Unsupervised Pretraining for Sequence to Sequence Learning
Open accessThe DOI matches its Crossref record; 40 of the 41 references it cites were found in a registry or an index.
Abstract
This work presents a general unsupervised learning method to improve the accuracy of sequence to sequence (seq2seq) models.In our method, the weights of the encoder and decoder of a seq2seq model are initialized with the pretrained weights of two language models and then fine-tuned with labeled data.We apply this method to challenging benchmarks in machine translation and abstractive summarization and find that it significantly improves the subsequent supervised models.Our main result is that pretraining improves the generalization of seq2seq models.We achieve state-of-theart results on the WMT English→German task, surpassing a range of methods using both phrase-based machine translation and neural machine translation.Our method achieves a significant improvement of 1.3 BLEU from the previous best models on both WMT'14 and WMT'15 English→German.We also conduct human evaluations on abstractive summarization and find that our method outperforms a purely supervised learning baseline in a statistically significant manner.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The claim comes from somewhere else, not from this source.
In the citing paper
“They initialized LSTMs with a language model (LM) or a sequence autoencoder, and found the pre-training can improve the training and generalization of LSTMs in many text classification tasks.”
From Pre-trained models for natural language processing: A survey (2020) · cites this paper as [35]
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredIn the citing paper
“Dai and Le [35] proposed the first successful instance of PTM for NLP.”
From Pre-trained models for natural language processing: A survey (2020) · cites this paper as [35]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Adam: A Method for Stochastic Optimization
Kingma, D. P., & Ba, J. (2014) · arXiv · doi:10.48550/arxiv.1412.6980 · Cited by 98,946 · cited 1×
“We used the Adam optimizer (Kingma and Ba, 2015) and train with asynchronous SGD on 16 GPUs for speed.”
MethodGlove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071 · cited 2×
“pretrained GloVe embedding vectors (Pennington et al., 2014) had more impact.”
BackgroundLearning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation
Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1179 · Cited by 25,545 · cited 2×
“Sequence to sequence (seq2seq) models (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner and Blunsom, 2013; Allen, 1987; Ñeco and Forcada, 1997) are extremely effective on a variety of tasks that require a mapping between a variable-length input sequence to a variable-length output sequence.”
BackgroundDistributed Representations of Words and Phrases and their Compositionality
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1310.4546 · Cited by 24,533 · cited 2×
“Interestingly, they use pretrained word2vec (Mikolov et al., 2013) vectors to initialize their word em- beddings.”
BLEU: a method for automatic evaluation of machine translation
Papineni, K., Roukos, S., Ward, T., & Zhu, W.-J. (2001) · Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL '02 · doi:10.3115/1073083.1073135 · Cited by 21,344 · cited 2×
“Evaluation on the validation set was with case-sensitive BLEU (Papineni et al., 2002) on tokenized text using multi-bleu.”
MethodSequence to Sequence Learning with Neural Networks
Sutskever, I., Vinyals, O., & Le, Q. V. (2014) · arXiv · doi:10.48550/arxiv.1409.3215 · Cited by 17,766 · cited 2×
“Sequence to sequence (seq2seq) models (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner and Blunsom, 2013; Allen, 1987; Ñeco and Forcada, 1997) are extremely effective on a variety of tasks that require a mapping between a variable-length input sequence to a variable-length output sequence.”
BackgroundSelf-citationNeural Machine Translation by Jointly Learning to Align and Translate
Bahdanau, D., Cho, K., & Bengio, Y. (2014) · arXiv · doi:10.48550/arxiv.1409.0473 · Cited by 16,213 · cited 1×
“Multi-layer attention: In all our models, we use an attention mechanism (Bahdanau et al., 2015), where the model attends over both top and first layer (see Figure 2-b).”
Method(2004) · Cited by 9,718
On the difficulty of training recurrent neural networks
Pascanu, R., Mikolov, T., & Bengio, Y. (2012) · arXiv · doi:10.48550/arxiv.1211.5063 · Cited by 4,310 · cited 2×
“0 (Pascanu et al., 2013), and dropout of 0.”
MethodInfluentialContext-dependent pre-trained deep neural networks for large-vocabulary speech recognition
Dahl, G. E., Yu, D., Deng, L., & Acero, A. (2012) · IEEE Transactions on Audio Speech and Language Processing · doi:10.1109/tasl.2011.2134090 · Cited by 3,011 · cited 2×
“Unsupervised pretraining has been intensively studied in the past years, most notably is the work by Dahl et al. (2012) who found that pretraining with deep belief networks improved feedforward acoustic models.”
BackgroundLong short-term memory recurrent neural network architectures for large scale acoustic modeling
Sak, H., Senior, A., & Beaufays, F. (2014) · Interspeech 2014 · doi:10.21437/interspeech.2014-80 · Cited by 2,973 · cited 2×
“, 2016) We used a 1 layer 4096 dimensional LSTM with the hidden state projected down to 1024 units (Sak et al., 2014) and trained for one week on 32 Tesla K40 GPUs.”
MethodTeaching Machines to Read and Comprehend
Hermann, K. M., Kočiský, T., Grefenstette, E., Espeholt, L., Kay, W., Suleyman, M., & Blunsom, P. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1506.03340 · Cited by 2,521 · cited 1×
“Dataset and Evaluation: For a low-resource abstractive summarization task, we use the CNN/Daily Mail corpus from (Hermann et al., 2015).”
MethodAbstractive Text Summarization using Sequence-to-sequence RNNs and Beyond
Nallapati, R., Zhou, B., dos Santos, C., Gulcehre, C., & Xiang, B. (2016) · Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning · doi:10.18653/v1/k16-1028 · Cited by 2,361
Recurrent Neural Network Regularization
Zaremba, W., Sutskever, I., & Vinyals, O. (2014) · arXiv (Cornell University) · doi:10.48550/arxiv.1409.2329 · Cited by 2,123 · cited 2×
“We used a learning rate of 5e-5 which is multiplied by 0.8 every 50K steps after an initial 400K steps, gradient clipping with norm 5.0 (Pascanu et al., 2013), and dropout of 0.2 on non-recurrent connections (Zaremba et al., 2014).”
MethodInfluentialRecurrent Continuous Translation Models
Kalchbrenner, N., & Blunsom, P. (2013) · Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d13-1176 · Cited by 1,232 · cited 2×
“Sequence to sequence (seq2seq) models (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner and Blunsom, 2013; Allen, 1987; Ñeco and Forcada, 1997) are extremely effective on a variety of tasks that require a mapping between a variable-length input sequence to a variable-length output sequence.”
BackgroundAn Empirical Investigation of Catastrophic Forgetting in Gradient-Based Neural Networks
Goodfellow, I. J., Mirza, M., Xiao, D., Courville, A., & Bengio, Y. (2013) · arXiv · doi:10.48550/arxiv.1312.6211 · Cited by 1,089 · cited 1×
“However, this procedure may lead to catastrophic forgetting, where the model’s performance on the language modeling tasks falls dramatically after fine-tuning (Goodfellow et al., 2013).”
BackgroundSemi-supervised Sequence Learning
Dai, A. M., & Le, Q. V. (2015) · arXiv · doi:10.48550/arxiv.1511.01432 · Cited by 994 · cited 1×
“Dai and Le (2015); Radford et al. (2017) are amongst the rare studies which showed the benefits of pretraining in a semi-supervised learning setting.”
BackgroundSelf-citationNeural Machine Translation of Rare Words with Subword Units
Sennrich, R., Haddow, B., & Birch, A. (2016) · Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p16-1162 · Cited by 974 · cited 6×
“Experimental settings: We use subword units (Sennrich et al., 2015b) with 31500 merges, resulting in a vocabulary size of about 32000.”
MethodInfluentialNeural Machine Translation of Rare Words with Subword Units
Sennrich, R., Haddow, B., & Birch, A. (2015) · arXiv · doi:10.48550/arxiv.1508.07909 · Cited by 950 · cited 6×
“Experimental settings: We use subword units (Sennrich et al., 2015b) with 31500 merges, resulting in a vocabulary size of about 32000.”
MethodInfluentialExploring the Limits of Language Modeling
Józefowicz, R., Vinyals, O., Schuster, M., Shazeer, N., & Wu, Y. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1602.02410 · Cited by 823 · cited 3×
“The LM is a one-layer LSTM of size 1024 trained in a similar fashion to Jozefowicz et al. (2016).”
MethodTransfer Learning for Low-Resource Neural Machine Translation
Zoph, B., Yuret, D., May, J., & Knight, K. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1163 · Cited by 788 · cited 1×
“Various forms of transfer or multitask learning with seq2seq framework also have the flavors of our algorithm (Zoph et al., 2016; Luong et al., 2015; Firat et al., 2016).”
BackgroundOn Using Monolingual Corpora in Neural Machine Translation
Gülçehre, Ç., Fırat, O., Xu, K., Cho, K., Barrault, L., Lin, H.-C., Bougares, F., Schwenk, H., & Bengio, Y. (2015) · HAL (Le Centre pour la Communication Scientifique Directe) · doi:10.48550/arxiv.1503.03535 · Cited by 466 · cited 1×
“Perhaps most closely related to our method is the work by Gulcehre et al. (2015), who combined a language model with an already trained seq2seq model by fine-tuning additional deep output layers.”
MethodAchieving Human Parity in Conversational Speech Recognition
Xiong, W., Droppo, J., Huang, X., Seide, F. T. B., Seltzer, M., Stolcke, A., Yu, D., & Zweig, G. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1610.05256 · Cited by 451 · cited 1×
“More recent acoustic models have found pretraining unnecessary (Xiong et al., 2016; Zhang et al., 2016; Chan et al., 2015), probably because the reconstruction objective of deep belief networks is too easy.”
BackgroundVery deep convolutional networks for end-to-end speech recognition
Zhang, Y., Chan, W., & Jaitly, N. (2017) · 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) · doi:10.1109/icassp.2017.7953077 · Cited by 407 · cited 1×
“More recent acoustic models have found pretraining unnecessary (Xiong et al., 2016; Zhang et al., 2016; Chan et al., 2015), probably because the reconstruction objective of deep belief networks is too easy.”
BackgroundLearning to Generate Reviews and Discovering Sentiment
Radford, A., Jozefowicz, R., & Sutskever, I. (2017) · arXiv · doi:10.48550/arxiv.1704.01444 · Cited by 330 · cited 1×
“Dai and Le (2015); Radford et al. (2017) are amongst the rare studies which showed the benefits of pretraining in a semi-supervised learning setting.”
BackgroundChan, W., Jaitly, N., Le, Q. V., & Vinyals, O. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1508.01211 · Cited by 284
Self-citationFindings of the 2015 Workshop on Statistical Machine Translation
Bojar, O., Chatterjee, R., Federmann, C., Haddow, B., Huck, M., Hokamp, C., Koehn, P., Logacheva, V., Monz, C., Negri, M., et al. (2015) · Proceedings of the Tenth Workshop on Statistical Machine Translation · doi:10.18653/v1/w15-3001 · Cited by 277
Exploiting Source-side Monolingual Data in Neural Machine Translation
Zhang, J., & Zong, C. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1160 · Cited by 277 · cited 1×
“Similarly, Zhang and Zong (2016) found it useful to add an additional task of sentence reordering of sourceside monolingual data for neural machine translation.”
BackgroundZero-Resource Translation with Multi-Lingual Neural Machine Translation
Firat, O., Sankaran, B., Al-Onaizan, Y., Yarman Vural, F. T., & Cho, K. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1026 · Cited by 261 · cited 1×
“Various forms of transfer or multitask learning with seq2seq framework also have the flavors of our algorithm (Zoph et al., 2016; Luong et al., 2015; Firat et al., 2016).”
BackgroundToward Human Parity in Conversational Speech Recognition
Xiong, W., Droppo, J., Huang, X., Seide, F., Seltzer, M. L., Stolcke, A., Yu, D., & Zweig, G. (2017) · IEEE/ACM Transactions on Audio, Speech, and Language Processing · doi:10.1109/taslp.2017.2756440 · Cited by 245
Montreal Neural Machine Translation Systems for WMT’15
Jean, S., Firat, O., Cho, K., Memisevic, R., & Bengio, Y. (2015) · Proceedings of the Tenth Workshop on Statistical Machine Translation · doi:10.18653/v1/w15-3014 · Cited by 134 · cited 1×
“7 Supervised NMT (Jean et al., 2015) single 22.”
Improving LSTM-based Video Description with Linguistic Knowledge Mined from Text
Venugopalan, S., Hendricks, L. A., Mooney, R., & Saenko, K. (2016) · Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing · doi:10.18653/v1/d16-1204 · Cited by 131 · cited 2×
“To be even more efficient, the softmax of the decoder is initialized with the softmax of the pretrained target side LM.”
BackgroundMethodSequence-to-Sequence RNNs for Text SummarizationPreprint
Nallapati, R. M., Xiang, B., & Zhou, B. (2016) · arXiv (Cornell University) · Cited by 120
Semi-Supervised Learning for Neural Machine TranslationPreprint
Cheng, Y., Xu, W. H., He, Z., He, W. P., Wu, H., Sun, M., & Liu, Y. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1606.04596 · Cited by 61 · cited 2×
“In our experiments, we benchmark our method on machine translation, where other unsupervised methods are shown to give promising results (Sennrich et al., 2015a; Cheng et al., 2016).”
MethodSeveral Studies on Natural Language ·and Back-Propagation
Allen, R. B. (1987) · Cited by 60
Asynchronous translations with recurrent neural nets
Ñeco, R. P., & Forcada, M. L. (2002) · doi:10.1109/icnn.1997.614693 · Cited by 58 · cited 2×
“Sequence to sequence (seq2seq) models (Sutskever et al., 2014; Cho et al., 2014; Kalchbrenner and Blunsom, 2013; Allen, 1987; Ñeco and Forcada, 1997) are extremely effective on a variety of tasks that require a mapping between a variable-length input sequence to a variable-length output sequence.”
BackgroundMulti-task Sequence to Sequence Learning
Luong, M.-T., Le, Q. V., Sutskever, I., Vinyals, O., & Kaiser, Ł. (2015) · arXiv (Cornell University) · doi:10.48550/arxiv.1511.06114 · Cited by 56 · cited 1×
“Various forms of transfer or multitask learning with seq2seq framework also have the flavors of our algorithm (Zoph et al., 2016; Luong et al., 2015; Firat et al., 2016).”
BackgroundSelf-citationEdinburgh's Statistical Machine Translation Systems for WMT16
Williams, P. J., Sennrich, R., Nădejde, M., Huck, M., Haddow, B., & Bojar, O. (2016) · Conference on Machine Translation (WMT) · doi:10.18653/v1/w16-2327 · Cited by 28 · cited 1×
“BLEU System ensemble? newstest2014 newstest2015 Phrase Based MT (Williams et al., 2016) 21.”
BackgroundThe Edit Distance Transducer in Action: The University of Cambridge English-German System at WMT16
Stahlberg, F., Hasler, E., & Byrne, B. (2016) · Conference on Machine Translation (WMT) · doi:10.18653/v1/w16-2324 · Cited by 10 · cited 2×
“4 Edit Distance Transducer NMT (Stahlberg et al., 2016) single 21.”
BackgroundUnsupervised Learning of Predictors from Unpaired Input-Output SamplesPreprint
Chen, J., Huang, P.-S., He, X., Gao, J., & Deng, L. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1606.04646 · Cited by 6 · cited 1×
“Related to our approach in principle is the work by Chen et al. (2016) who proposed a two-term, theoretically motivated unsupervised objective for unpaired input-output samples.”
Background
Research works printed but not linked: 3
- Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S. Cor- rado, and Jeff Dean. 2013. Distributed representa- tions of words and phrases and their compositional- ity. In NIPS.
- Courtney Napoles, Matthew Gormley, and Benjamin Van Durme. 2012. Annotated gigaword. In Pro- ceedings of the Joint Workshop on Automatic Knowl- edge Base Construction and Web-scale Knowledge Extraction. ACL.
- Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015a. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709. · arXiv
Cited by
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
“We include the language modeling objective due to its historic use as a pre-training objective [Dai and Le, 2015; Ramachandran et al., 2016; Howard and Ruder, 2018; Radford et al., 2018; Peters et al., 2018] as well as its natural fit for the language model architectures we consider.”
BackgroundMethodSelf-citationPre-trained models for natural language processing: A survey
Qiu, X., Sun, T., Xu, Y., Shao, Y., Dai, N., & Huang, X. (2020) · Science China Technological Sciences · doi:10.1007/s11431-020-1647-3 · Cited by 1,591
Cross-lingual Language Model Pretraining
Lample, G., & Conneau, A. (2019) · arXiv · doi:10.48550/arxiv.1901.07291 · Cited by 1,586
“Ramachandran et al. (2016) show that language modeling pretraining can also provide significant improvements on machine translation tasks, even for high-resource language pairs such as English-German where there exists a signifi-cant amount of parallel data.”
BackgroundHierarchical Neural Story Generation
Fan, A., Lewis, M., & Dauphin, Y. (2018) · Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · doi:10.18653/v1/p18-1082 · Cited by 1,266
“The two models can be leveraged together without architectural modifications: Ramachandran et al. (2016) use language models to initialize the encoder and decoder side of the seq2seq model independently, and Chorowski and Jaitly (2016) combine the predictions of the language model and seq2seq model…”
MethodHierarchical Neural Story Generation
Fan, A., Lewis, M., & Dauphin, Y. (2018) · arXiv · doi:10.48550/arxiv.1805.04833 · Cited by 1,242
“The two models can be leveraged together without architectural modifications: Ramachandran et al. (2016) use language models to initialize the encoder and decoder side of the seq2seq model independently, and Chorowski and Jaitly (2016) combine the predictions of the language model and seq2seq model…”
Method
Metadata
- Authors
- Ramachandran, P., Liu, P., & Le, Q.
- Affiliations
- Google (United States) (US)
- ORCID
- Le, Quoc
- Venue
- Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing
- Published
- 2017
- Pages
- 383-391
- OpenAlex
- W2555428947
- Cited by
- 251
- References
- 41
- Citations by year
- 2026: 2 · 2025: 6 · 2024: 14 · 2023: 12 · 2022: 14 · 2021: 33 · 2020: 43 · 2019: 79 · 2018: 37 · 2017: 11
- FWCI
- 20.75
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- cc-by
- Locations
- Conference on Empirical Methods in Natural Language Processing (EMNLP) (publishedVersion, cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Speech Recognition and Synthesis
- Keywords
- unsupervised pre-training, sequence-to-sequence models, transfer learning, pre-trained language models, fine-tuning, machine translation, abstractive summarization, English-German translation, generalization
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.