Back

Improving Neural Machine Translation Models with Monolingual Data

Rico SennrichBarry HaddowAlexandra Birch

article2016ACL3,005 citations

Introduces back-translation to train neural machine translation models on target-side monolingual data without architecture changes, substantially increasing translation accuracy across standard and low-resource benchmarks.

Listen

Neural machine translation systems have reached strong performance using only parallel sentence pairs, yet the much larger volumes of monolingual text remain under-used. Earlier attempts to add monolingual data required separate language models and changes to the network, limiting their adoption. This paper set out to test whether monolingual target sentences could be incorporated directly into standard encoder-decoder training without any architectural modification.

The authors ran controlled experiments on the WMT 2015 English–German task and the lower-resource IWSLT 2014 Turkish–English task. They compared two ways of handling monolingual examples: pairing them with an empty source sentence or with a synthetic source sentence produced by back-translation. They also tested continued training on small amounts of in-domain monolingual or parallel data for domain adaptation. All systems were evaluated with case-sensitive BLEU on held-out test sets, and results were reported for both single models and ensembles.

Adding back-translated monolingual data produced the largest gains: 2.8–3.7 BLEU on English–German and 2.1–3.4 BLEU on Turkish–English, yielding new state-of-the-art scores. The dummy-source method gave smaller, less consistent improvements. Fine-tuning with in-domain synthetic data raised performance by a further 1.2 BLEU on TED-talk test sets. The gains came from three effects: reduced overfitting on limited parallel data, higher fluency of generated words, and better domain fit when the monolingual text matched the test distribution.

These results show that abundant monolingual text can now be turned into useful training signal for any encoder-decoder system, narrowing the gap between NMT and phrase-based systems that already exploit language models. The method is simple to apply and works even when parallel data are scarce or out of domain.

The paper recommends using larger or better-selected monolingual samples and exploring the same technique with source-side monolingual data. It notes that gains depend on back-translation quality and on how much the baseline model overfits; readers should therefore verify the approach on their own language pairs and data volumes before large-scale deployment.

University of Edinburghbilingualboostingencoder-decoderneural machine translation modelsmonolingual databack-translation quality
arXiv: 1511.06709
  • Paper: Neural Machine Translation of Rare Words with Subword Units, Rico Sennrich et al. (2016). This subsequent work extends the source's focus on neural machine translation data strategies by introducing subword segmentation to handle rare and out-of-vocabulary words.
  • Paper: Sequence-Level Knowledge Distillation, Yoon Kim et al. (2016). Building directly upon the back-translation and synthetic data techniques pioneered in the source, this paper explores sequence-level knowledge distillation to compress large translation models.
Cover for Improving Neural Machine Translation Models with Monolingual Data

Abstract

Neural Machine Translation (NMT) has obtained state-of-the art performance for several language pairs, while only using parallel data for training. Target-side monolingual data plays an important role in boosting fluency for phrase-based statistical machine translation, and we investigate the use of monolingual data for NMT. In contrast to previous work, which combines NMT models with separately trained language models, we note that encoder-decoder NMT architectures already have the capacity to learn the same information as a language model, and we explore strategies to train with monolingual data without changing the neural network architecture. By pairing monolingual training data with an automatic back-translation, we can treat it as additional parallel training data, and we obtain substantial improvements on the WMT 15 task English<->German (+2.8-3.7 BLEU), and for the low-resourced IWSLT 14 task Turkish->English (+2.1-3.4 BLEU), obtaining new state-of-the-art results. We also show that fine-tuning on in-domain monolingual and parallel data gives substantial improvements for the IWSLT 15 task English->German.

Topics

bilingualboostingencoder-decoderlanguage modellanguage pairsnetwork architectureparallel corporastatistical machine translation systemspeech translation

Keywords

neural machine translation modelsmonolingual databack-translation qualitysynthetic parallel textstatistical machine translationdeep fusiondomain adaptationparallel corpuslanguage models

Table of Contents

  • 1 Introduction
  • 2 Neural Machine Translation
  • 3 NMT Training with Monolingual Training Data
  • 3.1 Dummy Source Sentences
  • 3.2 Synthetic Source Sentences
  • 4 Evaluation
  • 4.1 Data and Methods
  • 4.1.1 English↔\leftrightarrowGerman
  • 4.1.2 Turkish→\toEnglish
  • 4.2 Results
  • 4.2.1 English→\toGerman WMT 15
  • 4.2.2 English→\toGerman IWSLT 15
  • 4.2.3 German→\toEnglish WMT 15
  • 4.2.4 Turkish→\toEnglish IWSLT 14
  • 4.2.5 Back-translation Quality for Synthetic Data
  • 4.3 Contrast to Phrase-based SMT
  • 4.4 Analysis
  • 5 Related Work
  • 6 Conclusion
  • References

Knowls

  1. Knowl 1 — Back-Translation for Integrating Monolingual Target Data in Neural Machine Translation

    model/method

    To train an attentional encoder-decoder Neural Machine Translation (NMT) model with target-language monolingual text without altering the network architecture, target monolingual data is converted into synthetic parallel data via back-translation.

    Given a human-translated parallel training dataset Dparallel={(xi,yi)}i=1ND_{\text{parallel}} = \{(x_i, y_i)\}_{i=1}^{N} and a target-language monolingual corpus Ymono={yj}j=1MY_{\text{mono}} = \{y_j\}_{j=1}^{M}, an auxiliary reverse translation model Mtarget→sourceM_{\text{target}\to\text{source}} is used to generate an automatic source translation x^j=Mtarget→source(yj)\hat{x}_j = M_{\text{target}\to\text{source}}(y_j) for each target sentence yjy_j.

    This forms a synthetic parallel dataset Dsynth={(x^j,yj)}j=1MD_{\text{synth}} = \{(\hat{x}_j, y_j)\}_{j=1}^{M} where the target side consists of genuine, well-formed human text and only the source side is synthetic. DsynthD_{\text{synth}} is mixed directly into DparallelD_{\text{parallel}}, and the forward NMT model Msource→targetM_{\text{source}\to\text{target}} is trained on the combined dataset using standard stochastic gradient descent without freezing any network parameters (such as the encoder or attention mechanism).

  2. Knowl 2 — Dummy Source Sentence Training for Target Monolingual Data and its Bottleneck

    model/method

    An alternative method for incorporating target-side monolingual data into an NMT model without architecture modifications is to treat monolingual sentences as parallel sentences with an uninformative, dummy source input. Each monolingual target sentence yy is paired with a single dummy token ⟨null⟩\langle\text{null}\rangle on the source side. Training batches alternate between authentic parallel data and dummy monolingual data (e.g., in a 1:1 ratio). For minibatches containing dummy source sentences, the network parameters of the encoder and attention model are frozen, forcing the decoder to rely exclusively on previous target words y<ty_{<t} rather than the context vector ctc_t.

    A key limitation of the dummy source sentence strategy is that it cannot scale to arbitrary proportions of monolingual data, nor can it be used for pure monolingual fine-tuning. Because the optimal output layer parameters differ between translation (which conditions on source context) and unconditional language modeling (where source context is absent), an excessive ratio of dummy source sentences causes the decoder's output layer to unlearn its conditioning on the source context vector.

  3. Knowl 3 — Translation Performance Improvements with Back-Translated Data on WMT 2015 English-German

    empirical result

    On the WMT 2015 English↔\leftrightarrowGerman translation tasks, augmenting parallel training data with synthetic parallel sentences generated via back-translation substantially outperforms both the parallel-only baseline and dummy source training across single models and ensembles.

    For English→\toGerman (WMT 2015 parallel: 4.2M sentences; German monolingual back-translated: 3.6M sentences), evaluated with case-sensitive BLEU on detokenized text:

    System newstest2014 newstest2015
    Single Ensemble-4 Single Ensemble-4
    Syntax-based baseline 22.6 - 24.4 -
    Parallel baseline 19.9 20.4 22.8 23.6
    + Monolingual (dummy source) 20.4 21.4 23.2 24.6
    + Synthetic (back-translation) 22.7 23.8 25.7 26.5

    For German→\toEnglish (WMT 2015 parallel: 4.2M sentences; English monolingual back-translated: 4.2M sentences):

    System newstest2014 BLEU newstest2015 BLEU
    PBSMT baseline 28.8 29.3
    NMT (Gülçehre et al., 2015 baseline) 23.6 -
    + Deep Fusion (Gülçehre et al., 2015) 24.0 -
    Parallel baseline 25.9 26.7
    + Synthetic (single model) 29.5 30.4
    + Synthetic (ensemble-4) 30.8 31.6

    Synthetic data yields gains of +2.8+2.8 to +3.4+3.4 BLEU on English→\toGerman and +3.6+3.6 to +3.7+3.7 BLEU on German→\toEnglish over parallel baselines.

  4. Knowl 4 — Low-Resource Turkish-English Machine Translation Performance with Synthetic Data

    empirical result

    In low-resource neural machine translation on the IWSLT 2014 Turkish→\toEnglish task (320,000 parallel sentence pairs from WIT3 and SETimes), adding 3.2 million synthetic sentence pairs back-translated from the English LDC Gigaword corpus achieves substantial gains over parallel training, dummy source training, and external language model fusion techniques.

    Evaluation results in case-sensitive tokenized BLEU across four evaluation sets:

    System / Training Data tst2011 tst2012 tst2013 tst2014
    Baseline (Gülçehre et al., 2015) 18.4 18.8 19.9 18.7
    Deep Fusion (Gülçehre et al., 2015) 20.2 20.2 21.3 20.6
    Parallel baseline (320k pairs) 18.6 18.2 18.4 18.3
    Parallel + parallelsynth\text{parallel}_{\text{synth}} (1:1) 19.9 20.4 20.1 20.0
    Parallel + Gigawordmono\text{Gigaword}_{\text{mono}} (dummy source 1:1) 18.8 19.6 19.4 18.2
    Parallel + Gigawordsynth\text{Gigaword}_{\text{synth}} (back-translated) 21.2 21.1 21.8 20.4

    Adding Gigawordsynth\text{Gigaword}_{\text{synth}} improves translation quality by an average of +2.7+2.7 BLEU over the parallel baseline and outperforms deep fusion by +0.5+0.5 BLEU on average. Back-translating the target side of the original parallel data (parallelsynth\text{parallel}_{\text{synth}}) provides an average gain of +1.7+1.7 BLEU, confirming a self-regularization effect, but novel monolingual text (Gigawordsynth\text{Gigaword}_{\text{synth}}) provides the largest improvements.

  5. Knowl 5 — Domain Adaptation of NMT Models via In-Domain Synthetic Monolingual Data

    empirical result

    Pre-trained NMT models trained on general-domain data (WMT news corpus) can be adapted to a specific target domain (TED talks from IWSLT) through fine-tuning on back-translated in-domain monolingual target text.

    Evaluating English→\toGerman models on IWSLT TED talk test sets using single models (case-sensitive detokenized BLEU):

    Model Fine-Tuning Dataset Instances tst2013 tst2014 tst2015
    Parallel (WMT) None - 25.2 22.6 24.0
    + Synthetic (WMT) None - 26.5 23.5 25.5
    + Synthetic (WMT) WMTparallel/WITmono\text{WMT}_{\text{parallel}} / \text{WIT}_{\text{mono}} (dummy) 200k/200k 26.6 23.6 25.4
    + Synthetic (WMT) WITsynth\text{WIT}_{\text{synth}} (back-translated) 200k 28.2 24.4 26.7
    + Synthetic (WMT) WITparallel\text{WIT}_{\text{parallel}} (authentic parallel) 200k 30.4 25.9 28.4

    A single epoch of fine-tuning on WITsynth\text{WIT}_{\text{synth}} (obtained by translating German TED talk text into English using an out-of-domain WMT-trained German→\toEnglish system) increases BLEU on tst2015 from 25.5 to 26.7 (+1.2+1.2 BLEU). In contrast, fine-tuning with dummy source sentences on monolingual TED data (extWITmono ext{WIT}_{\text{mono}}) yields no improvement (25.4 BLEU).

  6. Knowl 6 — Sensitivity of Target Translation Quality to Back-Translation System Quality

    empirical result

    The translation performance of a forward NMT system trained with synthetic parallel data is relatively robust to the quality of the backward translation system used to create the synthetic data.

    German monolingual text was back-translated into English using three distinct German→\toEnglish systems of varying quality, and forward English→\toGerman NMT models were trained on the resulting synthetic data. Performance is measured on WMT test sets:

    Back-Translation System DE→\toEN 2015 BLEU Forward EN→\toDE BLEU
    newstest2014 newstest2015
    None (parallel baseline) - 20.4 23.6
    Parallel model (greedy decoding) 22.3 23.2 26.0
    Parallel model (beam search, size 12) 25.0 23.8 26.5
    Synthetic model (beam search, size 12) 28.3 23.9 26.6
    Ensemble of 3 (1 per run) - 24.2 27.0
    Ensemble of 12 (4 per run) - 24.7 27.6

    A 6.0 BLEU increase in the back-translation system quality (from 22.3 to 28.3 BLEU on DE→\toEN newstest2015) results in only a 0.6–0.7 BLEU difference in the forward English→\toGerman model (from 26.0 to 26.6 BLEU on newstest2015).

  7. Knowl 7 — Distinct Roles of Back-Translated Data in Neural MT Versus Phrase-Based SMT

    empirical result

    While back-translated synthetic parallel data has been used in Phrase-Based Statistical Machine Translation (PBSMT), its utility and behavior differ fundamentally between PBSMT and NMT architectures.

    When evaluated on English→\toGerman test sets for both in-domain (WMT news, average of newstest2014/newstest2015) and out-of-domain (IWSLT TED talks, average of tst2013/tst2014/tst2015) conditions using synthetic data generated from News Crawl:

    Translation Framework / Data WMT BLEU (In-Domain) IWSLT BLEU (Out-of-Domain)
    PBSMT (Parallel only) 20.1 21.5
    PBSMT (+ Synthetic WMTsynth\text{WMT}_{\text{synth}}) 20.8 21.6
    PBSMT Gain +0.7 +0.1
    NMT (Parallel only) 21.4 23.9
    NMT (+ Synthetic WMTsynth\text{WMT}_{\text{synth}}) 24.3 25.1
    NMT Gain +2.9 +1.2

    In PBSMT, where monolingual data is already integrated via an external target nn-gram language model, synthetic parallel data acts almost exclusively as a translation model domain adaptation mechanism (+0.7 on in-domain, +0.1 on out-of-domain). In NMT, back-translation fulfills broader roles—improving target language modeling, preventing decoder overfitting, and improving fluency—yielding substantial gains even on out-of-domain evaluation (+1.2 BLEU).

  8. Knowl 8 — Regularization and Overfitting Mitigation in NMT via Synthetic Data

    empirical result

    In low-resource training scenarios, attentional NMT models are prone to severe overfitting on the parallel training set. Training with synthetic parallel data acts as a powerful regularizer.

    On the Turkish→\toEnglish translation task (320,000 parallel sentences), training cross-entropy on parallel data drops rapidly while development set cross-entropy on tst2010 begins to rise (overfit) early in training. Incorporating back-translated data—either by back-translating the target side of parallel training data (parallelsynth\text{parallel}_{\text{synth}}) or back-translating monolingual text (Gigawordsynth\text{Gigaword}_{\text{synth}})—delays the onset of overfitting, stabilizes training dynamics, and achieves lower cross-entropy and higher BLEU on the validation set throughout training.

  9. Knowl 9 — Improvement of Target Word-Level Fluency and Subword Compound Formation

    empirical result

    Training NMT models with additional target-side monolingual data improves word-level fluency and the generation of novel, well-formed compound words produced via subword units (Byte Pair Encoding).

    Evaluating German translations generated by English→\toGerman systems on newstest2015, focusing on words composed of subword units that never appeared in the parallel training corpus (reference human translations contain 1168 such novel words):

    System Novel Words Produced Attested in Monolingual Corpus Natural (Native Speaker Judged)
    Parallel baseline 1078 53.4% 74.9%
    + Monolingual (dummy) 994 61.6% 84.6%
    + Synthetic (back-translated) 1217 56.4% 82.5%

    Both dummy source and synthetic back-translated data increase the proportion of novel subword-assembled words that are verified as attested German words in external corpora (56.4%–61.6% vs. 53.4%) and judged as grammatically natural by human native speakers (82.5%–84.6% vs. 74.9%), preventing the production of nonsensical morpheme combinations.

Coverage note — Standard external tool details (such as Moses baseline hyperparameters, specific Groundhog settings, and Zemberek Turkish morphological analyzer specifics) were omitted as they reflect prior standard tools rather than the core contributions of this paper.

References

  1. 1.Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. Neural Machine Translation by Jointly Learning to Align and Translate. In Proceedings of the International Conference on Learning Representations (ICLR).
  2. 2.Nicola Bertoldi and Marcello Federico. 2009. Domain adaptation for statistical machine translation with monolingual resources. In Proceedings of the Fourth Workshop on Statistical Machine Translation StatMT 09. Association for Computational Linguistics.
  3. 3.Ondřej Bojar, Rajen Chatterjee, Christian Federmann, Barry Haddow, Matthias Huck, Chris Hokamp, Philipp Koehn, Varvara Logacheva, Christof Monz, Matteo Negri, Matt Post, Carolina Scarton, Lucia Specia, and Marco Turchi. 2015. Findings of the 2015 Workshop on Statistical Machine Translation. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 1–46, Lisbon, Portugal. Association for Computational Linguistics.
  4. 4.P.F. Brown, S.A. Della Pietra, V.J. Della Pietra, F. Jelinek, J.D. Lafferty, R.L. Mercer, and P.S. Roossin. 1990. A Statistical Approach to Machine Translation. Computational Linguistics, 16(2):79–85.
  5. 5.Mauro Cettolo, Christian Girardi, and Marcello Federico. 2012. WIT3 : Web Inventory of Transcribed and Translated Talks. In Proceedings of the 16th Conference of the European Association for Machine Translation (EAMT), pages 261–268, Trento, Italy.
  6. 6.Mauro Cettolo, Jan Niehues, Sebastian Stüker, Luisa Bentivogli, and Marcello Federico. 2014. Report on the 11th IWSLT Evaluation Campaign, IWSLT 2014. In Proceedings of the 11th Workshop on Spoken Language Translation, pages 2–16, Lake Tahoe, CA, USA.
  7. 7.Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1724–1734, Doha, Qatar. Association for Computational Linguistics.
  8. 8.Alex Graves. 2011. Practical Variational Inference for Neural Networks. In J. Shawe-Taylor, R.S. Zemel, P.L. Bartlett, F. Pereira, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems 24, pages 2348–2356. Curran Associates, Inc.
  9. 9.Çağlar Gülçehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Loïc Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2015. On Using Monolingual Corpora in Neural Machine Translation. CoRR, abs/1503.03535.
  10. 10.Barry Haddow, Matthias Huck, Alexandra Birch, Nikolay Bogoychev, and Philipp Koehn. 2015. The Edinburgh/JHU Phrase-based Machine Translation Systems for WMT 2015. In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 126–133, Lisbon, Portugal. Association for Computational Linguistics.
  11. 11.Geoffrey E. Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. 2012. Improving neural networks by preventing co-adaptation of feature detectors. CoRR, abs/1207.0580.
  12. 12.Sébastien Jean, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. 2015a. On Using Very Large Target Vocabulary for Neural Machine Translation. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1–10, Beijing, China. Association for Computational Linguistics.
  13. 13.Sébastien Jean, Orhan Firat, Kyunghyun Cho, Roland Memisevic, and Yoshua Bengio. 2015b. Montreal Neural Machine Translation Systems for WMT’15 . In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 134–140, Lisbon, Portugal. Association for Computational Linguistics.
  14. 14.Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer, Ondřej Bojar, Alexandra Constantin, and Evan Herbst. 2007. Moses: Open Source Toolkit for Statistical Machine Translation. In Proceedings of the ACL-2007 Demo and Poster Sessions, pages 177–180, Prague, Czech Republic. Association for Computational Linguistics.
  15. 15.Patrik Lambert, Holger Schwenk, Christophe Servan, and Sadaf Abdul-Rauf. 2011. Investigations on Translation Model Adaptation Using Monolingual Data. In Proceedings of the Sixth Workshop on Statistical Machine Translation, pages 284–293, Edinburgh, Scotland. Association for Computational Linguistics.
  16. 16.Minh-Thang Luong and Christopher D. Manning. 2015. Stanford Neural Machine Translation Systems for Spoken Language Domains. In Proceedings of the International Workshop on Spoken Language Translation 2015, Da Nang, Vietnam.
  17. 17.Thang Luong, Hieu Pham, and Christopher D. Manning. 2015. Effective Approaches to Attention-based Neural Machine Translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1412–1421, Lisbon, Portugal. Association for Computational Linguistics.
  18. 18.David McClosky, Eugene Charniak, and Mark Johnson. 2006. Effective Self-training for Parsing. In Proceedings of the Main Conference on Human Language Technology Conference of the North American Chapter of the Association of Computational Linguistics, HLT-NAACL ’06, pages 152–159, New York. Association for Computational Linguistics.
  19. 19.Henry Rowley, Shumeet Baluja, and Takeo Kanade. 1996. Neural Network-Based Face Detection. In Computer Vision and Pattern Recognition ’96.
  20. 20.Haşim Sak, Tunga Güngör, and Murat Saraçlar. 2007. Morphological Disambiguation of Turkish Text with Perceptron Algorithm. In CICLing 2007, pages 107–118.
  21. 21.Holger Schwenk. 2008. Investigations on Large-Scale Lightly-Supervised Training for Statistical Machine Translation. In International Workshop on Spoken Language Translation, pages 182–189.
  22. 22.Rico Sennrich and Barry Haddow. 2015. A Joint Dependency Model of Morphological and Syntactic Structure for Statistical Machine Translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2081–2087, Lisbon, Portugal. Association for Computational Linguistics.
  23. 23.Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural Machine Translation of Rare Words with Subword Units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL 2016), Berlin, Germany.
  24. 24.Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. Sequence to Sequence Learning with Neural Networks. In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, pages 3104–3112, Montreal, Quebec, Canada.
  25. 25.Alex Ter-Sarkisov, Holger Schwenk, Fethi Bougares, and Loïc Barrault. 2015. Incremental Adaptation Strategies for Neural Network Language Models. In Proceedings of the 3rd Workshop on Continuous Vector Space Models and their Compositionality, pages 48–56, Beijing, China. Association for Computational Linguistics.
  26. 26.Francis M. Tyers and Murat S. Alperen. 2010. SETimes: A parallel corpus of Balkan languages. In Workshop on Exploitation of multilingual resources and tools for Central and (South) Eastern European Languages at the Language Resources and Evaluation Conference, pages 1–5.

Citation

MLA
Sennrich, R., et al. “Improving Neural Machine Translation Models with Monolingual Data”. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2016, pp. 86–96, https://doi.org/10.18653/v1/P16-1009.
APA
Sennrich, R., Haddow, B., & Birch, A. (2016). Improving Neural Machine Translation Models with Monolingual Data. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 86–96. https://doi.org/10.18653/v1/P16-1009
Chicago
Sennrich, R., B. Haddow, and A. Birch. 2016. “Improving Neural Machine Translation Models with Monolingual Data”. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 86–96. https://doi.org/10.18653/v1/P16-1009.
Harvard
Sennrich, R., Haddow, B. and Birch, A. (2016) “Improving Neural Machine Translation Models with Monolingual Data”, Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp. 86–96. Available at: https://doi.org/10.18653/v1/P16-1009.
Vancouver
1. Sennrich R, Haddow B, Birch A (2016) Improving Neural Machine Translation Models with Monolingual Data. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, pp 86–96

BibTeX

@inproceedings{sennrich-etal-2016-improving,
    title = "Improving Neural Machine Translation Models with Monolingual Data",
    author = "Sennrich, Rico  and
      Haddow, Barry  and
      Birch, Alexandra",
    editor = "Erk, Katrin  and
      Smith, Noah A.",
    booktitle = "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2016",
    address = "Berlin, Germany",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/P16-1009/",
    doi = "10.18653/v1/P16-1009",
    pages = "86--96"
}
Metadata:ACL Anthology

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by-nc-sa/4.0/