Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Cross-lingual transfer among five Turkic languages is strongest between close sibling pairs, depends on transfer direction, and shifts when the fixed translation target changes.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 08:39 UTC pith:TWDLPWIV

load-bearing objection Useful, well-run Turkic transfer study with a normalization problem in its headline metric; the qualitative map likely holds, but the quantitative rankings need a raw-score reanalysis. the 3 major comments →

arxiv 2607.29355 v1 pith:TWDLPWIV submitted 2026-07-31 cs.CL cs.AI

Cross-Lingual Transfer for Machine Translation in Turkic Languages

Omer Burak Cinar , Mehmet Mert Dalkilic , Cagri Toraman This is my paper
classification cs.CL cs.AI
keywords cross-lingual transfermachine translationTurkic languagestransfer matrixrecovery ratelow-resource MTscript normalization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper studies cross-lingual transfer in machine translation across five Turkic languages—Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz—by fine-tuning models on one source–target pair and evaluating them on other source languages with the same translation target. It reports that transfer is strongest between closely related pairs, especially Turkish–Azerbaijani and Kazakh–Kyrgyz, that transfer direction matters, and that the same transfer source–target pair behaves differently under different translation targets. It also finds that Latinization of script improves BLEU and chrF in most script-mismatched settings but not uniformly across metrics or directions. The relative ordering of transfer sources is stable across test datasets, continual pretraining, back-translation data, and model architecture, suggesting a systematic family-internal structure rather than a single-set artifact.

Core claim

The paper constructs fixed-target transfer matrices for five Turkic languages, where each cell reports the recovery rate: the score of a model fine-tuned on transfer source i translating into a fixed target t, evaluated on transfer target j, divided by the score of the same-source model (j→t) and multiplied by 100. Averaging over translation targets different from both languages gives a language-pair transfer score. Using this measure, the paper finds that Turkish–Azerbaijani and Kazakh–Kyrgyz show the highest recovery (e.g., 57.19% and 48.99% BLEU), that recovery is directional (az→ky is not ky→az), and that the same pair’s recovery changes with the translation target (e.g., az→ky varies fr

What carries the argument

The fixed-target transfer matrix with recovery rate R_m(i,j,t)=S_m(i,j,t)/S_m(j,j,t)×100, averaged over translation targets t distinct from both the transfer source i and transfer target j. The matrix rows are transfer sources, columns are transfer targets, and the diagonal gives the same-source reference score; the recovery rate is the percentage of that reference retained when the transfer source differs from the transfer target. This normalization is what makes pairwise transfer comparable across languages and metrics, and every matrix-level conclusion in the paper is read from it.

Load-bearing premise

The load-bearing measurement assumption is that the recovery rate of Eq. (2) faithfully measures transfer; because it divides cross-source scores by the same-source diagonal score, small diagonal denominators can inflate apparent transfer beyond 100% (seen in COMET and COMETKiwi), and no confidence intervals or multi-seed variance are reported.

What would settle it

Recompute the transfer matrices with the same data but replacing the diagonal reference with the mean of several fine-tuning seeds (or a model fine-tuned on all five languages), and check whether Turkish–Azerbaijani and Kazakh–Kyrgyz still rank highest; if the top pairs change or recovery cells above 100% reorder the matrix, the central claim collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • For low-resource Turkic machine translation, the best transfer donor is typically a sibling language in the same Turkic subgroup, so building data and models around Oghuz and Kipchak pairs should outperform using a high-resource non-Turkic source like English.
  • Transfer direction cannot be ignored: reversing the transfer source and transfer target can change recovery by more than 15 percentage points, so any transfer result should be reported with direction and fixed translation target specified.
  • Latinizing Cyrillic-script inputs can improve BLEU and chrF by large margins (up to 63–64%) for script-mismatched transfer pairs, but should not be applied blindly because it can hurt ky→kk and because COMET can decrease even when surface metrics rise.
  • The relative ranking of transfer donors is stable enough across datasets, continual pretraining regimes, back-translation data, and model architectures that a single evaluation setting can be used to select donors for other Turkic languages.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The recovery-rate normalization divides by the same-source diagonal score; cells above 100% (e.g., COMETKiwi over 180%) show that some off-diagonal pairs outscore their reference, so the strongest-pair ranking could be an artifact of low diagonal denominators rather than actual transfer superiority.
  • The target-dependence result implies that cross-lingual transfer is not a static property of a language pair; donor selection for a new translation direction should be conditioned on the output language, which the paper demonstrates but leaves as a design implication.
  • The strong raw-cell correlation between encoder–decoder and decoder-only models may reflect shared training data or evaluation distribution rather than linguistic structure; testing on independently human-translated data would separate those sources of stability.
  • A concrete extension would be to re-run the recovery analysis using a multi-seed mean or a joint fine-tuned reference on the diagonal; if the Turkish–Azerbaijani and Kazakh–Kyrgyz blocks survive that renormalization, the subgroup-transfer claim is far more secure.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper studies cross-lingual transfer among five Turkic languages (tr, az, uz, kk, ky) in a fixed-translation-target MT setup. For each translation target t, models are fine-tuned on each source i→t and evaluated zero-shot on all other sources j→t. Transfer is quantified by the recovery rate R_m(i,j,t)=S_m(i,j,t)/S_m(j,j,t)×100 (Eq. 2), averaged over translation targets different from both i and j (Eq. 3). The main findings are: transfer is strongest for the closely related pairs tr–az and kk–ky; transfer direction matters; the same transfer source–target pair behaves differently under different translation targets; Latinization improves BLEU/chrF mainly in script-mismatched directions but not uniformly; and donor preferences are stable across XWMT/Tatoeba, with/without CPT, and between mT5 and Qwen3 (raw-cell correlations r≈0.95). The Limitations section appropriately recognizes that the fine-tuning data is largely synthetic and scopes the claims to relative transfer structure under a fixed pipeline.

Significance. If the quantitative claims survive reanalysis, this would be a useful fixed-target transfer matrix for an understudied language family, with a reproducible pipeline (scripts, matrices, and model releases). The paper's clear strengths are the detailed appendix tables, the cross-dataset and cross-architecture stability analyses based on raw donor–evaluation matrices, and the high raw-cell correlations (Table 6: mT5–Qwen3 BLEU r=0.953; Table 5: allCPT–noCPT BLEU r=0.967). However, the central quantitative instrument—the recovery-rate normalization in Eq. (2)—confounds transfer strength with same-source diagonal scores, and no uncertainty quantification is provided. Because the headline pair rankings and direction-asymmetry magnitudes are read directly from this normalized quantity, the quantitative claims as currently stated are not fully supported.

major comments (3)
  1. [Eq. (2), Table 1, Appendix A.3.1] The recovery rate R = S(i,j,t)/S(j,j,t)×100 divides by the same-source diagonal score, which varies widely across languages (raw BLEU diagonals 12.86–13.43 for tr/az vs 3.44–6.70 for kk/ky; Appendix A.3.1). A fixed raw transfer score therefore produces much higher recovery when the denominator is low. Concretely, Table 1 reports kk→ky (48.99%) as much stronger than ky→kk (39.32%), but multiplying the Appendix A.3.1 cells by their diagonals gives nearly equal raw BLEU averages (~2.16 vs ~2.06). The >100% COMET/COMETKiwi cells (uz→az COMET 103.9%; ky→kk COMETKiwi 180.3%) illustrate the same inflation. Since the main RQ1/RQ2 conclusions (strongest pairs, direction-asymmetry magnitudes, target-conditioning) are read from this normalized quantity, a reanalysis on raw scores—or an alternative normalization that does not depend on diagonal magnitude—is required before these quantitative claims
  2. [§A.2.1, Tables 1, 3] All experiments use a single seed (seed 42; Table 10) and no confidence intervals are reported. Many conclusions are ordering claims over small margins (e.g., Table 1 Avg. BLEU: tr 39.36, az 39.10, uz 39.33) and recovery-rate differences of a few points (e.g., Figure 1). With 400 test samples per pair and one fine-tuning run, these differences may not be statistically reliable. Please report multi-seed variance or bootstrap/paired significance tests for the main recovery and raw-score comparisons, particularly for the direction-asymmetry and strongest-pair claims.
  3. [Eq. (3), Table 1 Avg. column] The language-level summary T_m(i,j) averages recovery rates over translation targets t∈L\{i,j}, and the English row is averaged over t∈L\{j} (four targets) rather than three. As a result, the Avg. column mixes incomparable quantities: recovery denominators vary across t, and different target sets are used for Turkic vs English rows. The claim that Kazakh is the strongest transfer source on average (BLEU 42.22 vs ~39 for tr/az/uz) may be an artifact of these choices. Please report raw-score averages with identical target sets, and make the row definitions directly comparable.
minor comments (5)
  1. [Table 1 caption] The caption should disclose that the English row averages recovery over four translation targets while Turkic rows average over three; the main text mentions this, but the caption itself does not.
  2. [Section 5.3, Table 2] Please define whether the reported percentages are computed relative to original-script raw scores or recovery rates. Appendix A.3.3 reports raw deltas; the main text should state the formula used for Table 2 and reconcile the two presentations.
  3. [Eq. (2)] State explicitly that recovery is unbounded and can exceed 100%. This occurs for COMET and COMETKiwi in Appendix A.3.1 and will surprise readers; a one-sentence clarification would help.
  4. [Tables 5–6] Clarify exactly which cells enter each off-diagonal correlation (all i≠j cells across all targets? per-target subsets?) and why the 'best donor' counts in Table 5 differ from the 20-column procedure used in Table 3.
  5. [Typos] Minor editorial issues: 'transfer sourceis' in Section 1; 'V olume' appears in references [14], [19], [31], [40], and [46]; 'Suzhuo' in reference [27]. Appendix A.1.2 mentions 'Gemini 3.1 Pro'; please verify the model name/version.

Circularity Check

0 steps flagged

No circular derivation found; recovery-rate normalization is a metric-validity issue, not a construction.

full rationale

The paper's central claims rest on measured translation scores on external XWMT/Tatoeba benchmarks, not on parameters fitted to those claims. Equation (2) defines recovery rate as R_m(i,j,t)=S_m(i,j,t)/S_m(j,j,t), and Equation (3) averages it over translation targets; both are explicit, self-defined metrics computed from independent raw scores. The diagonal denominator is a measured same-source score, not a fitted constant, and the recovery values are not fed back to predict the same data. The appendix's >100% COMET/COMETKiwi cells and the wide spread of BLEU diagonals (tr/az ~12.9-13.4 vs kk/ky ~3.4-6.7) show that low denominators can inflate apparent transfer, but this is a statistical identifiability and robustness concern, not a circular reduction: the strongest-pair and directionality claims are not equivalent by construction to the input data. The stability analyses compare raw matrices (e.g., Table 6 reports r=0.953 on all raw BLEU cells), which is consistent with the claims being empirical rather than definitional. The self-citations [49,50] are methodological (CPT recipe; recovery-rate reference) and are not load-bearing: no uniqueness theorem, ansatz, or fitted parameter is imported from them to force the results. The Limitations section appropriately confines the conclusions to relative transfer structure under a fixed pipeline. No derivation step reduces to its own input, so no specific circular step is identified.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

No fitted constants in the mathematical sense; the central quantities are measured BLEU/chrF/COMET scores. The listed hand-chosen settings (corpus size, seed, test sample size) are load-bearing because the transfer matrices are measured under them and no sensitivity analysis covers them. No new theoretical entities are introduced.

free parameters (3)
  • Fine-tuning corpus size per direction = 100K pairs
    Appendix A.1.2: FT dataset completed to 100K pairs per language using back-translation, following Eronen et al. [13]. Transfer matrices are measured under this fixed size; no scaling analysis.
  • Random seed = 42
    Tables 9–11 set seed 42 for all CPT/FT/eval runs; no multi-seed variance is reported, so any seed sensitivity of recovery matrices is unmeasured.
  • Test sample size = 400 per pair
    Appendix A.1.2 and Table 10: MAX_TEST_SAMPLES=400 for XWMT and Tatoeba subsets; donor rankings are computed on these small samples without confidence intervals.
axioms (3)
  • domain assumption Synthetic back-translated and mined parallel pairs preserve the relative cross-lingual transfer structure of Turkic languages.
    Section 3 and Limitations: most fine-tuning pairs are synthetic; the authors explicitly limit claims to relative structure under a fixed pipeline, but the entire matrix is built from these pairs.
  • domain assumption Recovery rate (Eq. 2), normalized by the same-source diagonal score, is a valid and interpretable measure of transfer strength.
    Central method; Appendix A.3.1 shows >100% COMET/COMETKiwi recoveries, indicating the metric is not bounded and can be inflated by small denominators.
  • domain assumption Single-seed, 400-sample-per-pair evaluation is sufficient to establish donor ranking stability.
    Appendix A.1.2 and Table 10; no confidence intervals or multi-seed variance are reported, yet rankings are declared stable.

pith-pipeline@v1.3.0-daily-deepseek · 29008 in / 14017 out tokens · 125813 ms · 2026-08-03T08:39:13.530956+00:00 · methodology

0 comments
read the original abstract

Cross-lingual transfer is central to low-resource machine translation, but its behavior within closely related language families remains insufficiently characterized. We study transfer among five Turkic languages; Turkish, Azerbaijani, Uzbek, Kazakh, and Kyrgyz; using pairwise transfer matrices. In this setting, each model is fine-tuned with one transfer source and evaluated on a different transfer target while the translation target remains the same. Across mT5 experiments, we find that transfer is strongest between closely related Turkic pairs, especially Turkish-Azerbaijani and Kazakh-Kyrgyz. We also show that transfer direction matters, and that the same transfer source-transfer target pair can behave differently when the translation target changes. Latinization improves BLEU and chrF in several script-mismatched settings, but its effect is not uniform across metrics. Additional analyses show that transfer sources are mostly stable across different datasets and model settings.

Figures

Figures reproduced from arXiv: 2607.29355 by Cagri Toraman, Mehmet Mert Dalkilic, Omer Burak Cinar.

Figure 1
Figure 1. Figure 1: Paired BLEU recovery-rate comparison for mT5 Latin. Each y-axis label shows a pair of possible transfer source–transfer [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of Dataset Filtering Metrics In this work, we used pairs from publicly available datasets for CPT in transfer-coefficient sampling experiments, and we combined back-translated data with book-aligned pairs to create the FT data for transfer-matrix calculations. We used all aligned book pairs and completed the FT dataset size to 100K pairs for each language using back-translation [12], following t… view at source ↗
Figure 3
Figure 3. Figure 3: Full Latinized allCPT XWMT recovery matrices. Rows of panels correspond to fixed translation targets and metric [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Extended appendix version of the paired BLEU recovery-rate comparison for mT5 Latin. Each y-axis label shows two [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Latinized-minus-original allCPT XWMT raw-score deltas. Rows of panels correspond to fixed translation targets and [PITH_FULL_IMAGE:figures/full_fig_p025_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: allCPT-minus-noCPT raw-score deltas on Latinized XWMT. Rows of panels correspond to fixed translation targets and [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qwen3 0.6B recovery matrices. Rows of panels correspond to fixed translation targets and metric columns correspond to [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

62 extracted references · 6 linked inside Pith

  1. [1]

    Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoît Sagot. 2022. Towards a cleaner document-oriented multilingual crawled corpus. InProceedings of the Thirteenth Language Resources and Evaluation Conference, pages 4344–4355, Marseille, France. European Language Resources Association

    2022

  2. [2]

    Costa-jussà

    Belen Alastruey, João Maria Janeiro, Alexandre Allauzen, Maha Elbayad, Loïc Barrault, and Marta R. Costa-jussà. 2025. Interference matrix: Quantifying cross-lingual interference in trans- former encoders.arXiv preprint arXiv:2508.02256

    Pith/arXiv arXiv 2025

  3. [3]

    Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence embeddings for zero- shot cross-lingual transfer and beyond. InTransactions of the Association for Computational Linguistics

    2019

  4. [4]

    Mikko Aulamo, Nikolay Bogoychev, Shaoxiong Ji, Graeme Nail, Gema Ramírez-Sánchez, Jörg Tiedemann, Jelmer van der Linde, and Jaume Zaragoza. 2023. HPLT: High performance lan- guage technologies. InProceedings of the 24th Annual Conference of the European Association for Machine Translation, pages 517–518, Tampere, Finland. European Association for Machine ...

    2023

  5. [5]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural machine translation by jointly learning to align and translate. InProceedings of the 3rd International Conference on Learning Representations

    2015

  6. [6]

    Smith, and Luke Zettlemoyer

    Terra Blevins, Tomasz Limisiewicz, Suchin Gururangan, Margaret Li, Hila Gonen, Noah A. Smith, and Luke Zettlemoyer. 2024. Breaking the curse of multilinguality with cross-lingual expert language models. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 10822–10837, Miami, Florida, USA. Association for Computat...

    2024

  7. [7]

    Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. Enriching word vectors with subword information.Transactions of the Association for Computational Linguistics, 5:135–146

    2017

  8. [8]

    Chang, Catherine Arnett, Zhuowen Tu, and Benjamin K

    Tyler A. Chang, Catherine Arnett, Zhuowen Tu, and Benjamin K. Bergen. 2024. When is multilin- guality a curse? language modeling for 250 high- and low-resource languages. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4074–4096, Miami, Florida, USA. Association for Computational Linguistics

    2024

  9. [9]

    Everlyn Asiko Chimoto and Bruce A. Bassett. 2022. Very low resource sentence alignment: Luhya and Swahili. InProceedings of the Fifth Workshop on Technologies for Machine Translation of Low-Resource Languages (LoResMT 2022), pages 1–8, Gyeongju, Republic of Korea. Association for Computational Linguistics

    2022

  10. [10]

    Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. Unsupervised cross-lingual representation learning at scale. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 8440–8451, Online. Asso...

    2020

  11. [11]

    Costa-jussà, James Cross, Onur Çelebi, Francisco Guzmán, and 1 others

    Marta R. Costa-jussà, James Cross, Onur Çelebi, Francisco Guzmán, and 1 others. 2022. No lan- guage left behind: Scaling human-centered machine translation.arXiv preprint arXiv:2207.04672. 12 Cross-Lingual Transfer for MT in Turkic Languages Cinar et al

    Pith/arXiv arXiv 2022

  12. [12]

    Ona de Gibert, Joseph Attieh, Teemu Vahtola, Mikko Aulamo, Zihao Li, Raúl Vázquez, Tiancheng Hu, and Jörg Tiedemann. 2025. Scaling low-resource MT via synthetic data generation with LLMs. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 27674–27692, Suzhou, China. Association for Computational Linguistics

    2025

  13. [13]

    Juuso Eronen, Michal Ptaszynski, Karol Nowakowski, and Zheng Lin Chia. 2023. Improving Polish to English neural machine translation with transfer learning: Effects of data volume and language similarity. InProceedings of the 1st International Workshop on Multilingual, Multimodal and Multitask Language Generation, pages 40–46, Tampere, Finland. European As...

    2023

  14. [14]

    Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, and 1 others. 2021. Beyond english-centric multilingual machine translation.Journal of Machine Learning Research, 22(107):1– 48

    2021

  15. [15]

    Christian Federmann, Tom Kocmi, and Ying Xin. 2022. NTREX-128 – news test references for MT evaluation of 128 languages. InProceedings of the First Workshop on Scaling Up Multilingual Evaluation, pages 21–24, Online. Association for Computational Linguistics

    2022

  16. [16]

    Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. Language- agnostic BERT sentence embedding. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 878–891, Dublin, Ireland. Associa- tion for Computational Linguistics

    2022

  17. [17]

    Dirk Goldhahn, Thomas Eckart, and Uwe Quasthoff. 2012. Building large monolingual dictionaries at the Leipzig corpora collection: From 100 to 200 languages. InProceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 759–765, Istanbul, Turkey. European Language Resources Association (ELRA)

    2012

  18. [18]

    Isai Gordeev, Sergey Kuldin, and David Dale. 2024. FLORES+ translation and machine translation evaluation for the Erzya language. InProceedings of the Ninth Conference on Machine Translation, pages 614–623, Miami, Florida, USA. Association for Computational Linguistics

    2024

  19. [19]

    Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean

    Melvin Johnson, Mike Schuster, Quoc V . Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean

  20. [20]

    Yunsu Kim, Petre Petrov, Pavel Petrushkov, Shahram Khadivi, and Hermann Ney. 2019. Pivot-based transfer learning for neural machine translation between non-English languages. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pag...

    2019

  21. [21]

    Tom Kocmi. 2020. Exploring benefits of transfer learning in neural machine translation.arXiv preprint arXiv:2001.01622

    Pith/arXiv arXiv 2020

  22. [22]

    Tom Kocmi and Ondˇrej Bojar. 2018. Trivial transfer learning for low-resource neural machine translation. InProceedings of the Third Conference on Machine Translation: Research Papers, pages 244–252, Brussels, Belgium. Association for Computational Linguistics. 13 Cross-Lingual Transfer for MT in Turkic Languages Cinar et al

    2018

  23. [23]

    Tom Kocmi and Christian Federmann. 2023. GEMBA-MQM: Detecting translation quality error spans with GPT-4. InProceedings of the Eighth Conference on Machine Translation, pages 768–775, Singapore. Association for Computational Linguistics

    2023

  24. [24]

    Tom Kocmi and Christian Federmann. 2023. Large language models are state-of-the-art evaluators of translation quality. InProceedings of the 24th Annual Conference of the European Association for Machine Translation, pages 193–203, Tampere, Finland. European Association for Machine Translation

    2023

  25. [25]

    Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. In Proceedings of the First Workshop on Neural Machine Translation, pages 28–39, Vancouver. Association for Computational Linguistics

    2017

  26. [26]

    Sneha Kudugunta and 1 others. 2023. Madlad-400: A multilingual and document-level large audited dataset.arXiv preprint arXiv:2309.04662

    Pith/arXiv arXiv 2023

  27. [27]

    Danni Liu and Jan Niehues. 2025. Conditions for catastrophic forgetting in multilingual translation. InProceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), pages 347–359, Suzhuo, China. Association for Computational Linguistics

    2025

  28. [28]

    Shayne Longpre, Sneha Kudugunta, Niklas Muennighoff, I-Hung Hsu, Isaac Caswell, Alex Pentland, Sercan Arik, Chen-Yu Lee, and Sayna Ebrahimi. 2025. Atlas: Adaptive transfer scaling laws for multilingual pretraining, finetuning, and decoding the curse of multilinguality.arXiv preprint arXiv:2510.22037

    arXiv 2025

  29. [29]

    Kaiwen Lu, Yating Yang, Fengyi Yang, Rui Dong, Bo Ma, Aihetamujiang Aihemaiti, Abibilla Atawulla, Lei Wang, and Xi Zhou. 2025. Low-resource language expansion and translation capacity enhancement for LLM: A study on the Uyghur. InProceedings of the 31st International Conference on Computational Linguistics, pages 8360–8373, Abu Dhabi, UAE. Association for...

    2025

  30. [30]

    Yingfeng Luo, Ziqiang Xu, Yuxuan Ouyang, Murun Yang, Dingyang Lin, Kaiyan Chang, Tong Zheng, Bei Li, Peinan Feng, Quan Du, Tong Xiao, and Jingbo Zhu. 2025. Beyond english: Toward inclusive and scalable multilingual machine translation with llms.arXiv preprint arXiv:2511.07003

    Pith/arXiv arXiv 2025

  31. [31]

    Yan Meng and Christof Monz. 2024. Disentangling the roles of target-side transfer and regularization in multilingual machine translation. InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (V olume 1: Long Papers), pages 1828–1840, St. Julian’s, Malta. Association for Computational Linguistics

    2024

  32. [32]

    Jamshidbek Mirzakhalov, Anoop Babu, Duygu Ataman, Sherzod Kariev, Francis Tyers, Otabek Ab- duraufov, Mammad Hajili, Sardana Ivanova, Abror Khaytbaev, Antonio Laverghetta Jr., Bekhzodbek Moydinboyev, Esra Onal, Shaxnoza Pulatova, Ahsan Wahab, Orhan Firat, and Sriram Chellappan

  33. [33]

    Nguyen and David Chiang

    Toan Q. Nguyen and David Chiang. 2017. Transfer learning across low-resource, related languages for neural machine translation. InProceedings of the Eighth International Joint Conference on Natural Language Processing (V olume 2: Short Papers), pages 296–301, Taipei, Taiwan. Asian Federation of Natural Language Processing. 14 Cross-Lingual Transfer for MT...

    2017

  34. [34]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics

    2002

  35. [35]

    Maja Popovi´c. 2015. chrF: character n-gram F-score for automatic MT evaluation. InProceedings of the Tenth Workshop on Statistical Machine Translation, pages 392–395, Lisbon, Portugal. Association for Computational Linguistics

    2015

  36. [36]

    Sukannya Purkayastha, Sebastian Ruder, Jonas Pfeiffer, Iryna Gurevych, and Ivan Vuli ´c. 2023. Romanization-based large-scale adaptation of multilingual language models. InFindings of the As- sociation for Computational Linguistics: EMNLP 2023, pages 7996–8005, Singapore. Association for Computational Linguistics

    2023

  37. [37]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of Machine Learning Research, 21(140):5485–5551

    2020

  38. [38]

    Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. 2020. COMET: A neural framework for MT evaluation. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2685–2702, Online. Association for Computational Linguistics

    2020

  39. [39]

    Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G

    Ricardo Rei, Marcos Treviso, Nuno M. Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, José G. C. de Souza, Taisiya Glushkova, Duarte Alves, Luisa Coheur, Alon Lavie, and André F. T. Martins. 2022. CometKiwi: IST-unbabel 2022 submission for the quality estimation shared task. InProceedings of the Seventh Conference on Machine Translation (WMT),...

    2022

  40. [40]

    Lütfi Kerem Senel, Benedikt Ebing, Konul Baghirova, Hinrich Schuetze, and Goran Glavaš. 2024. Karde¸ s-NLU: Transfer to low-resource languages with the help of a high-resource cousin – a benchmark and evaluation for Turkic languages. InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics (V olume 1: L...

    2024

  41. [41]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 86–96, Berlin, Germany. Association for Computational Linguistics

    2016

  42. [42]

    Govind Soni and Pushpak Bhattacharyya. 2024. RoMantra: Optimizing neural machine translation for low-resource languages through Romanization. InProceedings of the 21st International Conference on Natural Language Processing (ICON), pages 157–168, AU-KBC Research Centre, Chennai, India. NLP Association of India (NLPAI)

    2024

  43. [43]

    David Stap, Vlad Niculae, and Christof Monz. 2023. Viewing knowledge transfer in multilingual machine translation through a representational lens. InFindings of the Association for Computa- tional Linguistics: EMNLP 2023, pages 14973–14987, Singapore. Association for Computational Linguistics

    2023

  44. [44]

    Simeng Sun, Angela Fan, James Cross, Vishrav Chaudhary, Chau Tran, Philipp Koehn, and Francisco Guzmán. 2022. Alternative input signals ease transfer in multilingual machine translation. InProceedings of the 60th Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 5291–5305, Dublin, Ireland. Association for Comp...

    2022

  45. [45]

    Ilya Sutskever, Oriol Vinyals, and Quoc V . Le. 2014. Sequence to sequence learning with neural networks. InProceedings of the 28th International Conference on Neural Information Processing Systems, pages 3104–3112

    2014

  46. [46]

    Maali Tars, Andre Tättar, and Mark Fišel. 2021. Extremely low-resource machine translation for closely related languages. InProceedings of the 23rd Nordic Conference on Computational Lin- guistics (NoDaLiDa), pages 41–52, Reykjavik, Iceland (Online). Linköping University Electronic Press, Sweden

    2021

  47. [47]

    Jörg Tiedemann. 2012. Parallel data, tools and interfaces in OPUS. InProceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), pages 2214–2218, Istanbul, Turkey. European Language Resources Association (ELRA)

    2012

  48. [48]

    Jörg Tiedemann. 2020. The tatoeba translation challenge – realistic data sets for low resource and multilingual MT. InProceedings of the Fifth Conference on Machine Translation, pages 1174–1182, Online. Association for Computational Linguistics

    2020

  49. [49]

    Cagri Toraman. 2024. Adapting open-source generative large language models for low-resource languages: A case study for Turkish. InProceedings of the F ourth Workshop on Multilingual Repre- sentation Learning (MRL 2024), pages 30–44, Miami, Florida, USA. Association for Computational Linguistics

    2024

  50. [50]

    Cagri Toraman, Furkan ¸ Sahinuç, and Eyup Yilmaz. 2022. Large-scale hate speech detection with cross-domain transfer. InProceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2215–2225, Marseille, France. European Language Resources Association

    2022

  51. [51]

    Iulia Turc, Kenton Lee, Jacob Eisenstein, Ming-Wei Chang, and Kristina Toutanova. 2021. Revisit- ing the primacy of English in zero-shot cross-lingual transfer.Preprint, arXiv:2106.16171

    Pith/arXiv arXiv 2021

  52. [52]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InProceedings of the 31st International Conference on Neural Information Processing Systems, pages 6000–6010

    2017

  53. [53]

    Qian Wang and Jiajun Zhang. 2022. Addressing asymmetry in multilingual neural machine translation with fuzzy task clustering. InProceedings of the 29th International Conference on Com- putational Linguistics, pages 5129–5141, Gyeongju, Republic of Korea. International Committee on Computational Linguistics

    2022

  54. [54]

    Xinyi Wang and Graham Neubig. 2019. Target conditioned sampling: Optimizing data selection for multilingual neural machine translation. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5823–5828, Florence, Italy. Association for Computational Linguistics

    2019

  55. [55]

    Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mT5: A massively multilingual pre-trained text-to-text transformer. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 483–498, Onlin...

    2021

  56. [56]

    Bilge Ka˘gan Yazar and Erdal Kiliç. 2025. Improving low-resource Kazakh-English and Turkish- English neural machine translation using transfer learning and part of speech tags.IEEE Access, 13:32341–32356. 16 Cross-Lingual Transfer for MT in Turkic Languages Cinar et al

    2025

  57. [57]

    Rustem Yeshpanov, Alina Polonskaya, and Huseyin Atakan Varol. 2024. KazParC: Kazakh parallel corpus for machine translation. InProceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9633–9644, Torino, Italia. ELRA and ICCL

    2024

  58. [58]

    Brian Yu, Hansen Lillemark, and Kurt Keutzer. 2023. Simple and effective input reformulations for translation. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10322–10334, Singapore. Association for Computational Linguistics

    2023

  59. [59]

    Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low- resource neural machine translation. InProceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1568–1575, Austin, Texas. Association for Computational Linguistics. A Appendix A.1 Dataset Creation Details A.1.1 Monolingual Corpus...

    2016

  60. [62]

    no meaning preserved

    including corpora such as OpenSubtitles, TED2020, and CCAligned. All parallel data underwent multi-stage filtering, including language detection via FastText [7], script normalization, removal of noisy or misaligned pairs, length ratio constraints, and deduplication. Since the collected data were highly imbalanced and contained few direct pairs between lo...

    1999

  61. [2017]

    Transactions of the Association for Computational Linguistics, 5:339–351

    Google’s multilingual neural machine translation system: Enabling zero-shot translation. Transactions of the Association for Computational Linguistics, 5:339–351

  62. [2021]

    InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5876–5890, Online and Punta Cana, Dominican Republic

    A large-scale study of machine translation in Turkic languages. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5876–5890, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics

    2021