Elamur Research⌕
☰
home / papers / 2013 / Recurrent Continuous Translation Models

Recurrent Continuous Translation Models

Summary

Phrase-based statistical machine translation treats phrase pairs with different surface forms as unrelated. It also depends on latent alignments. Together these cause sparsity and weak generalisation to rare or unseen phrases. Kalchbrenner and Blunsom propose Recurrent Continuous Translation Models (RCTM), which work only with continuous vectors for words, phrases and sentences and use no alignments or phrase tables. A recurrent language model generates the target sentence word by word with no Markov truncation. It is conditioned on the source sentence through a Convolutional Sentence Model (CSM):

  • RCTM I adds a single source-sentence vector to every hidden state.
  • RCTM II uses a truncated CSM to produce intermediate representations aimed at target words.

The whole system is trained jointly. On English→French, the models reach a perplexity more than 43% below a strong IBM Model 2 variant. They react strongly to scrambled source word order. When reranking n-best lists, they nearly match the cdec system. This matters because it shows a single end-to-end neural model can learn both translation and target-language modelling without hand-built alignment machinery.

Key points

  • Factorisation: P(f|e) is broken into per-word conditionals, each depending on all earlier target words and the full source sentence. Scoring cost grows linearly with source and target length.
  • Perplexity: the RCTMs beat IBM Model 1 by a wide margin and score more than 43% below a state-of-the-art IBM Model 2 variant (Dyer et al., 2013).
  • Word-order test: randomly permuting source words makes perplexity against the references clearly worse, so the model depends on word position.
  • Sampling test: RCTM II drew 2000 samples per sentence from its own distribution (French vocabulary of 34,831 words, top-5 restricted sampling). The outputs mostly agree with the source in number, tense and meaning.
  • Rescoring results: the task was reranking cdec 1000-best lists on four WMT-NT sets.
    • RCTM + word penalty scored 19.7–19.8 / 21.1 / 22.5 / 21.5–21.7 Bleu.
    • The full cdec system with 12 features scored 19.9 / 21.2 / 22.6 / 21.8.
    • Adding a separate monolingual RLM gave no gain.

Interesting ideas

  • I1. Parser-free hierarchical encoding: the convolutional encoder combines local n-grams first and then works up to the whole sentence, with no parse tree. This suits languages without good parsers, such as many low-resource ones, and the output never depends on a choice of parse.
  • I2. Order-shuffling as a diagnostic: comparing perplexity on original and shuffled source input is a cheap way to check whether a conditional model actually uses word order, and it applies to any sequence encoder.
  • I3. Inspecting a model by sampling: drawing candidates from a restricted top-k distribution and ranking them by model probability needs no external decoder. It is a lightweight way to see what a generative model has learned.
  • I4. Minimal-feature reranking test: pairing the model's log-probability with only a word penalty, then comparing against a richly engineered system, shows whether one learned score can stand in for many hand-made features.
  • I5. Translation models as conditioned language models: treating translation as a language model conditioned on a source vector is a flexible design. The authors note it could take extra context, such as discourse or multiple source languages, or use character-level recurrence for morphologically rich languages like Ukrainian.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Topics: Machine translation

Cited by

365 papers of our library cite it (of the 140884 whose bibliographies we read); 61 of them have a page here, 304 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{kalchbrenner2013recurrent,
  title = {{Recurrent Continuous Translation Models}},
  author = {Kalchbrenner, Nal and Blunsom, Phil},
  booktitle = {Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing},
  year = {2013},
  doi = {10.18653/v1/d13-1176},
  url = {https://doi.org/10.18653/v1/d13-1176}
}

References

18 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yoshua Bengio, Réjean Ducharme, Pascal Vincent, and Christian Janvin. 2003. A neural probabilistic language model. Journal of Machine Learning Research, 3:1137–1155. find
  2. Peter F. Brown, Vincent J.Della Pietra, Stephen A. Della Pietra, and Robert. L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19:263–311. find
  3. R. Collobert and J. Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In International Conference on Machine Learning, ICML. find
  4. John Duchi, Elad Hazan, and Yoram Singer. 2011. Adaptive subgradient methods for online learning and stochastic optimization. J. Mach. Learn. Res., 12:2121–2159, July. find
  5. Chris Dyer, Jonathan Weese, Hendra Setiawan, Adam Lopez, Ferhan Ture, Vladimir Eidelman, Juri Ganitkevitch, Phil Blunsom, and Philip Resnik. 2010. cdec: A decoder, alignment, and learning framework for finite-state and context-free translation models. In Proceedings of the ACL 2010 System Demonstrat… find our card
  6. Chris Dyer, Victor Chahuneau, and Noah A. Smith. 2013. A simple, fast, and effective reparameterization of ibm model 2. In Proc. of NAACL. find
  7. Edward Grefenstette, Mehrnoosh Sadrzadeh, Stephen Clark, Bob Coecke, and Stephen Pulman. 2011. Concrete sentence spaces for compositional distributional models of meaning. CoRR, abs/1101.0309. find
  8. Karl Moritz Hermann and Phil Blunsom. 2013. The Role of Syntax in Vector Space Models of Compositional Semantics. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Sofia, Bulgaria, August. Association for Computational Linguistics. Fo… find
  9. Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent Convolutional Neural Networks for Discourse Compositionality. In Proceedings of the Workshop on Continuous Vector Space Models and their Compositionality, Sofia, Bulgaria, August. Association for Computational Linguistics. find our card
  10. Hai Son Le, Alexandre Allauzen, and François Yvon. 2012. Continuous space translation models with neural networks. In HLT-NAACL, pages 39–48. find
  11. Tomas Mikolov and Geoffrey Zweig. 2012. Context dependent recurrent neural network language model. In SLT, pages 234–239. find
  12. Tomas Mikolov, Martin Karafiát, Lukas Burget, Jan Cernocký, and Sanjeev Khudanpur. 2010. Recurrent neural network based language model. In Takao Kobayashi, Keikichi Hirose, and Satoshi Nakamura, editors, INTERSPEECH, pages 1045–1048. ISCA. find
  13. Tomas Mikolov, Stefan Kombrink, Lukas Burget, Jan Cernocký, and Sanjeev Khudanpur. 2011. Extensions of recurrent neural network language model. In ICASSP, pages 5528–5531. IEEE. find
  14. Holger Schwenk, Daniel Déchelotte, and Jean-Luc Gauvain. 2006. Continuous space language models for statistical machine translation. In ACL. find
  15. Holger Schwenk. 2012. Continuous space translation models for phrase-based statistical machine translation. In COLING (Posters), pages 1071–1080. find
  16. Richard Socher, Eric H. Huang, Jeffrey Pennin, Andrew Y. Ng, and Christopher D. Manning. 2011. Dynamic pooling and unfolding recursive autoencoders for paraphrase detection. In J. Shawe-Taylor, R.S. Zemel, P. Bartlett, F.C.N. Pereira, and K.Q. Weinberger, editors, Advances in Neural Information Proc… find
  17. Richard Socher, Brody Huval, Christopher D. Manning, and Andrew Y. Ng. 2012. Semantic Compositionality Through Recursive Matrix-Vector Spaces. In Proceedings of the 2012 Conference on Empirical Methods in Natural Language Processing (EMNLP). find our card
  18. Ilya Sutskever, James Martens, and Geoffrey E. Hinton. 2011. Generating text with recurrent neural networks. In Lise Getoor and Tobias Scheffer, editors, ICML, pages 1017–1024. Omnipress. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/kalchbrenner-2013-recurrent-continuous-translation-models/