval, comprising: The acquisition module is used to acquire enhanced information associated with the source sentence to be translated. The enhanced information includes at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs. The first implicit correction module, used to perform the first stage of implicit correction, includes: A comparative demonstration context is constructed, comprising at least one comparative example unit, which consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated for the retrieved example source statement based on a large language model; and a standard reference translation corresponding to the retrieved example source statement. The source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context are input into the large language model to generate the first-stage corrected translation; The second implicit correction module is used to perform the second stage of explicit correction, including: Based on the large language model, and according to the preset multidimensional quality measurement standards, the first-stage revised translation is subjected to error detection and annotation, and a structured quality feedback signal is generated. The source sentence to be translated, the first-stage corrected translation, and the quality feedback signal are input again into the large language model to generate the second-stage corrected translation.

[0016] Preferably, the acquisition module includes: The first acquisition submodule is used to map the source sentence to be translated and the source sentences in the external corpus into high-dimensional vectors based on a multilingual sentence embedding model using a dense retrieval method, and calculate the cosine similarity; the two parallel sentence pairs with the highest similarity are selected to form a dense retrieval subset; The second acquisition submodule is used to calculate the literal similarity between the source sentence to be translated and the source sentences in the external corpus using a sparse retrieval method, select the parallel sentence pair with the highest similarity to form a sparse retrieval subset, and take the union of the dense retrieval subset and the sparse retrieval subset as the retrieved similar parallel sentence pair.

[0017] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0018] The beneficial effects of this invention are as follows: At the knowledge extraction and retrieval level, this invention not only retrieves similar parallel sentence pairs from external corpora, but also extracts themes and keywords from source sentences and retrieved example sentences as auxiliary knowledge through specific algorithms to construct a consistent information structure across sentences. At the implicit correction level, a comparative demonstration context containing "source sentence - auxiliary knowledge - initial model translation - reference translation" is constructed. Unlike traditional methods that only provide correct examples, this invention displays the model's own "flawed initial translation" and "high-quality reference" side by side, utilizing the context learning capabilities of a large language model to guide the model to automatically capture its own error patterns and perform implicit self-correction, generating higher-quality intermediate translations. At the explicit correction level, a multidimensional quality metric (MQM) analysis and iterative feedback mechanism based on a large language model are introduced. Specifically, a large language model is used as a quality evaluator to perform fine-grained MQM error annotation on the implicitly corrected intermediate translation, accurately locating and identifying specific error types (such as accuracy and fluency errors) and their locations. Subsequently, the generated MQM annotation information is injected as a structured feedback signal to guide the model to perform multiple rounds of targeted rewriting and optimization of the translation, thereby eliminating residual fine-grained semantic biases. Attached Figure Description

[0019] The disclosure of this invention will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Furthermore, similar numbers in the drawings are used to denote similar components, wherein: Figure 1 This represents the average marginal benefit of the number of iterations in one embodiment of the present invention.

[0020] Figure 2 This is a comparative example of English-to-low-resource language (such as Armenian) translation error correction according to an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of a large-model, low-resource language translation method based on the fusion of reasoning and retrieval, according to an embodiment of the present invention. Detailed Implementation

[0022] Some embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0023] Example 1 like Figure 1-3 As shown, this invention discloses a large-model, low-resource language translation method based on the fusion of reasoning and retrieval, comprising the following steps: Step S1: Obtain enhanced information associated with the source sentence to be translated. The enhanced information includes at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs, so as to maintain the consistency of the information structure and provide knowledge clues for context learning.

[0024] In this embodiment, existing public datasets for low-resource machine translation are selected to construct a bilingual parallel corpus retrieval library for providing external examples and for evaluation. The collected corpus undergoes data preprocessing, format standardization, and feature extraction to ensure the standardization of the input data.

[0025] In one embodiment, step S1, obtaining enhancement information associated with the source statement to be translated, includes: Step S11: Employ a hybrid retrieval strategy to retrieve similar parallel sentence pairs from an external corpus, including: Using a dense retrieval method, deep semantic representations are extracted from an external retrieval pool to obtain two parallel sentence pairs that are most similar in semantic space. Specifically, based on a multilingual sentence embedding model, the source sentence to be translated and the source sentences in the external corpus are mapped into high-dimensional vectors, and cosine similarity is calculated. The two most similar parallel sentence pairs are selected to form a dense search subset.

[0026] Step S12: Utilize sparse retrieval to perform precise matching at the word level to bridge specific vocabulary gaps in low-resource languages ​​and extract the most relevant sentence pair in terms of literal features. Specifically, calculate the literal similarity between the source sentence to be translated and the source sentences in the external corpus, and select the parallel sentence pair with the highest similarity to form a sparse retrieval subset; The union of the dense retrieval subset and the sparse retrieval subset is taken as the retrieved similar parallel sentence pairs. That is... This independent hybrid retrieval design ensures both the semantic relevance of the context and the accurate targeting of low-frequency words.

[0027] In this embodiment, to address the scarcity of low-resource language data and the potential noise interference from single-source retrieval, a hybrid retrieval strategy of "external corpus retrieval + internal knowledge extraction" is established. Specifically, this mechanism includes two parallel information acquisition paths: One approach is the external retrieval path, which uses a multilingual sentence embedding model to map the source sentence to be translated into a high-dimensional vector, calculates the cosine similarity in a bilingual parallel corpus, and retrieves the Top-K semantically most similar parallel sentence pairs (including the source sentence and the reference translation), thereby providing external contextual examples.

[0028] Secondly, the internal extraction path leverages the semantic understanding and generation capabilities of the large language model. Through specific prompting engineering, it extracts two types of auxiliary knowledge from both the "source sentence to be translated" and the "retrieved source sentence": first, general topical information to constrain the scope of the translation domain; and second, keywords with entity meaning to maintain consistency of proper nouns. This collaborative strategy of "retrieval providing examples and knowledge providing constraints" ensures that the information input to the model is not only semantically similar but also highly consistent with the original sentence in terms of information structure and knowledge background, thus effectively avoiding the negative impact of retrieval noise.

[0029] In one embodiment, the dense retrieval method calculates the cosine similarity between the source sentence to be translated and the source sentences in the external corpus based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus. This indicates a feature encoding operation.

[0030] In one embodiment, the sparse retrieval method calculates the literal similarity between the source sentence to be translated and the source sentences in the external corpus based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus.

[0031] In one embodiment, the auxiliary knowledge includes the first topic information, the first keyword information, the second topic information, and the second keyword information; Obtaining enhanced information associated with the source statement to be translated also includes: The source sentence to be translated is semantically parsed using a large language model to extract the first topic information and the first keyword information; A large language model is used to perform semantic analysis on the source sentences of each retrieval example in the retrieved similar parallel sentence pairs, and to extract the second topic information and the second keyword information.

[0032] In this embodiment, the semantic understanding capabilities of a large language model are utilized to analyze the source sentence to be translated. and the retrieved set of parallel examples The model extracts general thematic information and keywords with entity meaning. These knowledge signals ensure the consistency of information structure between the input sentence and the retrieved example, effectively constraining the scope of the translation domain. Simultaneously, in zero-shot scenarios, the large language model performs preliminary decoding on both the source sentence to be translated and the source sentence of the retrieved example, generating an initial model translation with potential flaws.

[0033] Step S2: Perform the implicit correction in the first stage, including: Step S21: Construct a comparison demonstration context, which includes at least one comparison example unit. The comparison example unit consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated based on a large language model for the retrieved example source statement; and a standard reference translation corresponding to the retrieved example source statement. Step S22: Input the source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context into the large language model to generate the first-stage corrected translation.

[0034] In this embodiment, an implicit correction context based on comparative examples is constructed: To activate the self-correction capability of the large language model without fine-tuning, a four-tuple structure containing a comparison of "source sentence - extracted knowledge - incorrect translation - correct translation" is built. Specifically, for each retrieved parallel sentence pair, the large language model first performs a zero-sample initial translation, generating an initial translation with potential errors. Subsequently, the "retrieved source sentence," "extracted auxiliary knowledge," "model's initial translation," and "standard reference translation" are combined into a comparative example unit. During the inference phase, several such examples are input into the model as context. This step aims to show the model a dynamic correction path from "flawed initial translation" to "perfect translation" through a context learning mechanism, guiding the model to implicitly avoid its own common error patterns when generating the target translation.

[0035] In one embodiment, the comparison demonstration context : in, The source language sentence to be translated. This represents the set of auxiliary knowledge extracted. and These represent the initial translation and the actual reference translation of the search example, respectively. This is a preliminary translation of the sentence to be translated. For the k-th sentence retrieved, The index of the retrieved sentence.

[0036] Step S3: Perform the second phase of explicit correction, including: Step S31: Based on the large language model, according to the preset multidimensional quality measurement standards, perform error detection and annotation on the first stage of the revised translation, and generate a structured quality feedback signal. Step S32: Input the source sentence to be translated, the first-stage corrected translation, and the quality feedback signal back into the large language model to generate the second-stage corrected translation.

[0037] In this embodiment, to eliminate fine-grained semantic biases remaining after implicit correction, an explicit quality-oriented feedback mechanism based on Multidimensional Quality Metric (MQM) is introduced. This mechanism prompts the large language model to identify residual errors in the first-stage corrected translation in a structured manner, including potential error types, error ranges, and brief explanations. These identified issues serve as correction cues, guiding the model to improve translation quality in a targeted manner. This explicit feedback and rewriting process is executed iteratively, outputting the final high-quality target translation after three iterations.

[0038] Specifically, a large language model is used as a quality evaluator to scan the intermediate translation according to a predefined MQM standard. The model needs to output structured feedback signals to accurately locate the specific error types and positions in the translation. Subsequently, a contextual input model is constructed, containing "original input - intermediate translation - MQM error annotations". Based on the above structured MQM error annotation results and combined with correction instructions, the large language model uses its internal autoregressive decoding mechanism to rewrite the annotated errors and regenerate a complete and correct translation. This process is iterated multiple times until a preset number of rounds is reached, thereby ensuring the fidelity and fluency of the final translation.

[0039] Furthermore, the MQM evaluation system adopted includes three core dimensions: Accuracy: The focus is on detecting omissions, illusions, or errors in the translation of entities.

[0040] Fluency is assessed by checking whether the target language conforms to grammatical rules.

[0041] Terminology consistency checks whether the translation of specialized terms is consistent across contexts. Errors are categorized into three levels: critical errors, which hinder text comprehension; major errors, which disrupt the flow of the text but whose intended meaning remains understandable; and minor errors, technical mistakes that do not disrupt the flow or impede understanding.

[0042] Through MQM feedback, interpretable and intervention-friendly translation quality optimization can be achieved.

[0043] In one embodiment, to balance the translation quality of the final translation with the computational overhead of the reasoning stage, the explicit correction in the second stage further includes: Using the second-stage revised translation as the current translation, the steps of error detection and annotation, as well as generating the second-stage revised translation, are repeated to form an iterative revision process; The output of the (t+1)th round of the iterative correction process Represented as: in, The generation and processing of a large language model. Indicates the first Explicit feedback signals generated in each round The source language sentence to be translated. This is the second-stage correction translation for the output of round t. Extensive experimental verification shows that the marginal performance improvement of the model gradually converges as the number of iterations increases; therefore, the maximum number of iterations is set to [value missing]. This achieves an optimal balance between performance and efficiency. The sentence output after three rounds of iteration is a high-quality target language translation that integrates hybrid information enhancement and two-stage correction.

[0044] Based on steps S1-S3, this method uses externally retrieved bilingual parallel sentence pairs and their initial model translations as a foundation. It combines the source sentences with auxiliary knowledge such as thematic information and keywords from the retrieval examples to construct a structured comparative example prompt template. This guides the large language model to implicitly capture error correction patterns and update the initial translation, thereby effectively improving the targeting of translation error correction and the ability to guide semantic constraints.

[0045] By introducing a two-stage correction mechanism, the first-stage translation obtained through implicit correction is further input into a quality-oriented feedback module based on multidimensional quality metric (MQM). This enables hierarchical correction modeling of macro-level major translation errors and fine-grained residual errors (such as sufficiency, fluency, and vocabulary selection). Subsequently, a high-quality target translation is generated through multiple iterative rewriting strategies, enhancing the model's translation expression ability and robustness in scenarios with extremely low resource data scarcity and complex professional fields.

[0046] The first-stage implicit correction mechanism in this invention not only retains the contextual advantage of providing external parallel sentence pairs in retrieval-enhanced generation, but also imposes stronger semantic constraints on the specific domain scope and proper nouns of the source sentence by introducing auxiliary knowledge guidance strategies based on topic information and keywords. This effectively suppresses lexical illusion and domain drift phenomena commonly seen in low-resource translation. In the construction of contextual cue templates, addressing the problem that traditional large language models are prone to semantic biases and difficulty in self-correction during single generation, this method effectively stimulates the model's self-correction potential without parameter updates by structurally comparing the model's initial translation in the retrieval examples with high-quality reference translations. Simultaneously, to bridge the contextual differences between external static retrieval data and the current text to be translated, this invention further combines auxiliary knowledge cues with comparative error correction signals, enabling static examples in the retrieval database to be transformed into dynamic correction guidance during context learning. This solves the problem that translation error patterns are difficult for the model to effectively capture and eliminate under a single retrieval generation path.

[0047] Furthermore, in the second stage of explicit correction learning, this invention designs an iterative feedback loop based on Multidimensional Quality Metric (MQM), effectively eliminating the sufficiency, fluency, and fine-grained vocabulary selection flaws that remain after implicit correction. This enhances the model's robustness in discriminating multiple semantic conflicts when dealing with complex and long sentences or dense technical terms (such as medical texts). Through the aforementioned two-stage correction strategy that integrates implicit contrastive context with explicit quality-oriented feedback, this invention can more accurately characterize and simulate the semantic transformation path of a professional translator—"initial translation-contrast-review-modification"—improving the model's ability to output high-fidelity translations even in the absence of large-scale parallel corpora and facing complex linguistic transfer challenges. This significantly improves the shortcomings of existing low-resource machine translation methods in terms of semantic accuracy, terminology consistency, and cross-domain generalization ability.

[0048] In one embodiment, a vector index library supporting multilingual alignment is built before performing step S1.

[0049] In this embodiment, an encoder model pre-trained using multilingual contrastive learning is selected to perform offline vectorization processing on the low-resource corpus. Strict alignment of the source and target language spaces in the vector dimension is required to ensure semantic relevance in cross-language retrieval.

[0050] Figure 2 This paper presents a comparative example of the initial generation error and the final correction result of the proposed method in low-resource machine translation. Currently, conventional benchmark methods for low-resource machine translation using large language models mainly include single-generation paradigms such as zero-shot direct translation and retrieval-enhanced generation (RAG). Zero-shot translation directly utilizes the pre-trained knowledge within the large language model to perform source-to-target language conversion without providing any external parallel corpus examples. For example... Figure 2 As shown in the upper part, when the input contains complex long sentences with state reversals or multiple modifications, zero-shot translation models often suffer from cognitive overload due to the lack of explicit domain knowledge guidance and grammatical constraints of the target language in extremely low-resource scenarios. This results in the output of an initial translation with severe semantic bias, incorrectly translating it into a self-contradictory statement (i.e.,...). Figure 1 (The semantic conflict section marked in red at the top center).

[0051] While some studies have attempted to introduce Retrieval-Enhanced Generation (RAG) to piece together external translation examples to mitigate errors, both zero-shot direct translation and basic RAG essentially belong to the single-generation paradigm. These conventional methods have significant limitations in practical applications: first, they lack a mechanism for dynamically reviewing and adjusting the initially generated translation; second, when faced with complex word order and logical relationships in low-resource languages, single decoding is prone to generating and difficult to self-eliminate severe semantic and logical illusions; and third, the lack of fine-grained error feedback loops limits the model's translation fidelity when handling long-distance dependencies or dense terminology.

[0052] In comparison, such as Figure 2 As shown in the lower half and the middle of the correction process, this invention, by introducing a multi-stage rewriting mechanism based on error analysis, can accurately identify and correct residual errors in the initial translation (such as...). Figure 2 The translation below has been successfully corrected to the logically coherent green highlighted section, thus significantly improving the translation accuracy and robustness of the large language model in complex contexts.

[0053] In one possible embodiment, Figure 3 This paper presents a flowchart illustrating the specific framework of the proposed low-resource language translation method based on a large-scale model that integrates reasoning and retrieval. The method combines retrieval-enhanced generation with a progressive translation correction mechanism, aiming to comprehensively improve the quality of machine translation in low-resource scenarios.

[0054] First, in the initial preparation phase, given a source sentence, the model retrieves a small subset of relevant parallel sentence pairs from an external bilingual corpus. Then, a large language model is used to perform an initial translation of the source sentence with zero samples. Simultaneously, the same large language model is used to translate the source language side of the retrieved parallel examples, generating corresponding model translation results. Furthermore, the model extracts auxiliary knowledge signals, including topic information and keywords, from both the source sentence and the retrieved examples.

[0055] Next, we move to the first stage: implicit correction based on retrieved parallel examples. This stage guides the model to perform contextual correction by constructing structured prompt templates. Specifically, for each retrieved example, the initial translation generated by the model is compared with the corresponding reference translation, and comparative examples are formed by combining extracted topic and keyword knowledge. By demonstrating the differences between the model output and high-quality reference translations under similar contextual conditions, the large language model can implicitly capture error patterns in the examples and apply similar error correction strategies to the initial translation of the current source sentence using contextual learning capabilities, thereby completing the initial translation optimization.

[0056] The second phase then begins: explicit correction guided by a Multidimensional Quality Metric (MQM). While the implicit correction phase effectively corrected major translation errors, subtle issues such as adequacy, fluency, or vocabulary selection may still exist in the translation, which are difficult to resolve through contextual learning alone. Therefore, this phase introduces a quality-oriented explicit correction step, employing MQM-style cues to guide the large language model in analyzing translation quality in a structured manner. The model is required to identify potential error types and error ranges according to criteria and provide brief explanations.

[0057] Finally, these clearly identified specific error feedbacks are used as correction cues to guide the model to further improve the translation results. Since a single correction is often insufficient to resolve all residual issues, this invention iteratively applies this explicit correction process, re-inputting the updated translation into the explicit correction loop, typically for three iterations, to output a final high-quality translation result. Through this two-stage correction architecture, this invention not only utilizes retrieved knowledge and semantic constraints for implicit optimization but also combines iterative explicit self-correction, effectively bridging the gap between external knowledge and the model's internal capabilities, significantly improving the model's robustness and accuracy in handling complex semantics and specialized terminology when dealing with low-resource languages.

[0058] As shown in Table 1, this invention uses three publicly available standard benchmark datasets—FLORES-200, NTREX-128, and TICO-19—as experimental corpora (test sets). These datasets cover multi-domain mixed text, news text, and highly specialized medical text, and are widely used to evaluate machine translation performance from English or Chinese to low-resource languages. These datasets contain varying degrees of domain complexity, are highly representative and challenging, and can be used to comprehensively evaluate the effectiveness, cross-domain adaptability, and robustness of the proposed two-stage correction framework in low-resource machine translation tasks. Table 1 presents the statistical information for the three benchmark datasets used in this experiment.

[0059] Table 1

[0060] The method proposed in this invention underwent systematic comparative experiments on the FLORES-200, NTREX-128, and TICO-19 evaluation datasets, and its performance was compared with several existing mainstream benchmark models. The experimental evaluation employed two reference-translation-based metrics, XCOMET and BLEURT, as measurement standards. These two metrics accurately measure translation quality from two complementary perspectives: semantic sufficiency and overall fluency.

[0061] Table 2

[0062] The experimental results in Table 2 show that the proposed method achieves competitive performance on the FLORES-200 multi-domain dataset, demonstrating that it can effectively improve semantic fidelity and translation robustness in low-resource scenarios.

[0063] Table 3

[0064] The experimental results in Table 3 show that, on the NTREX-128 dataset in the news domain, our algorithm outperforms existing methods in most languages.

[0065] Table 4

[0066] The experimental results in Table 4 show that our method achieves the best XCOMET scores across all four language tasks on the TICO-19 dataset in the medical field, demonstrating its excellent cross-domain robustness.

[0067] Table 5

[0068] The experimental results in Table 5 show that, on the TICO-19 dataset, which faces the dual challenges of dense medical terminology and Chinese source language translation, the method proposed in this invention still demonstrates excellent generalization ability and domain adaptability. In all four low-resource target language translation tasks, this method achieved the highest XCOMET score.

[0069] To explore the functional differences and synergistic effects between the first and second stages, this invention compares the translation performance of performing only initial implicit correction versus fully introducing explicit correction guided by MQM (Multidimensional Quality Metric). The results in Table 6 show that while the contextual demonstration in the first stage effectively corrects major initial translation errors, the introduction of explicit quality feedback in the second stage resulted in sustained performance improvements across all tested low-resource languages.

[0070] Table 6

[0071] To verify the specific contribution of auxiliary knowledge guidance in the implicit correction stage, this invention conducted experiments on the FLORES-200 dataset to examine the output results of the first stage. Table 7 shows that after removing topic- and keyword-based knowledge guidance, the average values ​​of XCOMET and BLEURT decreased across eight low-resource languages, indicating that the additional auxiliary contextual information can effectively guide the large language model to generate more accurate and high-quality translations that are consistent with the target domain.

[0072] Table 7

[0073] Regarding the setting of the number of feedback iterations in the second stage, this invention further analyzes the average marginal performance gain of all eight languages ​​under different number of iterations. Figure 1 This indicates that even a single initial iteration can lead to substantial improvements in translation quality. More importantly, with a moderate extension of the iteration process, the model gains additional stable performance gains through repeated self-checking and error correction, confirming the effectiveness of the iterative feedback mechanism in deep translation optimization. However, after more than three iterations, the model performance gradually stabilizes, and the marginal benefits of continuous iteration begin to decrease significantly. Considering that excessive correction iterations will increase the computational overhead at the system inference level, this invention comprehensively weighs the final translation quality against the actual inference efficiency, ultimately determining three iterations as the optimal balance point, thereby maximizing the overall translation performance while strictly controlling inference costs.

[0074] Example 2 This invention provides a large-model, low-resource language translation system based on the fusion of reasoning and retrieval, comprising: The acquisition module is used to acquire enhanced information associated with the source sentence to be translated. The enhanced information includes at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs. The first implicit correction module, used to perform the first stage of implicit correction, includes: A comparative demonstration context is constructed, comprising at least one comparative example unit, which consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated for the retrieved example source statement based on a large language model; and a standard reference translation corresponding to the retrieved example source statement. The source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context are input into the large language model to generate the first-stage corrected translation; The second implicit correction module is used to perform the second stage of explicit correction, including: Based on the large language model, and according to the preset multidimensional quality measurement standards, the first-stage revised translation is subjected to error detection and annotation, and a structured quality feedback signal is generated. The source sentence to be translated, the first-stage corrected translation, and the quality feedback signal are input again into the large language model to generate the second-stage corrected translation.

[0075] The acquisition module includes: The first acquisition submodule is used to map the source sentence to be translated and the source sentences in the external corpus into high-dimensional vectors based on a multilingual sentence embedding model using a dense retrieval method, and calculate the cosine similarity; the two parallel sentence pairs with the highest similarity are selected to form a dense retrieval subset; The second acquisition submodule is used to calculate the literal similarity between the source sentence to be translated and the source sentences in the external corpus using a sparse retrieval method, select the parallel sentence pair with the highest similarity to form a sparse retrieval subset, and take the union of the dense retrieval subset and the sparse retrieval subset as the retrieved similar parallel sentence pair.

[0076] Example 3 The present invention also provides a computer-readable storage medium. In one embodiment of the computer-readable storage medium according to the present invention, the computer-readable storage medium can be configured to store a program that performs the large-model low-resource language translation method based on reasoning and retrieval fusion described in the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described large-model low-resource language translation method based on reasoning and retrieval fusion. For ease of explanation, only the parts related to the embodiments of the present invention are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. The computer-readable storage medium can be a storage device comprising various electronic devices. Optionally, in the embodiments of the present invention, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0077] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the original technical features, and the technical solutions resulting from these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. A large-scale, low-resource language translation method based on the fusion of reasoning and retrieval, characterized in that, Includes the following steps: Obtain enhanced information associated with the source sentence to be translated, the enhanced information including at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs; The first phase of implicit modification includes: A comparative demonstration context is constructed, comprising at least one comparative example unit, which consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated for the retrieved example source statement based on a large language model; and a standard reference translation corresponding to the retrieved example source statement. The source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context are input into the large language model to generate the first-stage corrected translation; Perform the second phase of explicit modifications, including: Based on the large language model, and according to the preset multidimensional quality measurement standards, the first-stage revised translation is subjected to error detection and annotation, and a structured quality feedback signal is generated. The source sentence to be translated, the first-stage corrected translation, and the quality feedback signal are input again into the large language model to generate the second-stage corrected translation.

2. The method according to claim 1, characterized in that, Obtaining enhanced information associated with the source statement to be translated includes: A hybrid retrieval strategy is employed to retrieve similar parallel sentence pairs from external corpora, including: Using a dense retrieval method, based on a multilingual sentence embedding model, the source sentence to be translated and the source sentences in the external corpus are mapped into high-dimensional vectors respectively, and the cosine similarity is calculated. Select the two most similar parallel sentence pairs to form a dense search subset; Using a sparse retrieval method, the literal similarity between the source sentence to be translated and the source sentences in the external corpus is calculated, and the parallel sentence pair with the highest similarity is selected to form a sparse retrieval subset; The union of the dense retrieval subset and the sparse retrieval subset is used as the retrieved similar parallel sentence pairs.

3. The method according to claim 2, characterized in that, The auxiliary knowledge includes the first topic information, the first keyword information, the second topic information, and the second keyword information; Obtaining enhanced information associated with the source statement to be translated also includes: The source sentence to be translated is semantically parsed using a large language model to extract the first topic information and the first keyword information; A large language model is used to perform semantic analysis on the source sentences of each retrieval example in the retrieved similar parallel sentence pairs, and to extract the second topic information and the second keyword information.

4. The method according to claim 1, characterized in that, The comparison demonstration context : in, The source language sentence to be translated. This represents the set of auxiliary knowledge extracted. and These represent the initial translation and the actual reference translation of the search example, respectively. This is a preliminary translation of the sentence to be translated. For the k-th sentence retrieved, The index of the retrieved sentence.

5. The method according to claim 1, characterized in that, The explicit corrections performed in the second phase also include: Using the second-stage revised translation as the current translation, the steps of error detection and annotation, as well as generating the second-stage revised translation, are repeated to form an iterative revision process; The output of the (t+1)th round of the iterative correction process Represented as: in, The generation and processing of a large language model. Indicates the first Explicit feedback signals generated in each round The source language sentence to be translated. This is the second stage of the translation correction for the output of round t.

6. The method according to claim 2, characterized in that, In the dense retrieval method, the cosine similarity between the source sentence to be translated and the source sentences in the external corpus is calculated based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus. This indicates a feature encoding operation.

7. The method according to claim 2, characterized in that, In the sparse retrieval method, the literal similarity between the source sentence to be translated and the source sentences in the external corpus is calculated based on the following scoring function: in, The source language sentence to be translated. The source sentences are from an external corpus.

8. A large-model, low-resource language translation system based on the fusion of reasoning and retrieval, characterized in that, include: The acquisition module is used to acquire enhanced information associated with the source sentence to be translated. The enhanced information includes at least: similar parallel sentence pairs retrieved from an external corpus, and auxiliary knowledge extracted from the source sentence to be translated and the retrieved parallel sentence pairs. The first implicit correction module, used to perform the first stage of implicit correction, includes: A comparative demonstration context is constructed, comprising at least one comparative example unit, which consists of the following: a retrieved example source statement; the auxiliary knowledge extracted from the retrieved example source statement; an initial translation generated for the retrieved example source statement based on a large language model; and a standard reference translation corresponding to the retrieved example source statement. The source sentence to be translated, the auxiliary knowledge, the initial translation corresponding to the source sentence to be translated, and the comparison demonstration context are input into the large language model to generate the first-stage corrected translation; The second implicit correction module is used to perform the second stage of explicit correction, including: Based on the large language model, and according to the preset multidimensional quality measurement standards, the first-stage revised translation is subjected to error detection and annotation, and a structured quality feedback signal is generated. The source sentence to be translated, the first-stage corrected translation, and the quality feedback signal are input again into the large language model to generate the second-stage corrected translation.

9. The system according to claim 8, characterized in that, The acquisition module includes: The first acquisition submodule is used to map the source sentence to be translated and the source sentences in the external corpus into high-dimensional vectors based on a multilingual sentence embedding model using a dense retrieval method, and calculate the cosine similarity; the two parallel sentence pairs with the highest similarity are selected to form a dense retrieval subset; The second acquisition submodule is used to calculate the literal similarity between the source sentence to be translated and the source sentences in the external corpus using a sparse retrieval method, select the parallel sentence pair with the highest similarity to form a sparse retrieval subset, and take the union of the dense retrieval subset and the sparse retrieval subset as the retrieved similar parallel sentence pair.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images