Doc2FRC: Length-Consistent Document-Level Machine Translation via Fixed-Range Chunking

• Direct Doc2Doc translation remains prone to n-gram repetition and progressive quality degradation as sequence lengths increase.Source: §1 · p. 1
• FRC uses dynamic programming to keep document chunks within a target length interval, aligning training and inference length distributions despite short residual fragments from conventional chunking.Source: §1 · pp. 1–2
• A lightweight dual-boundary matching algorithm aligns chunks in document-level parallel corpora that are not fully sentence-aligned.Source: §1 · p. 2
• GlobVDoc is a manually curated, 10-language test set independent of mainstream DocMT training distributions, while a proposed LTCR variant estimates consistency through reference–hypothesis word alignments.Source: §1 · p. 2
• FRC-based training outperforms the Doc2Doc baseline and achieves an IWSLT2017 d-BLEU improvement exceeding existing DocMT methods.Source: §1 · p. 2

• Ramos et al. (2025) segment documents into {1,2,4} parts and use MRD2D and CAPT, while FRC specializes within a targeted length range.Source: §2 · p. 2
• Existing Doc2Sent methods align input lengths but lack sufficient in-context information, while fixed-part chunk training cannot guarantee training–inference length-distribution alignment.Source: §2 · pp. 2–3
• Training-free DocMT research includes document-derived prompts, multi-turn refinement, specialized agents, translation memory, and discourse knowledge graphs.Source: §2 · p. 2

3.1 Fixed-Range Chunking

• FRC partitions pre-segmented units into chunks within [m, M], using strict optimization first and relaxed optimization with at most K violations if strict packing fails.Units longer than M are isolated before dynamic programming, and the piecewise cost minimizes deviations when boundary violations are unavoidable.Source: §3.1 · p. 3

3.2 Chunk Alignment

• Dual-boundary matching addresses the lack of perfect sentence alignment by matching source-chunk boundary units to proximal target units and using dynamic programming to select an alignment path.The heuristic may cause marginal data loss but is presented as a scalable alternative for large datasets.Source: §3.2 · p. 3

3.3 FRC-based Training Formats

• Five formats—Doc, 0c1t, 1c1t, 2c1t, and FS4—vary document extent and the preceding-chunk context available during training.Source: §3.3 · p. 3, Figure 1 · p. 4
• Four model types include direct Doc2Doc fine-tuning, separate format-specific models, STAIR variants, and a unified mixed-format model.The figure depicts training-data construction for four model configurations from one document.Source: §3.3 · p. 4, Figure 1 · p. 4

• GlobVDoc is a 10-language document-level test set from Global Voices, designed to be independent of mainstream DocMT training sources.Source: §4 · p. 4
• The dataset includes manually selected English-to-nine-language document pairs, with strict sentence alignment and standardized quality and structure guidelines.The documents span interviews, biographies, real-time news reports, and podcasts.Source: §4 · p. 4
• An appendix quality-estimation function evaluates autocorrelation within each test dataset.Source: §4 · p. 4

5.1 Datasets

• Evaluation uses IWSLT2017 and BWB test data plus GlobVDoc; IWSLT2017 covers six en↔de, fr, it, ko, nl, and zh pairs.BWB uses continuous web-novel chapters, concatenating adjacent chapters for zh→en evaluation.Source: §5.1 · p. 5

5.1.1 Experimental Setups

• QWEN2.5-7B-INSTRUCT and TOWER-7B-INSTRUCT are evaluated with separate [256, 512] chunk-pair sets measured using each model's tokenizer.Stanza delineates units and LaBSE cosine similarity supports chunk alignment; both backbones support 32k-token contexts.Source: §5.1.1 · p. 5

5.1.2 Evaluation Methods

• Evaluation includes d-BLEU, ds-BLEU, chunk-level d-COMET, LTCR, and GEMBA-DA.The proposed LTCR variant uses monolingual hypothesis-reference alignment with Needleman–Wunsch rather than cross-lingual source-hypothesis alignment.Source: §5.1.2 · p. 5

5.2 Results on IWSLT2017 and BWB

• Chunking generally beats d2d for both backbones; STAIR-FS4 surpasses D2DFT on IWSLT2017 and is comparable on BWB, while STAIR-1c1t and 2c1t exceed FS4 in d(s)-BLEU on IWSLT2017.D2DFT's d2d decoding has higher ds-BLEU but lower d-BLEU than 0c1t, indicating chunking is more effective for longer documents; STAIR-FS4's training and optimal decoding formats can mismatch, particularly for QWEN2.5.Source: §5.2 · pp. 5–7
• Against Ramos et al., SEP-2c1t reports IWSLT2017 d-BLEU 12.44 (+6.44), 32.24 (+6.24), 11.26 (-2.24), and 56.84 (+9.84), respectively, outperforming in three cases.The order is QWEN en-xx, TOWER en-xx, QWEN xx-en, and TOWER xx-en; SEP achieves the best overall IWSLT2017 results, with d(s)-BLEU and d-COMET parity with large-scale LLMs.Source: Table 1 · p. 6, §5.2 · pp. 6–7

5.3 Benchmark on GlobVDoc

• On GlobVDoc, STAIR-2c1t has the highest overall QWEN2.5 accuracy, but most fine-tuned models outperform GRAFT and DELTA while trailing GPT-4.1-mini.The largest stated deficiencies are for Portuguese, Dutch, and Korean; D2DFT also degrades for English-to-German, Dutch, and Russian translations.Source: Table 2 · p. 7, §5.3 · p. 7
• On GlobVDoc, Continuous Memory-trained D2DFT and STAIR-FS4 have superior LTCR, with STAIR-FS4 matching or surpassing large-scale LLMs.Other models can achieve high accuracy while showing diminished terminology consistency.Source: §5.3 · p. 7

6.1 Extrapolating Position Embedding

• SEP-1c1t performs best against the PE-based baselines on accuracy and n-gram repetition, with the clearest advantage on IWSLT.Figure 2 reports LTCR scores by dataset for different models.Source: §6.1 · p. 8, Figure 2 · p. 8

6.2 Chunking Strategy

• The proposed FRC strategy outperforms conventional FW and FN in almost all settings; on OOD GlobVDoc, all chunking strategies perform similarly, while GEMBA-DA gives FRC an approximately 1.0-point-or-more advantage on IWSLT2017 and BWB.Source: §6.2 · p. 9

6.3 Fixed Range

• Across fixed ranges, en-xx performance differs only marginally, while [256, 512] has a clear advantage for xx-en, underscoring contextual dependence in that direction.Table 5 reports d-BLEU and LTCR for [0, 256], [256, 512], and [512, 768].Source: §6.3 · p. 9, Table 5 · p. 9

6.4 Model Scale

• From 3B to 14B, SEP-1c1t’s d-BLEU and d-COMET both increase strictly across QWEN2.5 model scales.Figure 3 reports performance under the [256, 512] FRC setting.Source: Figure 3 · p. 9, §6.4 · p. 9

6.6 Statistical Significance Analysis

• On IWSLT2017, SEP-2c1t and STAIR-2c1t significantly outperform STAIR-FS4, DOCFT, and the original models across backbones and directions (p ≤10−4); their mutual advantage varies by setting.Tests use one-sided paired document-bootstrap resampling; the SEP-2c1t versus STAIR-2c1t difference is significant only for QWEN2.5-7B on en-xx (p = 0.004).Source: Figure 4 · p. 10, §6.6 · p. 10

6.5 Trade-off: Accuracy, Training Cost and Deployment Efficiency

• SEP-2c1t (LoRA) offers the best effectiveness-efficiency balance, STAIR-2c1t (Full) minimizes deployment complexity, and SEP-2c1t (Full) achieves highest accuracy at greater overhead.LoRA has lower d-BLEU on IWSLT xx-en but competitive d-COMET and surpasses full fine-tuning on GlobVDoc; rank 64 slightly improves in-domain performance but degrades on GlobVDoc.Source: §6.5 · pp. 9–10, Table 6 · p. 10

• FRC uses dynamic programming to unify chunk-length intervals across training and test data, addressing length discrepancies caused by document-length variability.Source: §7 · p. 10
• The systematic comparison of training paradigms finds separate training more stable on in-domain test data and analyzes trade-offs in quality, training efficiency, and deployment cost.Source: §7 · p. 10
• On the paper’s GlobVDoc benchmark, the fine-tuned models outperform QWEN3-30B-A3B-INSTRUCT-based agent methods in out-of-distribution performance.Source: §7 · p. 10

Limitations

• The authors excluded GuoFeng test data after finding overlap with DocBlocks training data and WMT shared-task Test 3.The overlap affects a small portion of the data.Source: Limitations · p. 11
• Because Ramos et al. (2025) did not release trained parameters, the comparison uses relative improvements over corresponding baselines and may not be fully equivalent.Source: Limitations · p. 11
• FRC-versus-FW/FN comparisons may understate differences because intermediate data processing can vary and SEP-2c1t may encourage homogeneous representations.Source: Limitations · p. 11
• Optimizing context selection for FRC remains future work, despite prior research suggesting context selection can influence model precision.Source: Limitations · p. 11
• Full-parameter fine-tuning is computationally costly, so most analyses use one backbone and a specific training setting.The examples given are TOWER under D2DFT and SEP-1c1t.Source: Limitations · p. 11

Ethical Statement

• The authors state that study materials are publicly accessible for research and that GlobVDoc source material carries a Creative Commons license.Source: Ethical Statement · p. 11

A Evaluation Procedures

• d-COMET is computed on aligned source, hypothesis, and reference chunk triplets with 50% overlap, while GEMBA-DA uses GPT-4.1 for document-level assessment.The d-COMET procedure uses three alignment-oriented chunking strategies; GEMBA-DA scores are averaged over the test set.Source: A Evaluation Procedures · pp. 16–17
• d-COMET combines scores from three segmentation methods, and d-BLEU tokenization varies by translation direction and target language.Source: A Evaluation Procedures · pp. 16–17

B Interest Word Selection for LTCR

• LTCR interest words are reference-text named entities that occur at least twice in a document.The retained entity labels include people, organizations, locations, and other listed categories.Source: B Interest Word Selection for LTCR · p. 17

C Details of GlobVDoc

• Document pairs were reviewed by at least two volunteers, with disagreements resolved collectively; multimedia-heavy articles were excluded.The team also cross-verified alignments for languages without native-speaker contributors.Source: C Details of GlobVDoc · p. 17
• GlobVDoc articles were manually sentence-segmented, with related short segments merged to preserve discourse integrity and semantic correspondence prioritized for complex mappings.Source: C Details of GlobVDoc · p. 18
• The appendix lists GlobVDoc’s document and sentence counts, word totals, language pairs, and document metadata.Source: Table 7 · p. 17, C Details of GlobVDoc · p. 18
• GlobVDoc’s scale is comparable to commonly used DocMT test sets; GuoFeng Test 3 is omitted because its references were unavailable.Table 8 reports dataset statistics, including document, sentence, and English-word counts.Source: C Details of GlobVDoc · p. 18, Table 8 · p. 18

D Quality Estimation of Test Data

• The corpus-level d-ACF is a weighted average of the first four lags, using sentence-embedding cosine similarity or sentence-level chrF.A length-penalty weight mitigates effects of document-length variation; the analysis uses English-side sentences and FRC pairs.Source: D Quality Estimation of Test Data · pp. 18–19
• GlobVDoc has the highest cosine-based d-ACF among the compared datasets at both sentence and chunk levels, and a slight advantage under lexical similarity.Figure 5 presents d-ACF results across four datasets.Source: D Quality Estimation of Test Data · p. 19, Figure 5 · p. 19
• GlobVDoc and IWSLT2017 show U-shaped long-range autocorrelation, whereas BWB and GuoFeng gradually rise after an initial decline.For IWSLT2017, the increase after approximately lag k = 290 suggests some documents may strongly correspond at their beginnings and endings.Source: D Quality Estimation of Test Data · p. 19, Figure 6 · p. 19

E.1 Hyperparameters for FRC and Chunk Alignment

• FRC allows up to eight chunks outside the target range during relaxed optimization; the reported setting produced no invalid training or test samples.Dual-Boundary Matching uses λ = 0.5, σ = 0.05, and 256 target-unit candidates per source boundary.Source: §E.1 · pp. 19–20

E.2 Training Data Filtering and Statistics

• Chunk pairs are filtered by source-length threshold and language-pair median length ratio; outlier samples are then used as breakpoints for further document segmentation.Source: §E.2 · p. 20
• LaBSE-based alignment retains more samples than SentenceBLEU-based alignment and performs consistently better after filtering, although SentenceBLEU is not far behind on BWB.The comparison is shown in Figure 7; the authors associate SentenceBLEU’s higher outlier rate, especially for Chinese, with challenging web-novel data.Source: §E.2 · pp. 20–21, Figure 7 · p. 21
• Tokenizer choice yields similar FRC sample attrition under LaBSE similarity but substantially different cross-lingual length ratios, especially for Korean and Chinese.Source: §E.2 · p. 21
• SEP-2c1t, STAIR-2c1t, and UNI-2c1t have more compact, better-aligned train-test input-length distributions than D2DFT and STAIR-FS4.The three SEP-2c1t submodels have particularly sharp, narrowly bounded distributions in Figure 8.Source: §E.2 · p. 21, Figure 8 · p. 22

E.3 Training and Decoding Setups

• The models use supervised fine-tuning with loss computed only on target-side tokens after FRC-based chunk alignment.Source: §E.3 · p. 21

STAIR-2c1t

• At inference, 0c1t translates without context, 1c1t uses the preceding chunk, and 2c1t uses the two preceding chunks after staged initialization.The first chunk starts in 0c1t mode, and the second in 1c1t mode.Source: STAIR-2c1t · p. 23

F.1 Training Process of UNI-2c1t

• UNI-2c1t’s IWSLT2017 test performance gradually rises and plateaus without discernible overfitting, while BWB shows a temporary test dip during the epoch 1.0-to-1.2 format transition.The authors note that generalization may already be limited given the model’s exceptionally high training-data fit.Source: §F.1 · p. 23

F.2 Analysis of Context Sensitivity

• Replacing test context with unrelated content substantially degrades SEP-2c1t and STAIR-2c1t across test sets, but causes only marginal drops for UNI-2c1t.Table 13 reports d-BLEU under original and randomly replaced context, averaged over five runs.Source: §F.2 · p. 24, Table 13 · p. 24
• On sampled training instances, context replacement drops STAIR-2c1t performance by approximately 20%, while UNI-2c1t shows negligible degradation.UNI-2c1t also attains nearly perfect d-BLEU across the 0c1t, 1c1t, and 2c1t decoding formats.Source: §F.2 · p. 24, Table 14 · p. 24

F.3 Training Cost

• SEP sub-models take 113h 51m total to train, versus 119h 35m for UNI-2c1t on the same hardware.The SEP sub-models can also be trained simultaneously in parallel.Source: §F.3 · p. 24

G.1 Translation Quality–Efficiency Trade-off

• Source-MT context improves terminological consistency and marginally improves en-xx d-BLEU, while Source-Only is more efficient and has an xx-en accuracy advantage.Source-MT decoding latency averages approximately 2.5 times that of Source-Only because it requires sequential decoding; Figure 10 encodes runtime inversely in marker size.Source: §G.1 · p. 25, Figure 10 · p. 25
• 1c+1t shows comprehensive LTCR superiority, particularly on BWB, where it approaches STAIR-FS4’s peak and surpasses large-scale LLMs.The paper attributes the stronger terminological-consistency gains to the higher frequency of domain-specific terminology in BWB web novels.Source: §G.1 · p. 25

G.2 Exposure Bias: A Case Study

• 89% (17/19) of non-terminal repetition chunks failed to recover, and repetition chunks frequently triggered cascading deterioration.The analysis defines broken chunks as having log probabilities below −5; some repetition chunks triggered cascades without crossing that threshold.Source: §G.2 · p. 26

Error Propagation and Recovery Across Chunks

• In an En→Ko case, repetition began at chunk 2, chunk 3 collapsed, and chunks 4–5 returned to normal output.The case is labeled Partial Recovery.Source: Error Propagation and Recovery Across… · p. 27, Figure 11 · p. 27

H.1 Length Analysis at Test Time

• STAIR-FS4 and D2DFT d-BLEU rises with larger FRC intervals, but degrades once test lengths exceed a threshold; STAIR-2c1t and SEP-2c1t peak when test FRC matches training at [256, 512].The analyses conclude that performance is bounded by the training length distribution, while matching the training configuration leaves test-length variation with little impact.Source: §H.1 · p. 26, Error Propagation and Recovery Across… · p. 27, Figure 12 · p. 27

H.2 Impact of Unit Selection on FRC

• The study compares line-break, embedding-based semantic, and Meta-chunking units, using the resulting units to form fixed-range chunks.Semantic chunking uses adjacent-sentence mE5-large embeddings and retains the top 20% of connection points; Meta-chunking selects breakpoints by perplexity.Source: §H.2 · p. 27
• Unit selection perturbs IWSLT en-xx more than BWB or GlobVDoc, predominantly harming both STAIR models while slightly improving SEP-2c1t d-BLEU.Across configurations, higher NRR and a greater proportion of problematic document lengths correlate with lower d-BLEU; BWB and GlobVDoc have zero NRR in all three settings.Source: §H.2 · pp. 27–28, Table 17 · p. 28
• SEP-2c1t has the lowest NRR and remains comparatively stable across FRC unit selections, while both STAIR models show decoding volatility.Despite typically exceeding 1k input tokens, SEP-2c1t has relatively low NRR; the paper suggests its single training length distribution may help explain the contrast with STAIR-2c1t, trained on three distributions.Source: §H.2 · p. 28, Table 18 · p. 28
• Meta-chunking yields marginal LTCR improvements of +1.6 for STAIR-FS4 and +0.6 for STAIR-2c1t, with both unit strategies giving STAIR-FS4 notable gains.The paper reports that these gains surpass Gemini-2.5-Pro; Figure 13 presents LTCR results across unit selections and models on BWB zh-en.Source: §H.2 · p. 28, Figure 13 · p. 28

I Results of LoRA Fine-Tuning

• Increasing LoRA rank from r = 16 to r = 64 adds trainable parameters but yields relatively limited gains on in-distribution test sets.Source: I Results of LoRA Fine-Tuning · p. 29

J AI Assistance Usage

• The authors used ChatGPT, Google Gemini, and Claude to support code development and writing refinement.Source: J AI Assistance Usage · p. 29

K Prompts

• Translation prompt templates specify source and target languages and text, with context variants containing one to three source-context segments.Source: K Prompts · p. 29

I Results of LoRA Fine-Tuning

• LoRA fine-tuning substantially improves the original SEP-2c1t model but does not match full-parameter fine-tuning.Table 19 compares the original, full-parameter fine-tuned, and LoRA fine-tuned models; the supplied text gives no metric values.Source: I Results of LoRA Fine-Tuning · p. 28, Table 19 · p. 30

K Prompts

• The scoring prompt asks for a 0–100 translation score against a human reference and requires JSON output containing an integer score.Source: K Prompts · p. 30

L Algorithm

• The dual-boundary matching algorithm aligns source chunks with target units by maximizing the sum of their matching scores, with failure as a possible outcome.Source: L Algorithm · p. 30
• Fixed-range chunking uses a two-stage dynamic program: it first seeks segments within [m, M], then permits up to K violations if needed.The algorithm takes a target length derived from [m, M] and outputs packed segments.Source: L Algorithm · p. 31
• The passages list article titles and translations, rather than experimental findings about LoRA fine-tuning.Examples include “A new aid regime for Gaza: Humanitarian facade, military core” and “43 years in Syria’s prisons for refusing to bomb a city.”Source: L Algorithm · p. 33
• GlobVDoc document metadata includes document ID, source and translated titles, source and target languages, and author and translator names.Source: Table 20 · p. 36