pith. sign in

arxiv: 1907.05019 · v1 · pith:7OLACY42new · submitted 2019-07-11 · 💻 cs.CL · cs.LG

Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges

Naveen Arivazhagan , Ankur Bapna , Orhan Firat , Dmitry Lepikhin , Melvin Johnson , Maxim Krikun , Mia Xu Chen , Yuan Cao
show 5 more authors
George Foster Colin Cherry Wolfgang Macherey Zhifeng Chen Yonghui Wu
This is my paper

Pith reviewed 2026-05-24 23:22 UTC · model grok-4.3

classification 💻 cs.CL cs.LG
keywords multilingual neural machine translationuniversal translationtransfer learninglow-resource languagesmassively multilingual NMTjoint training103 languages
0
0 comments X

The pith

A single neural model translates between any pair among 103 languages while matching bilingual quality on high-resource pairs and improving it on low-resource ones.

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

The paper constructs and evaluates one shared neural machine translation system trained jointly on data from 103 languages using more than 25 billion sentence pairs. The system uses a common architecture and training procedure to enable transfer of knowledge from high-resource to low-resource language pairs. This produces measurable quality gains for low-resource directions without measurable loss on high-resource directions relative to separate bilingual models. The authors also examine multiple design choices that affect overall quality and practicality, and they document specific remaining problems. Readers should care because the work tests whether one model can replace many separate systems for broad language coverage.

Core claim

We introduce our efforts towards building a universal neural machine translation system capable of translating between any language pair. We set a milestone towards this goal by building a single massively multilingual NMT model handling 103 languages trained on over 25 billion examples. Our system demonstrates effective transfer learning ability, significantly improving translation quality of low-resource languages, while keeping high-resource language translation quality on-par with competitive bilingual baselines. We provide in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT.

What carries the argument

A single shared neural network trained jointly across all 103 language pairs to support transfer learning.

If this is right

  • Low-resource language pairs receive quality gains from shared parameters and data.
  • High-resource language pairs retain quality levels comparable to dedicated bilingual models.
  • One model suffices for translation coverage across 103 languages instead of separate models per pair.
  • Design choices such as data sampling and model capacity directly affect whether transfer succeeds.
  • Practical universal systems require further work on the issues identified in the analysis.

Where Pith is reading between the lines

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

  • Extending the same joint-training approach to additional languages would require explicit mechanisms to limit interference.
  • The observed transfer benefits could apply to other sequence tasks that benefit from multilingual data sharing.
  • Replacing many bilingual models with one multilingual model would reduce total parameter count and inference overhead in deployment.
  • Language pairs that are typologically distant may still exhibit hidden interference that only appears at larger scale.

Load-bearing premise

A single shared model architecture and training procedure can balance performance across diverse language pairs without significant negative transfer or interference between languages.

What would settle it

After training, measure that translation quality on at least one high-resource language pair falls measurably below the corresponding bilingual baseline.

Figures

Figures reproduced from arXiv: 1907.05019 by Ankur Bapna, Colin Cherry, Dmitry Lepikhin, George Foster, Maxim Krikun, Melvin Johnson, Mia Xu Chen, Naveen Arivazhagan, Orhan Firat, Wolfgang Macherey, Yonghui Wu, Yuan Cao, Zhifeng Chen.

Figure 1
Figure 1. Figure 1: Per language pair data distribution of the training dataset used for our multilingual ex￾periments. The x-axis indicates the language pair index, and the y-axis depicts the number of train￾ing examples available per language pair on a log￾arithmic scale. Dataset sizes range from 35k for the lowest resource language pairs to 2 billion for the largest. language pair. (ii) European parliamentary doc￾uments (K… view at source ↗
Figure 2
Figure 2. Figure 2: Quality (measured by BLEU) of in￾dividual bilingual models on all 204 supervised language pairs, measured in terms of BLEU (y￾axes). Languages are arranged in decreasing order of available training data from left to right on the x-axes (pair ids not shown for clarity). Top plot reports BLEU scores for translating from English to any of the other 102 languages. Bottom plot reports BLEU scores for translatin… view at source ↗
Figure 3
Figure 3. Figure 3: Effect of sampling strategy on the per￾formance of multilingual models. From left to right, languages are arranged in decreasing order of available training data. While the multilingual models are trained to translate both directions, Any→En and En→Any, performance for each of these directions is depicted in separate plots to highlight differences. Results are reported rela￾tive to those of the bilingual b… view at source ↗
Figure 4
Figure 4. Figure 4: Temperature based data sampling strate￾gies overlaid on the data distribution. We repeat the experiment in Section 4.1 with temperature based sampling, setting T = 5 for 7 [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of varying the sampling temper￾ature on the performance of multilingual models. From left to right, languages are arranged in de￾creasing order of available training data. Results are reported relative to those of the bilingual base￾lines (2). Performance on individual language pairs is reported using dots and a trailing aver￾age is used to show the trend. The colors cor￾respond to the following sam… view at source ↗
Figure 6
Figure 6. Figure 6: Effect of increasing the number of lan￾guages on the translation performance of multi￾lingual models. From left to right, languages are arranged in decreasing order of available training data. Results are reported relative to those of the bilingual baselines (2). The colors correspond to the following groupings of languages: (i) Blue: 10 languages ↔ En, (ii) Red: 25 languages ↔ En, (iii) Yellow: 50 languag… view at source ↗
Figure 7
Figure 7. Figure 7: Results comparing the performance of models trained to translate English to and from all languages to two separate from and to English models. From left to right, languages are arranged in decreasing order of available training data. Re￾sults are reported relative to those of the bilingual baselines (2). The colors correspond to the fol￾lowing models: (i) Green: dedicated (individual) En→Any model for top … view at source ↗
Figure 8
Figure 8. Figure 8: Average number of sentence-piece to [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Effect of increasing capacity on the per [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
read the original abstract

We introduce our efforts towards building a universal neural machine translation (NMT) system capable of translating between any language pair. We set a milestone towards this goal by building a single massively multilingual NMT model handling 103 languages trained on over 25 billion examples. Our system demonstrates effective transfer learning ability, significantly improving translation quality of low-resource languages, while keeping high-resource language translation quality on-par with competitive bilingual baselines. We provide in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT. While we prototype a high-quality universal translation system, our extensive empirical analysis exposes issues that need to be further addressed, and we suggest directions for future research.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The manuscript reports the construction of a single shared NMT model covering 103 languages and trained on more than 25 billion sentence pairs. It claims that the model achieves effective positive transfer to low-resource pairs while remaining competitive with strong bilingual baselines on high-resource pairs, and supplies an extensive empirical analysis of training choices, practical trade-offs, and remaining open challenges rather than asserting a fully solved universal system.

Significance. If the reported empirical outcomes hold under the described conditions, the work constitutes a substantial scaling milestone for multilingual NMT. The explicit cataloguing of practical issues (language balancing, negative transfer risks, inference efficiency) alongside the positive transfer results provides a useful reference point for subsequent research. The scale of the experiment itself is a notable strength.

minor comments (3)
  1. [Abstract and §4 (Experimental Results)] The abstract states that high-resource performance remains 'on-par with competitive bilingual baselines' but does not name the precise evaluation metric (e.g., BLEU, chrF) or the data conditions used for those baselines; the main experimental section should make these quantities explicit for reproducibility.
  2. [§3 (Model and Training)] Language sampling ratios are listed among the free parameters; an ablation or sensitivity analysis showing how performance changes when these ratios are varied would strengthen the claim that the chosen schedule successfully mitigates interference.
  3. [§4 and §5 (Analysis)] Several figures compare multilingual versus bilingual performance; adding per-language-pair variance estimates or statistical significance markers would make the 'on-par' and 'significant improvement' statements easier to evaluate.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the supportive summary, recognition of the scaling milestone, and recommendation of minor revision. No major comments appear in the provided report, so we have no specific points requiring point-by-point rebuttal. We will handle any minor editorial or clarification requests in the revised version.

Circularity Check

0 steps flagged

No significant circularity; empirical report only

full rationale

The paper reports empirical results from training one shared multilingual NMT model on >25B sentence pairs across 103 languages and compares BLEU scores against bilingual baselines. No equations, fitted parameters renamed as predictions, or derivation chain appear in the abstract or described claims. Central milestone is an achieved training outcome with accompanying practical analysis; no self-definitional, fitted-input, or self-citation load-bearing steps reduce any result to its own inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

Empirical deep learning paper; relies on standard NN training assumptions and domain assumptions about transfer in multilingual settings. No new entities postulated.

free parameters (2)
  • language sampling ratios
    Likely tuned to balance high- and low-resource languages during training.
  • model capacity and architecture hyperparameters
    Standard tuning for NMT models to achieve reported quality.
axioms (2)
  • domain assumption Gradient-based optimization converges to useful shared representations across languages
    Implicit foundation for the transfer learning claim.
  • standard math Standard backpropagation and mini-batch training apply without modification
    Background assumption for all reported training.

pith-pipeline@v0.9.0 · 5688 in / 1129 out tokens · 19662 ms · 2026-05-24T23:22:21.368079+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Parallel-SFT: Improving Zero-Shot Cross-Programming-Language Transfer for Code RL

    cs.CL 2026-04 unverdicted novelty 7.0

    Parallel-SFT mixes parallel programs across languages during SFT to produce more transferable RL initializations, yielding better zero-shot generalization to unseen programming languages.

  2. Unsupervised Cross-lingual Representation Learning at Scale

    cs.CL 2019-11 conditional novelty 7.0

    XLM-R, pretrained on 100 languages with 2TB of CommonCrawl data, improves average XNLI accuracy by 14.6 points and MLQA F1 by 13 points over mBERT while matching strong monolingual models on GLUE.

  3. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer

    cs.LG 2019-10 unverdicted novelty 7.0

    T5 casts all NLP tasks as text-to-text generation, systematically explores pre-training choices, and reaches strong performance on summarization, QA, classification and other tasks via large-scale training on the Colo...

  4. Knowledge Transfer Scaling Laws for 3D Medical Imaging

    cs.CV 2026-05 conditional novelty 6.0

    Transfer-aware data allocation derived from observed power-law scaling laws for asymmetric knowledge transfer in 3D medical imaging outperforms standard proportional sampling by up to 58% and generalizes to new budgets.

  5. COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling

    cs.LG 2026-04 unverdicted novelty 6.0

    COMPASS uses semantic clustering on multilingual embeddings to select auxiliary data for PEFT adapters, outperforming linguistic-similarity baselines on multilingual benchmarks while supporting continual adaptation.

  6. StarCoder 2 and The Stack v2: The Next Generation

    cs.SE 2024-02 accept novelty 6.0

    StarCoder2-15B matches or beats CodeLlama-34B on code tasks despite being smaller, and StarCoder2-3B outperforms prior 15B models, with open weights and exact training data identifiers released.

  7. The False Promise of Imitating Proprietary LLMs

    cs.CL 2023-05 conditional novelty 6.0

    Finetuning open LMs on ChatGPT outputs creates models that mimic style and fool human raters but fail to close the performance gap to proprietary systems on tasks not well-represented in the imitation data.

  8. Junk DNA Hypothesis: Pruning Small Pre-Trained Weights Irreversibly and Monotonically Impairs "Difficult" Downstream Tasks in LLMs

    cs.LG 2023-09 unverdicted novelty 5.0

    Pruning small-magnitude weights from pre-trained LLMs causes monotonic irreversible performance degradation on difficult downstream tasks, supporting the Junk DNA Hypothesis that these weights hold essential knowledge.

  9. Opportunities and Challenges of Large Language Models for Low-Resource Languages in Humanities Research

    cs.CL 2024-11 unverdicted novelty 2.0

    This survey paper identifies opportunities for LLMs in low-resource language humanities research along with challenges in data accessibility, model adaptability, and cultural sensitivity.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · cited by 9 Pith papers · 21 internal anchors

  1. [1]

    Massively Multilingual Neural Machine Translation

    Massively multilingual neural machine translation. CoRR, abs/1903.00089. Maruan Al-Shedivat and Ankur P Parikh. 2019. Consistency by agreement in zero-shot neu- ral machine translation. arXiv preprint arXiv:1904.02338. Naveen Arivazhagan, Ankur Bapna, Orhan Fi- rat, Roee Aharoni, Melvin Johnson, and Wolf- gang Macherey. 2019. The missing ingredient in zer...

    work page internal anchor Pith review Pith/arXiv arXiv 1903

  2. [2]

    On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization

    On the optimization of deep networks: Implicit acceleration by overparameterization. arXiv preprint arXiv:1802.06509. Mikel Artetxe, Gorka Labaka, Eneko Agirre, and Kyunghyun Cho. 2017. Unsupervised neural machine translation. arXiv preprint arXiv:1710.11041. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly l...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  3. [3]

    Character-based Neural Machine Translation

    Character-based neural machine transla- tion. CoRR, abs/1603.00810. Matthieu Courbariaux, Yoshua Bengio, and Jean- Pierre David. 2014. Training deep neural networks with low precision multiplications. arXiv preprint arXiv:1412.7024. Josep Maria Crego, Jungi Kim, Guillaume Klein, et al. 2016. Systran’s pure neural machine translation systems. CoRR, abs/161...

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  4. [4]

    Multi-domain learning by confidence- weighted parameter combination. Mach. Learn., 79(1-2):123–149. Sergey Edunov, Myle Ott, Michael Auli, and David Grangier. 2018. Understanding back-translation at scale. arXiv preprint arXiv:1808.09381. Carlos Escolano, Marta R Costa-jussà, and José AR Fonollosa. 2019. Towards interlin- gua neural machine translation. ar...

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    Multi-Way, Multilingual Neural Machine Translation with a Shared Attention Mechanism

    Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning-Volume 70, pages 1126–1135. JMLR. org. Orhan Firat, Kyunghyun Cho, and Yoshua Ben- gio. 2016a. Multi-way, multilingual neural ma- chine translation with a shared attention mech- anism. arXiv preprint arXiv:1601.01073. ...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  6. [6]

    APE at Scale and its Implications on MT Evaluation Biases

    Text repair model for neural machine translation. arXiv preprint arXiv:1904.04790. Jonas Gehring, Michael Auli, David Grangier, Denis Yarats, and Yann N. Dauphin. 2017. 20 Convolutional sequence to sequence learning. CoRR, abs/1705.03122. Dan Gillick, Cliff Brunk, Oriol Vinyals, and Amarnag Subramanya. 2015. Multilingual lan- guage processing from bytes. ...

    work page internal anchor Pith review Pith/arXiv arXiv 1904

  7. [7]

    Shampoo: Preconditioned Stochastic Tensor Optimization

    Shampoo: Preconditioned stochas- tic tensor optimization. arXiv preprint arXiv:1802.09568. David Ha, Andrew Dai, and Quoc V Le. 2016a. Hypernetworks. arXiv preprint arXiv:1609.09106. Thanh-Le Ha, Jan Niehues, and Alexander Waibel. 2016b. Toward multilingual neural ma- chine translation with universal encoder and de- coder. arXiv preprint arXiv:1611.04798....

    work page internal anchor Pith review Pith/arXiv arXiv

  8. [8]

    Achieving Human Parity on Automatic Chinese to English News Translation

    Achieving human parity on automatic chinese to english news translation. arXiv preprint arXiv:1803.05567. Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kian- inejad, Md. Mostofa Ali Patwary, Yang Yang, and Yanqi Zhou. 2017. Deep learning scal- ing is predictable, empirically. arXiv preprint arXiv:1712.00409. Chris Hokam...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  9. [9]

    One Model To Learn Them All

    Google’s multilingual neural machine translation system: Enabling zero-shot transla- tion. Transactions of the Association of Com- putational Linguistics, 5(1):339–351. Mahesh Joshi, Mark Dredze, William W. Cohen, and Carolyn Rose. 2012. Multi-domain learn- ing: When do domains matter? In Proceed- ings of the 2012 Joint Conference on Empiri- cal Methods i...

    work page internal anchor Pith review Pith/arXiv arXiv 2012

  10. [10]

    In Proceedings of the 28th International Conference on International Conference on Machine Learning , ICML’11, pages 521–528, USA

    Learning with whom to share in multi- task feature learning. In Proceedings of the 28th International Conference on International Conference on Machine Learning , ICML’11, pages 521–528, USA. Omnipress. Eliyahu Kiperwasser and Miguel Ballesteros

    work page

  11. [11]

    Learning to Segment Inputs for NMT Favors Character-Level Processing

    Scheduled multi-task learning: From syntax to translation. Transactions of the Asso- ciation for Computational Linguistics , 6:225– 240. James Kirkpatrick, Razvan Pascanu, Neil Ra- binowitz, et al. 2017. Overcoming catas- trophic forgetting in neural networks. Pro- ceedings of the national academy of sciences , 114(13):3521–3526. Philipp Koehn. 2005. Euro...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  12. [12]

    Transfer Learning in Multilingual Neural Machine Translation with Dynamic Vocabulary

    Transfer learning in multilingual neural machine translation with dynamic vocabulary. CoRR, abs/1811.01137. Guillaume Lample, Ludovic Denoyer, and Marc’Aurelio Ranzato. 2017. Unsupervised machine translation using monolingual corpora only. arXiv preprint arXiv:1711.00043. Giwoong Lee, Eunho Yang, and Sung Hwang

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  13. [13]

    Character-based Neural Machine Translation

    Asymmetric multi-task learning based on task relatedness and loss. In International Conference on Machine Learning , pages 230– 238. Jason Lee, Kyunghyun Cho, and Thomas Hof- mann. 2017. Fully character-level neural ma- chine translation without explicit segmentation. Transactions of the Association for Computa- tional Linguistics, 5:365–378. Xilai Li, Yi...

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  14. [14]

    A neural interlingua for multilingual machine translation

    A neural interlingua for multilin- gual machine translation. arXiv preprint arXiv:1804.08198. Minh-Thang Luong, Quoc V Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2015. Multi- task sequence to sequence learning. arXiv preprint arXiv:1511.06114. Minh-Thang Luong and Christopher D Manning

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  15. [15]

    Achieving Open Vocabulary Neural Machine Translation with Hybrid Word-Character Models

    Achieving open vocabulary neural ma- chine translation with hybrid word-character models. arXiv preprint arXiv:1604.00788. Jiaqi Ma, Zhe Zhao, Xinyang Yi, Jilin Chen, Lichan Hong, and Ed H. Chi. 2018a. Modeling task relationships in multi-task learning with multi-gate mixture-of-experts. In Proceedings of the 24th ACM SIGKDD International Con- ference on ...

    work page internal anchor Pith review Pith/arXiv arXiv 1930

  16. [16]

    On First-Order Meta-Learning Algorithms

    On first-order meta-learning algorithms. arXiv preprint arXiv:1803.02999. 23 Myle Ott, Sergey Edunov, David Grangier, and Michael Auli. 2018. Scaling neural machine translation. arXiv preprint arXiv:1806.00187. S. J. Pan and Q. Yang. 2010. A survey on trans- fer learning. IEEE Transactions on Knowledge and Data Engineering, 22(10):1345–1359. Kishore Papine...

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  17. [17]

    Continual Lifelong Learning with Neural Networks: A Review

    Continual lifelong learning with neural networks: A review. CoRR, abs/1802.07569. Anastasia Pentina, Viktoriia Sharmanska, and Christoph H. Lampert. 2015. Curriculum learn- ing of multiple tasks. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). A. Phillips and M Davis. 2009. Tags for Identify- ing Languages. RFC 5646, RFC Editor. ...

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  18. [18]

    Contextual Parameter Generation for Universal Neural Machine Translation

    Contextual parameter generation for universal neural machine translation. arXiv preprint arXiv:1808.08493. Emmanouil Antonios Platanios, Otilia Stretcu, Graham Neubig, Barnabás Póczos, and Tom M. Mitchell. 2019. Competence-based curriculum learning for neural machine translation. CoRR, abs/1903.09848. Maja Popovi ´c. 2015. chrf: character n-gram f- score ...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  19. [19]

    An embarrassingly simple approach to zero-shot learning. In ICML. Michael T Rosenstein, Zvika Marx, Leslie Pack Kaelbling, and Thomas G Dietterich. 2005. To transfer or not to transfer. In NIPS 2005 work- shop on transfer learning , volume 898, pages 1–4. Aurko Roy, Ashish Vaswani, Arvind Neelakan- tan, and Niki Parmar. 2018. Theory and experi- ments on v...

    work page internal anchor Pith review Pith/arXiv arXiv 2005

  20. [20]

    Improving Neural Machine Translation Models with Monolingual Data

    Parameter sharing methods for multilin- gual self-attentional translation models. Pro- ceedings of the Third Conference on Machine Translation. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural machine trans- lation models with monolingual data. arXiv preprint arXiv:1511.06709. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016....

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  21. [21]

    In AAAI Spring Sym- posium: Lifelong Machine Learning

    Lifelong machine learning systems: Be- yond learning algorithms. In AAAI Spring Sym- posium: Lifelong Machine Learning. Shagun Sodhani, Sarath Chandar, and Yoshua Bengio. 2018. On training recurrent neu- ral networks for lifelong learning. CoRR, abs/1811.07017. Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2019. Mass: Masked sequence to sequ...

    work page arXiv 2018

  22. [22]

    Multilingual Neural Machine Translation with Knowledge Distillation

    Sequence to sequence learning with neu- ral networks. In Advances in neural informa- tion processing systems, pages 3104–3112. Xu Tan, Yi Ren, Di He, Tao Qin, Zhou Zhao, and Tie-Yan Liu. 2019. Multilingual neural ma- chine translation with knowledge distillation. arXiv preprint arXiv:1902.10461. Sebastian Thrun and Tom M. Mitchell. 1995. Lifelong robot le...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  23. [23]

    Denoising Neural Machine Translation Training with Trusted Data and Online Data Selection

    Matching networks for one shot learning. In Proceedings of the 30th International Con- ference on Neural Information Processing Sys- tems, NIPS’16, pages 3637–3645, USA. Curran Associates Inc. Wei Wang, Taro Watanabe, Macduff Hughes, Tet- suji Nakagawa, and Ciprian Chelba. 2018a. Denoising neural machine translation training with trusted data and online d...

    work page internal anchor Pith review Pith/arXiv arXiv 2016