A large model laotian translation method and system based on agent cooperation and cross-language semantic alignment
By employing agent collaboration and cross-linguistic semantic alignment, and leveraging the semantic and structural guidance of English and Thai, Lao translations are iteratively corrected through multiple rounds, thus resolving the semantic illusion problem in low-resource scenarios and achieving high-quality translation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- KUNMING UNIV OF SCI & TECH
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-05
AI Technical Summary
In low-resource language translation scenarios, existing technologies suffer from semantic illusion, and traditional methods heavily rely on high-quality parallel data and computing resources, leading to inconsistencies in semantic expression and logical errors in the translation results.
We employ an agent collaboration and cross-linguistic semantic alignment approach, using English as a semantic reference and combining it with Thai, which is linguistically similar, to provide lexical and syntactic guidance. Through a multi-agent interaction framework, we conduct multiple rounds of iterative correction of the translation results, including independent translation generation, cross-linguistic semantic alignment and correction, and semantic clarification and enhancement steps.
Without requiring additional data or model fine-tuning, it significantly improves the semantic fidelity and structural accuracy of low-resource translation, reduces semantic drift and logical errors, and improves translation quality.
Smart Images
Figure CN122154714A_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a large-scale Lao language translation method and system based on agent collaboration and cross-linguistic semantic alignment, belonging to the field of natural language processing and machine translation technology. Background Technology
[0002] In recent years, large language models have demonstrated powerful capabilities in general natural language processing tasks, significantly improving the fluency and overall quality of machine translation. However, in translation scenarios involving low-resource languages (such as Lao), the performance of large language models remains severely limited due to the scarcity of language resources and the lack of high-quality parallel corpora. This data scarcity often leads to a severe "semantic illusion" phenomenon in low-resource machine translation. That is, while the model-generated translation may appear grammatically fluent, its meaning differs greatly from the source text, deviating significantly from the true semantics of the original.
[0003] Traditional low-resource machine translation methods typically rely on large-scale, high-quality bilingual parallel corpora for model training, or employ hub languages and transfer learning strategies to alleviate data scarcity. However, when training data is limited or language differences are significant, the model's generalization ability remains constrained, and the consistency and stability of the semantic expression in the translation results are difficult to guarantee. Currently, machine translation optimization methods based on large language models mainly include prompting engineering, instruction fine-tuning, and post-editing strategies. However, these methods face significant challenges in low-resource scenarios: fine-tuning methods heavily rely on massive computational resources and large-scale, high-quality parallel data, making them difficult to implement under resource-scarce conditions; prompting engineering is highly sensitive to the design of prompt words, and a lack of sufficient few-sample examples further exacerbates the instability of model generation; while existing post-editing strategies are often limited to single-language optimization of the target language, and when the initial translation quality generated by the large model is extremely poor and contains logical errors, the effectiveness of single-language optimization alone is extremely limited.
[0004] Therefore, this invention proposes a large-scale Lao translation method and system based on agent collaboration and cross-lingual semantic alignment to address the insufficient optimization capabilities of single-language translation and semantic illusion in low-resource scenarios. This invention breaks through the traditional single-language correction framework, fully utilizing the advantage of the large-scale language model's "English-centric" nature to provide accurate semantic references, and introducing linguistically similar languages to provide guidance on lexical and syntactic structures. Through a cross-lingual multi-agent interaction framework, this invention significantly improves the semantic fidelity and structural accuracy of translation by iteratively correcting the output results of low-resource languages through multi-party feedback, without requiring model fine-tuning or the introduction of additional data. Summary of the Invention
[0005] This invention provides a large-scale Lao translation method and system based on agent collaboration and cross-language semantic alignment, which can be used to solve the semantic illusion problem caused by data scarcity in low-resource scenarios, as well as the problems of existing fine-tuning methods that highly depend on high-quality parallel corpora and computing resources, and the easy generation of logical errors in single-language optimization.
[0006] The technical solution of this invention is: a Lao machine translation method based on agent collaboration and cross-language semantic alignment, the method comprising the following specific steps:
[0007] Step 1: Independent Multilingual Translation Generation: Input the source language text (Chinese) into multiple translation agents based on a large language model, and generate initial translations for the reference language (English), a linguistically similar language (Thai), and the target language (Lao) independently. Calculate the semantic alignment score between the initial translation of the target language and the initial translation of the reference language. If the score meets the preset threshold, output it directly; otherwise, proceed to Step 2.
[0008] Step 2, Cross-language semantic alignment and correction: Using the translation of the reference language as a semantic reference and the translation of linguistically similar languages as lexical and syntactic guidance, modification suggestions for the target language are generated through information interaction between translation agents. The target language translation agent iteratively corrects the target language translation based on the modification suggestions. If the corrected semantic alignment score meets the preset threshold, it is output; otherwise, proceed to Step 3.
[0009] Step 3, Semantic Clarification and Enhancement: An analytical agent is introduced to perform semantic structure analysis on the source language text, extract key semantic fragments, and simplify the overall expression of the source language; the target language translation agent combines the source language text and the semantic analysis results of the analytical agent to generate and output the final target language translation.
[0010] Furthermore, in Step 1, the specific calculation process for generating independent multilingual translations includes:
[0011] Each translation agent independently generates its initial translation based on the large language model, and the formula is expressed as follows:
[0012]
[0013] in, This represents the input Chinese source language text, and 'i' represents the index set of the target language. These represent English, Thai, and Lao respectively; The parameter is Large model generation function; , and These represent the initial translations generated in English, Thai, and Lao, respectively.
[0014] As a preferred embodiment of the present invention, the specific steps of Step 1 are as follows:
[0015] Step 1.1: Assign an English translation agent Thai translation intelligent agent and Lao language translation intelligent agent Each uses a large language model to generate text for the Chinese source language. The initial translation is generated using the following formula:
[0016]
[0017] in, Represents the target language index and , Represents model parameters, and These represent the initial translations of English, Thai, and Lao, respectively.
[0018] Step 1.2: Calculate the initial translation in Lao. Initial translation of English Semantic alignment score between ,like If the preset threshold is met, output directly. If the result is not the final one, proceed to Step 2 for optimization.
[0019] Furthermore, the semantic alignment scores in Step 1 and Step 2 The calculation formula is as follows:
[0020]
[0021] in, Baseline cosine similarity calculated based on LaBSE embedding; It is the length ratio; It is the length smoothing index; For tail similarity; Weights for tail similarity; Minimum threshold; This is a conditional soft boosting factor that compensates for scores when overall similarity is high but tail similarity is low. This is the maximum score.
[0022] Furthermore, the specific steps and formulas for cross-language semantic alignment and correction in Step 2 include the following:
[0023] Step 2.1: In each optimization iteration k, the English translation agent... Thai translation AI According to the Lao translation of the current round Aggregate and generate modification suggestions The formula is:
[0024]
[0025] in, This indicates a suggested function to generate;
[0026] Step 2.2, Lao language translation intelligent agent Based on source language text and generated modification suggestions The formula for updating the translation results is:
[0027]
[0028] Where k is the iteration index. The model parameters representing the Lao translation agent, after at most two iterations, output the optimized translation version. Recalculate its semantic alignment score. If it meets the threshold, accept it; otherwise, proceed to Step 3.
[0029] Furthermore, the specific steps and formulas for semantic clarification and enhancement in Step 3 include the following:
[0030] Step 3.1: Analyze the intelligent agent Perform two core tasks: (1) extract key semantic fragments and generate concise explanations; (2) simplify the source language text. To understand the overall meaning of the sentence and reduce semantic omissions or excessive additions caused by complex structures;
[0031] Step 3.2: The Lao translation agent combines the source language text with the analysis output of the analysis agent. Generate the final translation result. The formula is:
[0032]
[0033] in, The model parameters represent the Lao language translation agent.
[0034] The present invention also provides a large-scale Lao translation system based on agent collaboration and cross-linguistic semantic alignment, the system comprising: a module for executing the large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment.
[0035] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the large-model Lao translation method based on agent collaboration and cross-language semantic alignment.
[0036] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the large-model Lao translation method based on agent collaboration and cross-language semantic alignment.
[0037] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the large-scale Lao translation method based on agent collaboration and cross-language semantic alignment.
[0038] The beneficial effects of this invention are:
[0039] 1. The cross-language multi-agent interaction framework proposed in this invention does not require task-specific model fine-tuning or the introduction of additional large-scale parallel corpus data. It can be flexibly deployed and significantly improve translation quality even under resource-scarce conditions.
[0040] 2. This invention fully utilizes the characteristic of most large language models being "English-centric", using English as a high-precision semantic reference benchmark, while introducing Thai, which has linguistic similarities with Lao, to provide lexical and syntactic guidance. Through the complementary exchange of cross-language information, it effectively reduces semantic drift and logical errors commonly found in monolingual optimization.
[0041] 3. This invention designs a multi-round iterative optimization strategy and a similarity-based conditional triggering mechanism, enabling the system to adaptively intercept and correct low-quality initial translations in multiple dimensions, fundamentally alleviating the "semantic illusion" phenomenon common in low-resource translation and improving the semantic fidelity of the translation.
[0042] 4. This invention significantly improves semantic accuracy while maintaining grammatical fluency. Experiments show that, compared with the baseline model, the Lao translation quality of this invention on the FLORES-101 dataset has achieved a significant leap. Attached Figure Description
[0043] Figure 1 This is the overall flowchart of the present invention;
[0044] Figure 2 This is a block diagram illustrating the principle of the present invention. Detailed Implementation
[0045] Example 1: As Figures 1-2A large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment is proposed, and its specific steps are as follows:
[0046] S1. Data Preparation and Basic Configuration: This embodiment uses the FLORES-101 test set as the evaluation benchmark. This dataset contains 1012 high-quality Chinese-Lao (zh-lo) parallel sentences and is widely used in low-resource machine translation research. This invention adopts a multi-agent architecture (implemented based on the LangGraph framework), assigning four agents with different roles: an English translation agent... Thai translation intelligent agent Lao language translation intelligent agent and analytical agents In this embodiment, all agents use GPT-4.1-mini as the basic large language model, and none of the agents are trained or fine-tuned for any specific task during operation.
[0047] S2, Independent Multilingual Translation Generation: For a given Chinese source sentence The system first indicates , and The system independently generates initial translations for a reference language (English), a linguistically similar language (Thai), and a target language (Lao). The specific generation formula is as follows:
[0048]
[0049] in, They represent English, Thai, and Lao respectively. Represents model parameters; The parameter is Large model generation function; , and These represent the initial translations generated in English, Thai, and Lao, respectively.
[0050] After generation, the system selects English translation. As a semantic evaluation benchmark, the initial translation of Lao was calculated. and Semantic alignment score between Semantic alignment score The calculation employs LaBSE embedding to compute cosine similarity, and combines length normalization and tail similarity control to enhance robustness, addressing common issues in low-resource translation such as length imbalance and tail redundancy. The specific formula is as follows:
[0051]
[0052] in, Baseline cosine similarity calculated based on LaBSE embedding; It is the length ratio;
[0053] In this embodiment, the parameter is initialized as: length smoothing index = 0.5, tail similarity weight =0.95, minimum threshold = 0.75, maximum score =0.95. Furthermore, when the overall similarity exceeds 0.65 but the tail similarity... When it is below 0.8, apply the conditional soft boosting factor. =+0.05. If the calculated If the value is greater than or equal to a preset threshold (set to 0.60 in this embodiment), then output directly. Otherwise, proceed to the next optimization stage.
[0054] S3. Cross-linguistic semantic alignment and correction: In this stage, English translation provides a semantic benchmark, and Thai translation provides lexical and syntactic structure guidance. Optimization is achieved through iterative interactions between agents. .
[0055] In each optimization iteration step k ( (maximum of two iterations) and First, based on the current Chinese source sentence Lao translation Generate specific modification suggestions :
[0056]
[0057] Subsequently, Lao language intelligent agent Incorporate suggestions and update translation results:
[0058]
[0059] An optimized version was obtained after two rounds of iteration. Calculate it again with semantic alignment score If the target is met, output the result; otherwise, trigger the final enhancement phase.
[0060] S4. Semantic Clarification and Reinforcement: For sentences with complex structures or a lot of implicit semantics (which are prone to omissions or incorrect additions in the first two stages), an analytical agent is introduced. assist .
[0061] first, Analyze the source sentence Extract key semantic fragments and generate concise explanations, while simplifying... The overall meaning of .
[0062] Then, Combining the source sentence and Analysis results Generate final translation :
[0063]
[0064] The final output is a high-quality Lao translation that has undergone triple protection.
[0065] The present invention also provides a large-scale Lao translation system based on agent collaboration and cross-linguistic semantic alignment, the system comprising: a module for executing the large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment.
[0066] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the large-model Lao translation method based on agent collaboration and cross-language semantic alignment.
[0067] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the large-model Lao translation method based on agent collaboration and cross-language semantic alignment.
[0068] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the large-scale Lao translation method based on agent collaboration and cross-language semantic alignment.
[0069] Experimental Results and Analysis: To verify the effectiveness of this invention, comparative experiments were conducted on the FLORES-101 dataset with representative baseline models. sp-BLEU, COMET, and chrF++ were used as evaluation metrics. Experimental results show that this invention significantly outperforms all baselines.
[0070] Table 1. Comparison of translation performance on the FLORES-101 dataset (Chinese-Lao)
[0071] Table 2 Case Analysis
[0072] As shown in Table 1, under zero fine-tuning conditions, this invention achieves significant improvements of +8.67 sp-BLEU, +7.92 chrF++, and +0.06 COMET compared to other baseline models. The case analysis in Table 2 reveals that, faced with complex source sentences containing adversative and concessive relationships, the baseline system lacks cross-linguistic support, misinterpreting them as conditional sentences and producing severe "semantic illusions." In contrast, this invention, by integrating semantic cues from English and structural cues from Thai, accurately captures syntactic relationships, avoids logical misunderstandings, and generates high-fidelity translation results.
[0073] To further verify the role of the two optimization steps in the CLAIF framework, this invention conducted ablation experiments. As shown in Table 3, when only the semantic clarification and enhancement steps are retained and the cross-lingual semantic alignment and correction steps are removed, the sp-BLEU index decreases significantly, indicating that this step is crucial for ensuring translation accuracy and detail preservation. When only the cross-lingual semantic alignment and correction steps are retained and the semantic clarification and enhancement steps are removed, the chrF++ index decreases significantly, indicating that this simplification step can improve character-level or sub-word-level structural coherence and consistency. Experimental results show that both optimization steps are indispensable for improving the overall translation effect.
[0074] Table 3 Ablation Experiment
[0075] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.
Claims
1. A large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment, characterized in that, The method includes the following steps: Step 1, Independent Multilingual Translation Generation: Input the source language text into multiple translation agents, and generate initial translations for the reference language, linguistically similar languages, and the target language independently; calculate the semantic alignment score between the initial translation of the target language and the initial translation of the reference language. If the score meets the preset threshold, output it directly; otherwise, proceed to Step 2. Step 2, Cross-language semantic alignment and correction: Using the translation results of the reference language as a semantic reference and the translation results of linguistically similar languages as lexical and syntactic guidance, modification suggestions for the target language are generated through information interaction between translation agents; The target language translation agent iteratively corrects the target language translation based on the modification suggestions. If the corrected semantic alignment score meets the preset threshold, it outputs the result; otherwise, it proceeds to Step 3. Step 3, Semantic Clarification and Enhancement: An analytical agent is introduced to perform semantic structure analysis on the source language text, extract key semantic fragments, and simplify the overall expression of the source language; the target language translation agent combines the source language text and the semantic analysis results of the analytical agent to generate and output the final target language translation.
2. The large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment according to claim 1, characterized in that, In Step 1, the specific calculation process for generating independent multilingual translations includes: Each translation agent independently generates its initial translation based on the large language model, and the formula is expressed as follows: ； in, This represents the input Chinese source language text, and 'i' represents the index set of the target language. These represent English, Thai, and Lao respectively; The parameter is Large model generation function; , and These represent the initial translations generated in English, Thai, and Lao, respectively.
3. The large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment according to claim 1, characterized in that, The semantic alignment scores in Step 1 and Step 2 The calculation formula is as follows: ； in, Baseline cosine similarity calculated based on LaBSE embedding; It is the length ratio; It is the length smoothing index; For tail similarity; Weights for tail similarity; The minimum threshold; This is a conditional soft boosting factor; This is the maximum score.
4. The large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment according to claim 1, characterized in that, The specific steps and formulas for cross-language semantic alignment and correction in Step 2 include the following: Step 2.1: In each optimization iteration k, the English translation agent... Thai translation AI According to the Lao translation of the current round Aggregate and generate modification suggestions The formula is: ； in, Indicates the suggested function; Step 2.2, Lao language translation intelligent agent Based on source language text and generated modification suggestions The formula for updating the translation results is: ； Where k is the iteration index. The model parameters representing the Lao translation agent, after at most two iterations, output the optimized translation version. Recalculate its semantic alignment score. If it meets the threshold, accept it; otherwise, proceed to Step 3.
5. The large-scale Lao translation method based on agent collaboration and cross-linguistic semantic alignment according to claim 1, characterized in that: The specific steps and formulas for semantic clarification and enhancement in Step 3 include the following: Step 3.1: Analyze the intelligent agent Perform two core tasks: (1) extract key semantic fragments and generate concise explanations; (2) simplify the source language text. To understand the overall meaning of the sentence and reduce semantic omissions or excessive additions caused by complex structures; Step 3.2: The Lao translation agent combines the source language text with the analysis output of the analysis agent. Generate the final translation result. The formula is: ； in, The model parameters represent the Lao language translation agent.
6. A large-scale Lao translation system based on agent collaboration and cross-linguistic semantic alignment, characterized in that: The system includes a module for performing the large-model Lao translation method based on agent collaboration and cross-lingual semantic alignment as described in any one of claims 1 to 5.
7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the program, it implements the large-scale Lao translation method based on agent collaboration and cross-language semantic alignment as described in any one of claims 1 to 5.
8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the large-scale Lao language translation method based on agent collaboration and cross-language semantic alignment as described in any one of claims 1 to 5.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the large-scale Lao language translation method based on agent collaboration and cross-language semantic alignment as described in any one of claims 1 to 5.