PaperlayerView exact revision on arXiv

Source-linked AI summary

STAR : Sentence Translation Alignment Rate for Document-to-Document Machine Translation

Yichen Dong, Hao Wang, Junhui Li, Linlong Xu, Longyue Wang, Weihua Luo

arXiv:2608.27161v1cs.CL

TL;DR

Doc2Doc translation often suffers from sentence-level structural misalignment, especially omissions and hallucinations, despite its potential for broader context. The paper introduces STAR to measure structural fidelity and StarPO to optimize document hypotheses with dynamic alignment masking. Experiments report improved translation quality and structural integrity, with compact models surpassing massive proprietary systems while using tokens more efficiently.

  • Problem

    Doc2Doc translation frequently violates sentence-level source-target correspondence through omissions, hallucinations, and other structural mismatches.

  • Method

    STAR measures document-level sentence alignment, while StarPO ranks hypotheses by structural quality and dynamically masks aligned sentences to optimize misaligned segments.

  • Results

    0.48 COMET improvement over standard CPO on LLaMA-3.1 was reported on average across News-Commentary results, while StarPO also outperformed GPT-4o and other strong systems.

  • Takeaways & Limitations

    StarPO provides a document-level translation approach that improves translation quality and token efficiency without requiring complex agentic workflows.

  • Takeaways & Limitations

    Validation currently focuses on compact 4B–9B models and high-to-medium resource languages, leaving scalability to larger architectures and low-resource languages open.

Abstract

from arXiv · show

Large Language Models (LLMs) have enabled a shift from sentence-level to document-to-document (Doc2Doc) machine translation, promising improved global coherence. However, document-to-document generation in a single pass frequently suffers from structural misalignment, manifesting as sentence omissions or hallucinations that violate the core requirement of source-target correspondence. To address this, we introduce Sentence Translation Alignment Rate (STAR), an auxiliary metric that explicitly quantifies sentence-level structural fidelity. Building on this, we propose STAR-masked Preference Optimization (StarPO), a framework that ranks document-level hypotheses by structural quality and utilizes a dynamic alignment mask to focus optimization on misaligned segments. Experimental results across news and literary domains demonstrate that StarPO significantly enhances translation quality and structural integrity. Notably, StarPO allows compact models to surpass the performance of massive proprietary systems like GPT-4o while maintaining superior token efficiency.

1 Introduction

Doc2Doc translation offers broader context but often breaks sentence-level source-target correspondence through omissions, hallucinations, and other structural mismatches. The paper introduces STAR and StarPO to measure and optimize structural fidelity while preserving the benefits of document-level generation.

  • Motivation: Doc2Doc translation can translate an entire source document in one pass, but often produces sentence omissions or hallucinations that violate source-target correspondence.These failures are especially pronounced in long or complex documents.
  • Motivation: Sent2Sent and Chunk2Chunk methods enforce local correspondence but limit global text planning and repeatedly encode overlapping source contexts.Their localized structure also incurs substantial computational costs.
  • Structural fidelity: Non-1-to-1 mappings can reflect legitimate merging, splitting, or reordering, so robust systems must distinguish fluent restructuring from omissions and hallucinations.Correcting these structural nuances can yield further performance gains.
  • Structural fidelity: Structural mismatches persist across model scales and even under explicit sentence-boundary constraints, while alignment rates gradually drop as input context length increases.GPT-4o also exhibits non-trivial structural mismatch in Doc2Doc and Chunk2Chunk settings.
  • Proposed approach: STAR measures sentence-level structural fidelity, and StarPO ranks document hypotheses by alignment quality while dynamically focusing optimization on misaligned segments.The mask downweights or excludes already aligned sentences and targets omissions and hallucinations.
  • Results: The paper reports consistent and robust improvements in document-level translation quality across multiple domains and models.Its stated contributions include mitigating sentence omissions and hallucinations through STAR-masked preference optimization.

2 Methodology

The framework quantifies sentence-level structural fidelity with STAR, uses STAR-ranked preference pairs, and masks preference optimization to concentrate learning on structurally problematic sentences.

  • Sentence Translation Alignment Rate (STAR): STAR measures sentence-level structural fidelity by computing the fraction of clean 1-to-1 alignment units among all alignment units.Documents are segmented, sentence-aligned, categorized into 1-to-1, deletion, insertion, and complex units, then scored using these categories.
  • Sentence Translation Alignment Rate (STAR): STARrelax treats complex alignments as positive, penalizing omissions and hallucinations while tolerating linguistically justified sentence merging or splitting.This variant distinguishes pathological structural errors from legitimate cross-lingual restructuring.
  • Preference Data Generation: Preference data pairs the highest-STAR candidate as chosen with the lowest-STAR candidate as rejected, retaining pairs only when their score difference exceeds τ = 0.1.For each source document, GPT-4o generates five candidates at temperature 1.0, with a reference translation added when available.
  • STAR-Masked Preference Optimization (StarPO): The training procedure first performs supervised fine-tuning on high-quality parallel corpora, then applies preference optimization to align the model with structural constraints.The preference stage uses a warmstarted policy and the STAR-masked preference loss.
  • STAR-Masked Preference Optimization (StarPO): StarPO replaces standard document-level likelihood with a sentence-level masked likelihood that excludes well-aligned 1-to-1 sentences from preference optimization.The mask assigns 0 to well-aligned sentences and 1 to sentences associated with structural mismatches, focusing learning on the latter.

3 Experimentation

Experiments evaluate StarPO on news and web-novel translation with multiple compact backbones and competitive systems. Across these settings, StarPO consistently improves translation quality over preference-optimization baselines while preserving valid complex mappings.

  • News-Commentary results: StarPO consistently outperforms standard CPO and achieves the best average performance across models on News-Commentary.On LLaMA-3.1, StarPO improves average COMET by 0.48 over standard CPO, with gains stable across model scales.
  • News-Commentary results: StarPO enables a compact model to outperform Tower-plus-9B, GPT-4o, and DeepSeek-R1 in the reported document-level translation comparisons.The comparison is reported as stable across model scales.
  • Guofeng results: On Guofeng, StarPO preserves evaluation of valid complex mappings while preventing structural pathologies during training.The reported Zh⇒De example restores coherence where LLaMA-3.1-base fails, with a cited 52.31 COMET score.
  • Optimization baselines: STAR significantly outperforms COMET, COMETKiwi, BLEU, and word-level alignment coverage for constructing preference data under an equalized data budget.The comparison supports using discourse-level structural fidelity rather than sentence-level alternatives for ranking preference pairs.
  • Ablations: Alignment-aware masking consistently surpasses random token- or sentence-level masking, while relaxed STAR and SFT are weaker alternatives in the reported ablations.The results attribute the stronger learning signal to focusing optimization on structurally misaligned segments.

4 Discussions

The discussion examines STAR’s alignment validity, alternatives, and fine-grained effects on semantic quality. Results show that structural alignment assessment benefits from precise sentence correspondence and that StarPO improves difficult restructuring cases.

  • STAR evaluation: STAR is evaluated against LLM-annotated alignment scores on WMT24 translations spanning multiple models and English-German, Chinese, Russian, and Spanish directions.Spearman correlation is computed under strict and relaxed matching settings.
  • STAR evaluation: STAR achieves superior correlation with structural noise compared with Align-then-Slide and SEGALE, highlighting the value of fine-grained sentence alignment.The discussion presents precise sentence alignment as a prerequisite for reliable document-level structural assessment.
  • Alternative metrics: Length-based metrics show weak correlations, with ρ < 0.2, and fail to detect structural noise when overall document length is preserved.The compared metrics include Token Count Ratio, Sentence Count Ratio, and Sentence Count Difference.
  • Robustness: Replacing STAR’s sentence tokenizer or LaBSE encoder causes only marginal performance fluctuations, indicating robustness to these component choices.The ablation substitutes Spacy for SaT and M3 for LaBSE.
  • LLM-as-a-judge evaluation: StarPO performs best across the reported fluency, content-error, and coherence-error dimensions and model families in the LLM-as-a-judge evaluation.It outperforms standard CPO, Tower+, and GPT-4o in the News-Commentary Zh⇒En setting.

5 Related Work

Related work distinguishes context-aware sentence-oriented translation from holistic Doc2Doc translation and situates StarPO among alignment evaluation and preference-optimization methods.

  • Document-level translation: LLM-based document translation generally follows Doc2Sent, which treats documents as sentence-level tasks, or Doc2Doc, which processes documents holistically.These paradigms differ in whether translation is organized around individual sentences or the whole document.
  • Doc2Sent approaches: Doc2Sent methods use context selection, self-refinement, memory-based agents, and fine-tuning data construction strategies.The cited approaches include both training-free prompting and supervised fine-tuning methods.
  • Doc2Doc approaches: Doc2Doc methods pursue holistic translation through long-context training, iterative or agentic refinement, and input optimization such as segmentation or knowledge fusion.Recent evaluation work predominantly relies on source-target sentence alignment for document-quality assessment.
  • Preference optimization: Reinforcement-learning methods for machine translation include online reward-based approaches and offline preference methods such as DPO and its variants.Online approaches commonly use quality-estimation models as rewards, whereas offline methods rely on pre-curated preference datasets.

6 Conclusion

The paper introduces STAR and StarPO to address sentence-level structural misalignment in Doc2Doc translation. Experiments report improved translation quality and token efficiency, including compact models surpassing massive proprietary systems.

  • Conclusion: STAR evaluates document-level structural fidelity, while StarPO uses dynamic masking to target omissions and hallucinations during preference optimization.The framework is presented as a method for addressing structural misalignment in Doc2Doc translation.
  • Conclusion: Experiments report that StarPO enables compact models to surpass massive proprietary systems in translation quality while improving token efficiency.The conclusion frames this as a robust Doc2Doc paradigm without complex agentic workflows.

Limitations

The paper identifies scope and deployment limitations alongside structural trade-offs in document-level translation. Its validation remains concentrated in compact models and higher-resource languages, while data construction still depends on proprietary APIs.

  • 1-to-1 alignment can mitigate hallucinations but may introduce structural rigidity that discourages valid complex mappings in stylized texts.The reported empirical impact of this rigidity is minimal.
  • Validation is concentrated on compact 4B–9B models and high-to-medium-resource languages, leaving scalability to 70B+ architectures and low-resource languages untested.The paper identifies these settings as a critical direction for future research.
  • The data-construction pipeline relies on proprietary models such as GPT-4o for candidate augmentation, preventing a fully end-to-end open-source workflow.The authors plan to replace this generation step with open-source alternatives for offline deployment.
  • Valid merging, splitting, and swapping can improve cohesion or readability but may still contain localized semantic inaccuracies.The paper therefore distinguishes stylistic restructuring from information loss during evaluation.
  • STAR alignment quality declines as input context length increases, motivating explicit alignment objectives for long-document translation.On Zh ⇒En News-Commentary with LLaMA-3-8B-Instruct, STAR falls from 100% in Sent2Sent to 96.69% at 256 tokens.
  • The margin threshold τ is relatively insensitive over 0.00–0.15, with performance variation below 0.6 points; τ = 0.10 is used by default.The default was selected because pilot studies showed a marginal but consistent advantage.

F Data Statistics

The experiments use News-Commentary and Guofeng bilingual parallel corpora spanning multiple language pairs, with document-level statistics reported for data splits, preference data, and document length.

  • The News-Commentary corpus covers Chinese–English, Chinese–German, Russian–English, German–English, and Spanish–English directions.
  • The Guofeng corpus primarily covers Chinese–Russian, Chinese–English, and Chinese–German language pairs.
  • Reported statistics include document counts across training, validation, and test splits, preference-optimization set size, average document tokens, and maximum document tokens.These statistics are presented in Tables 10 and 11.

G Implementation Details

The implementation fine-tunes models with LoRA using a lightweight training setup and evaluates STAR with a reported throughput of approximately 3 samples per second.

  • LoRA fine-tuning uses rank 8, alpha 16, one training epoch, AdamW, learning rate 1 × 10^-4, warmup ratio 0.05, and β = 0.1.
  • Training runs on one NVIDIA H100 GPU and requires approximately 1 hour.
  • STAR evaluation achieves a throughput of approximately 3 samples per second.

H STAR Scores

The paper evaluates STAR scores and their relationship to translation quality using automated and LLM-based assessment procedures. STAR is defined through sentence alignment, while broader quality comparisons report StarPO against multiple baselines.

  • Table 12 reports strict and relaxed STAR scores on the News-Commentary test set.The top table value is strict, while the bottom gray value is relaxed.
  • The LLM-based STAR judge splits source and target documents into sentences, aligns them, and calculates the Sentence Translation Alignment Rate.The automated metric is validated against this LLM implementation.
  • Figure 6 standardizes the document-level translation prompt across experiments, using <tgt_lang> for the target language and <src_doc> for the source document.The standardization is intended to eliminate prompt-engineering variance.
  • Gemini-2.5-Flash evaluates News-Commentary Zh ⇒En alignment results using the STAR-judging prompt.

J Details in Preference Optimization

The section analyzes STAR-masked preference optimization, showing how masking changes preference margins and preserves useful gradients while improving structural translation quality. It also compares STAR-based optimization with alternative reward signals and preference-learning methods.

  • Practical Advantages: STAR provides a more efficient and robust reward alternative to COMET-based training, calculating rewards at approximately 3 samples/s while supporting stable convergence.COMET-based GRPO/GSPO training is computationally expensive and unstable, while COMETKiwi may produce unintended target languages.
  • Score Distributions: Figure 7 compares STAR (Original) and STAR (Relaxed) score distributions to assess how structural constraints affect preference-pair construction.The top plot includes perfect matches, while the bottom excludes them for finer comparison.
  • Score Distributions: STAR (Original) maintains dense coverage among high-quality non-perfect candidates, whereas STAR (Relaxed) has sparse coverage near the perfect region.The relaxed setting concentrates many samples at score 1.0, making the non-perfect region less informative.
  • Masked Objective: The STAR-masked objective computes preference likelihood over selected sentences, focusing optimization on structural issues such as hallucinations, omissions, and complex alignments.The mask M(tj) determines whether each target sentence contributes to preference loss.
  • Margin Scaling: Masking reduces the accumulated preference margin relative to the full-document objective when the retained token ratio ρ is below 1.The masked margin is treated as a scaled version of the full margin when preference information is distributed across the document.
  • Initial Loss Magnitude: Masking increases the early CPO loss because the model perceives a smaller distance between preferred and dis-preferred candidates.The derivation attributes this behavior to the monotonic decrease of f(z) = −log σ(z) with the margin.
  • Gradient Flow: By masking easy-to-align regions, StarPO avoids gradient saturation and keeps optimization focused on complex, structurally critical segments.Reducing ΔSTAR keeps σ(βΔSTAR) farther from 1, preserving a significant gradient magnitude during training.
Loading 2608.27161v1…