An asymmetric machine translation synthetic data fine-tuning gradient correction method and system
By using a gradient correction method to fine-tune synthetic data in asymmetric machine translation, the problems of missing semantic anchors and insufficient adaptability in low-resource language translation are solved, which improves translation quality and efficiency, simplifies the training process, reduces costs, and promotes the application of low-resource language translation technology.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHENGZHOU UNIV
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies in low-resource language translation suffer from problems such as missing semantic anchors, easy semantic drift of the model, loss of useful gradient components, low training efficiency, and insufficient adaptability to low-resource scenarios due to symmetric gradient processing.
A gradient correction method based on asymmetric machine translation synthetic data is adopted. By constructing a hybrid training dataset, the gradients of real data and synthetic data are calculated separately. The gradient of real data is used as a benchmark for unidirectional judgment to detect gradient conflict states and perform asymmetric gradient correction. Finally, the parameters of the neural machine translation model are updated.
It improved translation quality, enhanced adaptability to low-resource machine translation scenarios, simplified the training process, reduced costs, improved translation quality and efficiency, and promoted the practical application of low-resource language translation technology.
Smart Images
Figure CN122133681A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of neural machine translation technology, and in particular to a method and system for fine-tuning gradient correction of synthetic data in asymmetric machine translation. Background Technology
[0002] The importance of fine-tuning gradient correction for synthetic data in asymmetric machine translation lies in addressing the semantic bias caused by synthetic data in low-resource language translation, ensuring that the translation model achieves a balance between fluency and semantic fidelity.
[0003] In existing technologies, there are two main approaches: one is the PCGrad method, which handles gradient conflicts in multi-task learning by symmetric gradient projection, and the other is the AlterBT method, which reduces the impact of noise by alternating between synthetic and real data for training.
[0004] First, the PCGrad method, due to its symmetric gradient processing, cannot distinguish the essential differences between real and synthetic data, resulting in a lack of semantic anchoring and easily causing semantic drift in the model. At the same time, its indiscriminate projection may inadvertently delete useful gradient components, reducing training efficiency. Second, the AlterBT method is prone to catastrophic forgetting during the fine-tuning stage on real data, losing useful knowledge in the synthetic data. Furthermore, its alternating training mode cannot ensure the effective removal of conflicting knowledge in the synthetic data, affecting the stability of translation performance. Finally, both methods are not well adapted to low-resource scenarios and struggle to maintain semantic fidelity when real data is extremely scarce, limiting their practical application in low-resource language translation.
[0005] Therefore, a method and system for fine-tuning gradient correction of synthetic data in asymmetric machine translation are provided to solve the above problems. Summary of the Invention
[0006] The purpose of this invention is to provide a method and system for fine-tuning gradient correction of synthetic data in asymmetric machine translation, which solves the problems of semantic anchoring loss, semantic drift of the model, loss of useful gradient components, low training efficiency and insufficient adaptability to low-resource scenarios caused by the symmetric multi-task gradient processing design of existing methods when applied to low-resource language neural machine translation.
[0007] To achieve the above objectives, this invention provides a method for fine-tuning gradient correction of asymmetric machine translation synthetic data, comprising the following steps: S1: Construct a hybrid training dataset; S2: Initialize the neural machine translation model and training parameters; S3: Calculate the gradients for the real data and the synthetic data respectively; S4: Perform unidirectional judgment based on the real data gradient to detect gradient conflict states; S5: Perform asymmetric gradient correction based on the gradient conflict state detection results; S6: Perform parameter updates and convergence verification on the neural machine translation model.
[0008] Preferably, step S1 specifically includes the following steps: S11: Collect manually annotated real parallel corpora of low-resource language pairs in the open-source and shared tasks of WMT and IWSLT, and set them as real anchor data; S12: Download monolingual data from NewsCrawl or CCNet, sample 1M target language monolingual sentences, and use the Qwen3-8B model to reverse translate the target language monolingual sentences into English to obtain the initial synthetic corpus; S13: Set COMETKiwi-22 as a quality filter to discard samples with quality scores below the threshold τ and retain high-quality synthetic corpora. The threshold τ is set to 0.82. S14: Construct training datasets based on real parallel corpora and high-quality synthetic corpora respectively to obtain a hybrid training dataset.
[0009] Preferably, in step S11, the number of sentence pairs in a single corpus is less than 1M sentences, the language pair direction of WMT is English (En) to Gujarati (Gu), Kazakh (Kk), Turkish (Tr), and Romanian (Ro), and the language pair direction of IWSLT is English (En) to Dutch (Nl) and Arabic (Ar).
[0010] Preferably, step S2 specifically includes the following steps: S21: Set Qwen3-8B as the neural machine translation model, and fine-tune the neural machine translation model based on LoRA technology, with the LoRA rank set to 256; S22: Apply different length constraints to the real parallel corpora and high-quality synthetic corpora in the mixed training dataset respectively. The sentence length of the real parallel corpora is no more than 192 words, and the sentence length of the high-quality synthetic corpora is no more than 256 words. S23: Initialize the iteration termination condition. Set the iteration termination condition every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and terminate training when the performance on the validation set does not improve for 5 consecutive training cycles.
[0011] Preferably, step S3 specifically includes the following steps: S31: Divide the mixed training dataset into predetermined batches and input them into the neural machine translation model for training in batches; S32: Calculate the gradient g of the real data corresponding to the real parallel data training task in a single training batch. parThe optimization objective for training on real parallel data is semantic fidelity; S33: Calculate the gradient g of the real data corresponding to the high-quality synthetic data training task in a single training batch. syn The optimization objective for high-quality synthetic data training tasks is fluency.
[0012] Preferably, step S4 specifically involves calculating the gradient g of the real data. par and the gradient of the real data g syn The inner product ρ is used to determine the gradient conflict state. If the inner product ρ is greater than 0, it is determined that the gradient directions are consistent and there is no conflict. If the inner product ρ is less than 0, it is determined that there is a gradient conflict.
[0013] Preferably, step S5 specifically includes the following steps: S51: When there is no conflict, calculate the gradient g for the final parameter update. final The final parameter update gradient g final Specifically set as follows: g final =g par +g syn ; S52: In case of conflict, calculate the final parameter update gradient g. final The final parameter update gradient g final Specifically set as follows: g final =g par +g ⊥ syn ; g ⊥ syn =g syn -(g par ·g syn / ||g par || 2 )·g par ; Among them, g ⊥ syn This represents the synthetic data gradient after removing conflicting components from the gradients of the real parallel data.
[0014] Preferably, step S6 specifically includes the following steps: S61: Update gradient g based on final parameters final The model parameters of the neural machine translation model are updated using the backpropagation algorithm; S62: Every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and repeat steps S3-S5 until the iteration termination condition is met to complete the training of the neural machine translation model. S63: The neural machine translation model is decoded using a beam search strategy with a beam width of 4, and evaluated using a test set of translation tasks; S64: Calculate the BLEU score of the neural machine translation model using the sacreBLEU toolkit and score the translation performance of the neural machine translation model using COMET-22.
[0015] A system for fine-tuning gradient correction of synthetic data in asymmetric machine translation includes a real parallel data acquisition module, a synthetic data generation and filtering module, a Qwen3-8B basic model, a gradient calculation module, a gradient conflict detection module, an asymmetric gradient correction module, and a parameter update module.
[0016] Therefore, the present invention employs the above-mentioned method and system for fine-tuning gradient correction of asymmetric machine translation synthetic data, which has the following beneficial effects: (1) This solution solves the semantic drift problem and improves translation quality: By anchoring real parallel data and asymmetric gradient correction, the interference of semantic bias of synthetic data on model training is eliminated from the root, ensuring that the model is always anchored to the correct semantic direction; (2) This solution enhances the adaptability of low-resource machine translation scenarios: It is optimized for low-resource scenarios where real data is scarce. Even if the proportion of synthetic data is as high as 90% or more, it can still stably guarantee the semantic fidelity of the translation, thus overcoming the problem of loss of real data knowledge that may be caused by the PCGrad method and the problem of catastrophic forgetting of valuable knowledge in synthetic data that may be caused by the AlterBT method. (3) This scheme simplifies the training process and lowers the threshold for use: it does not require repeated data switching and iteration to consume too much computing resources like the AlterBT method, nor does it require the use of symmetric logic to adjust the gradient on the real data while adjusting the gradient on the synthetic data, like the PCGrad method. It can be directly adapted to different low-resource language pairs and is easy to operate. (4) This solution improves economic benefits: reduces training time costs and manual parameter tuning costs, while improving translation quality, thereby reducing the development and application costs of low-resource language neural machine translation models by 20%-30%, and promoting the practical application of low-resource language translation technology.
[0017] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0018] Figure 1 This is a flowchart of a method for fine-tuning gradient correction of synthetic data in asymmetric machine translation according to the present invention; Figure 2 This is a geometric schematic diagram of an asymmetric machine translation synthetic data fine-tuning gradient correction method according to the present invention; Figure 3 This is a structural diagram of a system for fine-tuning gradient correction of synthetic data in asymmetric machine translation according to the present invention. Detailed Implementation
[0019] The method of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0020] Unless otherwise defined, the methodological or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0021] The terms "comprising" or "including" as used in this invention mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements. Terms such as "inner," "outer," "upper," and "lower" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this invention, unless otherwise explicitly specified and limited, the term "attached" and similar terms should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0022] Example like Figure 1 As shown, this invention provides a method for fine-tuning gradient correction of asymmetric machine translation synthetic data, comprising the following steps: S1: Construct a hybrid training dataset; Step S1 specifically includes the following steps: S11: Collect manually annotated real parallel corpora of low-resource language pairs in the open-source and shared tasks of WMT and IWSLT, and set them as real anchor data; Table 1: Basic Information on Training Data Sources and Scale
[0023] As shown in Table 1, in step S11, the number of sentence pairs in a single corpus is less than 1M sentences. The language pair direction of WMT is English (En) to Gujarati (Gu), Kazakh (Kk), Turkish (Tr), and Romanian (Ro). The language pair direction of IWSLT is English (En) to Dutch (Nl) and Arabic (Ar).
[0024] S12: Download monolingual data from NewsCrawl or CCNet, sample 1M target language monolingual sentences, and use the Qwen3-8B model to reverse translate the target language monolingual sentences into English to obtain the initial synthetic corpus; S13: Set COMETKiwi-22 as a quality filter to discard samples with quality scores below the threshold τ and retain high-quality synthetic corpora. The threshold τ is set to 0.82. S14: Construct training datasets based on real parallel corpora and high-quality synthetic corpora respectively to obtain a hybrid training dataset.
[0025] S2: Initialize the neural machine translation model and training parameters; Step S2 specifically includes the following steps: S21: Set Qwen3-8B as the neural machine translation model, and fine-tune the neural machine translation model based on LoRA technology, with the LoRA rank set to 256; S22: Apply different length constraints to the real parallel corpora and high-quality synthetic corpora in the mixed training dataset respectively. The sentence length of the real parallel corpora is no more than 192 words, and the sentence length of the high-quality synthetic corpora is no more than 256 words. S23: Initialize the iteration termination condition. Set the iteration termination condition every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and terminate training when the performance on the validation set does not improve for 5 consecutive training cycles.
[0026] S3: Calculate the gradients for the real data and the synthetic data respectively; Step S3 specifically includes the following steps: S31: Divide the mixed training dataset into predetermined batches and input them into the neural machine translation model for training in batches; S32: Calculate the gradient g of the real data corresponding to the real parallel data training task in a single training batch. par The optimization objective for training on real parallel data is semantic fidelity; S33: Calculate the gradient g of the real data corresponding to the high-quality synthetic data training task in a single training batch. syn The optimization objective for high-quality synthetic data training tasks is fluency.
[0027] S4: Unlike bidirectional indiscriminate detection, this embodiment uses the real data gradient as a benchmark for unidirectional judgment to detect gradient conflict states. Step S4 specifically involves calculating the gradient g of the real data. par and the gradient of the real data g synThe inner product ρ is used to determine the gradient conflict state. If the inner product ρ is greater than 0, it is determined that the gradient directions are consistent and there is no conflict. If the inner product ρ is less than 0, it is determined that there is a gradient conflict, which is caused by semantic deviation.
[0028] S5: Perform asymmetric gradient correction based on the gradient conflict state detection results; Step S5 specifically includes the following steps: S51: When there is no conflict, calculate the gradient g for the final parameter update. final The final parameter update gradient g final Specifically set as follows: g final =g par +g syn ; S52: In case of conflict, calculate the final parameter update gradient g. final The final parameter update gradient g final Specifically set as follows: g final =g par +g ⊥ syn ; g ⊥ syn =g syn -(g par ·g syn / ||g par || 2 )·g par ; Among them, g ⊥ syn This represents the synthetic data gradient after removing conflicting components from the gradients of the real parallel data.
[0029] like Figure 2 As shown, this embodiment does not use the bidirectional indiscriminate projection of PCGrad, but instead uses the gradient g of the real data. syn Gradient g projected unidirectionally onto the real data par The orthogonal complement space, the real data gradient g of the truly parallel data. par As a true anchor, the real data gradient g of high-quality synthetic data syn When the direction is consistent with the actual anchor point, it is represented as g. + syn In this embodiment, the two gradients are added to enrich the learning signal. When gradients conflict, it is represented as g. - syn In this embodiment, g - syn Projected onto g par In the orthogonal direction, to remove conflicting components, and only the orthogonal components g are included. ⊥syn With g par Add them together, and the final update gradient g final Ensure that the model optimization process is not affected by the conflict between synthetic data and real data.
[0030] S6: Perform parameter updates and convergence verification on the neural machine translation model.
[0031] Step S6 specifically includes the following steps: S61: Update gradient g based on final parameters final The model parameters of the neural machine translation model are updated using the backpropagation algorithm; S62: Every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and repeat steps S3-S5 until the iteration termination condition is met to complete the training of the neural machine translation model. S63: The neural machine translation model is decoded using a beam search strategy with a beam width of 4, and evaluated using a test set of translation tasks; S64: Calculate the BLEU score of the neural machine translation model using the sacreBLEU toolkit and score the translation performance of the neural machine translation model using COMET-22.
[0032] Table 2: Experimental Results of Model Training
[0033] As shown in Table 2, experimental results verified that the translations achieved good BLEU and COMET-22 scores on the six low-resource language pairs. Specifically, the English-Kazakh translation improved by 8.52 points in BLEU and 15.07 points in COMET compared to the supervised fine-tuning (SFT) baseline; the English-Turkish translation improved by 3% in BLEU compared to the PCGrad method, etc.
[0034] The method in this embodiment exhibits superior robustness compared to standard fine-tuning (SFT) and robust multilingual baselines (e.g., LLaMAX3, Aya-expanse-8B). In scenarios with very little real supervision (e.g., En-Gu and En-Kk), the method in this embodiment demonstrates particularly stable performance.
[0035] On En-Kk, the standard SFT suffered catastrophic degradation, with its BLEU score (2.17) even lower than the zero-sample baseline (4.85), likely due to overfitting to sparse and noisy parallel data. In contrast, the method in this embodiment effectively utilizes the synthetic signals while anchoring them to the real signals, achieving a significant BLEU score of 10.69, which far exceeds CeMAT (8.80) and LLaMAX3 (6.58).
[0036] Similarly, on En-Gu, the method of this invention outperforms SFT by more than 10 BLEU points and outperforms the strong gemma-3 baseline (25.75), which verifies that the gradient correction of this embodiment can effectively prevent such problems when other models are prone to semantic drift.
[0037] It is worth noting that on En-Ro, the method of this embodiment achieved 42.08 BLEU, surpassing SFT (+5.53) and the professional multilingual model Aya-expanse-8B (+4.25). Unlike Hunyuan-MT, which often exhibits high COMET scores but low BLEU scores, the method of this embodiment can typically deliver improvements on both metrics simultaneously and achieve the best results on average performance, indicating that it achieves a good balance between reference semantics and target language structure.
[0038] Table 3: Comparison of experimental results between this embodiment and reverse translation
[0039] As shown in Table 3, the method of this embodiment is compared with the standard SFT and the competitive reverse translation BT (simple synthetic data mixed with real data) baseline. On both tasks, the method of this invention consistently outperforms the reverse translation baseline (improving by 0.46 and 0.44 BLEU on the En-Kk and En-Tr tasks, respectively). The superior performance of the method of this embodiment shows that the asymmetric gradient projection of this embodiment can effectively filter out harmful synthetic updates, thereby achieving more accurate convergence.
[0040] Table 4: Comparison of Gradient Correction Methods for Different Data
[0041] As shown in Table 4, this embodiment compares the English-Turkish (En-Tr) task with other settings (including standard single-source training, multi-stage training strategies, and not using g). par The methods for updating parameters were compared.
[0042] Compared with the AlterBT method, the AlterBT method in this embodiment can achieve a COMET score (88.70) that is slightly closer to that of the method in this embodiment but still lower than that of the method in this invention. This shows that fine-tuning on parallel data can effectively avoid the negative impact of synthetic data. However, the method in this embodiment can bring better performance, especially in terms of the BLEU metric. This may also be due to the stylistic forgetting phenomenon: the model loses the diverse language patterns learned from synthetic data during the fine-tuning process and overfits to the relatively narrow style in the parallel corpus.
[0043] The dual role of parallel data anchors in the PCGrad method (without adding g) par The performance of the "SFT (using only synthetic data)" baseline (28.62 vs. 28.08) is better than that of the "SFT (using only synthetic data)" baseline, demonstrating that removing conflicting gradients by projection can effectively reduce noise.
[0044] However, the method in this embodiment further improves performance compared to the PCGrad method, which shows that parallel data plays a dual role: on the one hand, it corrects biases, and on the other hand, it provides necessary monitoring signals.
[0045] like Figure 3 As shown, a system for fine-tuning gradient correction of synthetic data in asymmetric machine translation includes a real parallel data acquisition module, a synthetic data generation and filtering module, a Qwen3-8B basic model, a gradient calculation module, a gradient conflict detection module, an asymmetric gradient correction module, and a parameter update module.
[0046] Therefore, the present invention employs the above-mentioned method and system for fine-tuning gradient correction of synthetic data in asymmetric machine translation. By constructing an asymmetric gradient correction mechanism anchored to real data, it effectively eliminates semantic bias in synthetic data in low-resource machine translation, significantly improving translation quality and scene adaptability.
[0047] Finally, it should be noted that the above embodiments are only used to illustrate the method of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the method of the present invention, and these modifications or equivalent substitutions should not cause the modified method to deviate from the spirit and scope of the method of the present invention.
Claims
1. A method for fine-tuning gradient correction of synthetic data in asymmetric machine translation, characterized in that, Includes the following steps: S1: Construct a hybrid training dataset; S2: Initialize the neural machine translation model and training parameters; S3: Calculate the gradients for the real data and the synthetic data respectively; S4: Perform unidirectional judgment based on the real data gradient to detect gradient conflict states; S5: Perform asymmetric gradient correction based on the gradient conflict state detection results; S6: Perform parameter updates and convergence verification on the neural machine translation model.
2. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11: Collect manually annotated real parallel corpora of low-resource language pairs in the open-source and shared tasks of WMT and IWSLT, and set them as real anchor data; S12: Download monolingual data from NewsCrawl or CCNet, sample 1M target language monolingual sentences, and use the Qwen3-8B model to reverse translate the target language monolingual sentences into English to obtain the initial synthetic corpus; S13: Set COMETKiwi-22 as a quality filter to discard samples with quality scores below the threshold τ and retain high-quality synthetic corpora. The threshold τ is set to 0.
82. S14: Construct training datasets based on real parallel corpora and high-quality synthetic corpora respectively to obtain a hybrid training dataset.
3. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 2, characterized in that, In step S11, the number of sentence pairs in a single corpus is less than 1M sentences. The language pair direction of WMT is English (En) to Gujarati (Gu), Kazakh (Kk), Turkish (Tr), and Romanian (Ro). The language pair direction of IWSLT is English (En) to Dutch (Nl) and Arabic (Ar).
4. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 3, characterized in that, Step S2 specifically includes the following steps: S21: Set Qwen3-8B as the neural machine translation model, and fine-tune the neural machine translation model based on LoRA technology, with the LoRA rank set to 256; S22: Apply different length constraints to the real parallel corpora and high-quality synthetic corpora in the mixed training dataset respectively. The sentence length of the real parallel corpora is no more than 192 words, and the sentence length of the high-quality synthetic corpora is no more than 256 words. S23: Initialize the iteration termination condition. Set the iteration termination condition every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and terminate training when the performance on the validation set does not improve for 5 consecutive training cycles.
5. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 4, characterized in that, Step S3 specifically includes the following steps: S31: Divide the mixed training dataset into predetermined batches and input them into the neural machine translation model for training in batches; S32: Calculate the gradient g of the real data corresponding to the real parallel data training task in a single training batch. par The optimization objective for training on real parallel data is semantic fidelity; S33: Calculate the gradient g of the real data corresponding to the high-quality synthetic data training task in a single training batch. syn The optimization objective for high-quality synthetic data training tasks is fluency.
6. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 5, characterized in that, Step S4 specifically involves calculating the gradient g of the real data. par and the gradient of the real data g syn The inner product ρ is used to determine the gradient conflict state. If the inner product ρ is greater than 0, it is determined that the gradient directions are consistent and there is no conflict. If the inner product ρ is less than 0, it is determined that there is a gradient conflict.
7. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 6, characterized in that, Step S5 specifically includes the following steps: S51: When there is no conflict, calculate the gradient g for the final parameter update. final The final parameter update gradient g final Specifically set as follows: g final =g par +g syn ; S52: In case of conflict, calculate the final parameter update gradient g. final The final parameter update gradient g final Specifically set as follows: g final =g par +g ⊥ syn ; g ⊥ syn =g syn -(g par ·g syn / ||g par || 2 )·g par ; Among them, g ⊥ syn This represents the synthetic data gradient after removing conflicting components from the gradients of the real parallel data.
8. The method for fine-tuning gradient correction of asymmetric machine translation synthetic data according to claim 7, characterized in that, Step S6 specifically includes the following steps: S61: Update gradient g based on final parameters final The model parameters of the neural machine translation model are updated using the backpropagation algorithm; S62: Every 1500 training steps, save the neural machine translation model with the best performance on the validation set, and repeat steps S3-S5 until the iteration termination condition is met to complete the training of the neural machine translation model. S63: The neural machine translation model is decoded using a beam search strategy with a beam width of 4, and evaluated using a test set of translation tasks; S64: Calculate the BLEU score of the neural machine translation model using the sacreBLEU toolkit and score the translation performance of the neural machine translation model using COMET-22.
9. A system for asymmetric machine translation synthetic data fine-tuning gradient correction method according to any one of claims 1-8, characterized in that, It includes a real parallel data acquisition module, a synthetic data generation and filtering module, a Qwen3-8B basic model, a gradient calculation module, a gradient conflict detection module, an asymmetric gradient correction module, and a parameter update module.