Two Intermediate Translations Are Better Than One: Fine-tuning LLMs for Document-level Translation Refinement

Yichen Dong, Xinglin Lyu, Junhui Li, Daimeng Wei, Min Zhang, Shimin Tao, Hao Yang

Abstract

Recent research has shown that large language models (LLMs) can enhance translation quality through self-refinement. In this paper, we build on this idea by extending the refinement from sentence-level to document-level translation, specifically focusing on document-to-document (Doc2Doc) translation refinement. Since sentence-to-sentence (Sent2Sent) and Doc2Doc translation address different aspects of the translation process, we propose fine-tuning LLMs for translation refinement using two intermediate translations, combining the strengths of both Sent2Sent and Doc2Doc. Additionally, recognizing that the quality of intermediate translations varies, we introduce an enhanced fine-tuning method with quality awareness that assigns lower weights to easier translations and higher weights to more difficult ones, enabling the model to focus on challenging translation cases. Experimental results across ten translation tasks with LLaMA-3-8B-Instruct and Mistral-Nemo-Instruct demonstrate the effectiveness of our approach.

Introduction: The Local-Global Trade-off

Machine translation has traditionally struggled with the trade-off between local precision and global context. Current methodologies generally fall into two categories, each with distinct drawbacks:

  1. Sentence-to-Sentence (Sent2Sent): These systems are excellent at capturing local semantic details but are "context-blind." This often leads to lexical inconsistency—for example, referring to a "treadmill" as a "running machine" in one sentence and a "treadmill" in the next—and poor overall discourse flow.
  2. Document-to-Document (Doc2Doc): These systems consider the entire text, which improves consistency and style. However, they are prone to "under-translation," where the model skips phrases or entire sentences because it loses track of the source during long-form generation.
Figure 1: An example of translation refinement. Sent2Sent captures the text but lacks consistency (using "running machine" then "treadmill"), while Doc2Doc maintains consistency but omits the final phrase "into new territory." The proposed refinement combines the strengths of both.

The goal of this research is to achieve the best of both worlds: the exhaustive coverage of Sent2Sent and the stylistic harmony of Doc2Doc.

Methodology: A Dual-Draft Refinement Framework

The core proposal is a refinement framework that utilizes an LLM as an "expert editor." The process is structured into three main phases:

1. Data Preparation

For a given source document, the researchers generate two intermediate versions: one via a Sent2Sent system and one via a Doc2Doc system. This provides the refinement model with two different "drafts" that likely contain complementary errors.

2. Two-Stage Fine-Tuning

The researchers fine-tune foundation models (specifically LLaMA-3-8B and Mistral-Nemo) using a sophisticated two-step approach:

  • Naive Fine-Tuning: The model learns to produce a high-quality reference translation given the source and the two intermediate drafts.
  • Quality-Aware (QA) Fine-Tuning: This is the "secret sauce" of the method. Recognizing that some translations are harder to refine than others, the authors assign higher loss weights to "hard" examples (where the intermediate translations are already high-quality) and lower weights to "easy" ones. This forces the model to focus on subtle, challenging improvements rather than just fixing obvious errors.
Figure 2: The proposed workflow, including (a) data preparation using COMET for quality weighting, (b) the two-stage fine-tuning process, and (c) the inference phase where two drafts are refined into a final translation.

3. Inference

During the translation task, the source and both intermediate drafts are fed into the fine-tuned LLM. To avoid "position bias"—where the model might favor the first draft it sees—the order of the two drafts is swapped during both training and inference.

Experimental Setup

The researchers utilized the News Commentary v18.1 dataset from WMT24, a high-quality parallel corpus that includes document boundaries. The evaluation covered ten translation directions, including English (En) ↔ German (De), Russian (Ru), Spanish (Es), Chinese (Zh), and French (Fr). Training sets ranged from approximately 7.3K to 9.7K documents per language pair.

Key Results and Analysis

The approach, dubbed "Ours," consistently outperformed all baselines, including direct Doc2Doc translation and refinement from a single source.

  • Metric Improvements: On LLaMA-3-8B-Instruct, the method achieved average improvements of 2.73 points in d-COMET scores over the raw Sent2Sent baseline and 1.80 points over the raw Doc2Doc baseline.
  • Discourse Metrics: The system showed significant gains in Coherence and LTCR (Lexical Translation Consistency), proving it successfully resolved the "running machine vs. treadmill" type of inconsistencies.
  • Error Reduction: Qualitative analysis showed a massive drop in mistranslations and omissions compared to using Doc2Doc or Sent2Sent alone. By cross-referencing two drafts, the LLM effectively acts as a verification mechanism.

Insights and Discussion

  • The "Two-Draft" Advantage: Refining from two distinct types of errors (inconsistency vs. omission) provides the LLM with a much richer signal than refining a single draft.
  • Quality Awareness Matters: Standard fine-tuning often plateaus because the model spends too much time "learning" to fix easy mistakes. By weighting the loss function based on COMET scores, the authors implement a curriculum that prioritizes difficult refinement cases.
  • Model Agnosticism: One of the most impressive findings is that an LLM fine-tuned on its own drafts can effectively refine translations from entirely different systems, such as GPT-4o-mini or the NLLB (No Language Left Behind) model.

Conclusion

This research demonstrates that document-level translation can be significantly improved by leveraging multiple intermediate perspectives. By combining Sent2Sent and Doc2Doc drafts with a quality-aware fine-tuning strategy, LLMs can produce translations that are both locally accurate and globally coherent. Future work will explore extending this framework to more specialized domains and developing more computationally efficient multilingual models.

Summary updated

Abstract

Recent research has shown that large language models (LLMs) can enhance translation quality through self-refinement. In this paper, we build on this idea by extending the refinement from sentence-level to document-level translation, specifically focusing on document-to-document (Doc2Doc) translation refinement. Since sentence-to-sentence (Sent2Sent) and Doc2Doc translation address different aspects of the translation process, we propose fine-tuning LLMs for translation refinement using two intermediate translations, combining the strengths of both Sent2Sent and Doc2Doc. Additionally, recognizing that the quality of intermediate translations varies, we introduce an enhanced fine-tuning method with quality awareness that assigns lower weights to easier translations and higher weights to more difficult ones, enabling the model to focus on challenging translation cases. Experimental results across ten translation tasks with LLaMA-3-8B-Instruct and Mistral-Nemo-Instruct demonstrate the effectiveness of our approach.
BibTeX
@inproceedings{dong2025two,
  author = {Yichen Dong and Xinglin Lyu and Junhui Li and Daimeng Wei and Min Zhang and Shimin Tao and Hao Yang},
  title = {{Two Intermediate Translations Are Better Than One: Fine-tuning LLMs for Document-level Translation Refinement}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure