Elamur Research⌕
☰
home / papers / 2017 / Learning to Translate in Real-time with Neural Machine Translation

Learning to Translate in Real-time with Neural Machine Translation

Summary

The paper addresses simultaneous machine translation, where output must start before the speaker finishes the sentence. Earlier systems mostly cut the input into segments and translated each one separately. The authors instead treat translation as a sequence of two interleaved actions: READ (take one more source word) and WRITE (emit one target word).

An NMT model with a unidirectional, left-to-right encoder is first pre-trained on full sentence pairs. It then acts as an environment for a recurrent agent. At each step the agent sees the decoder's current candidate word and decides whether to accept it or keep waiting. The agent is trained with policy gradient. Its reward combines a quality term, based on incremental smoothed BLEU, with a delay term that can be tuned toward a chosen target. The authors also adapt beam search to this streaming setting.

Experiments on EN-RU and EN-DE in both directions show clear gains over the NMT-based waiting heuristic of Cho and Esipova (2016) and over a conventional segmentation baseline. This makes it a learnable, tunable approach to real-time translation.

Key points

  • The encoder is a unidirectional RNN, because a bidirectional encoder needs the sentence's final word before it can start. The decoder is attention-based but attends only to the source prefix read so far.

  • The agent's observation joins three things: the current context vector, the decoder state, and the embedding of the greedy candidate word. The policy itself is a stochastic RNN.

  • Quality reward is the step-by-step difference in smoothed sentence-level BLEU. The brevity penalty is applied only at the end of the sentence, so partial outputs are not unfairly punished.

  • Two delay measures are used:

    • Average Proportion (AP) is global: the average share of the source that had been read when each target word was produced.
    • Consecutive Wait length (CW) is local: how many READs happen in a row between two WRITEs.

    Both can be steered toward target values (d*, c*).

  • Code and data are released at github.com/nyu-dl/dl4mt-simul-trans.

  • Qualitative analysis shows the main weakness. In DE→EN, German verbs often come late, and the agent sometimes gives up waiting and guesses the verb, for example writing "being paid" instead of "covered" for "gedeckt".

Interesting ideas

  • I1. Frozen model as an environment. The pre-trained NMT system stays fixed, and only a lightweight controller learns when to act. The same split could add latency control to other streaming tasks, such as incremental ASR or live captioning, without retraining the base model.
  • I2. Turning a sentence-level metric into per-step rewards. Rewarding the change in a partial metric at each step, and leaving whole-sentence penalties until the end, gives denser feedback. This trick transfers to any RL setup scored by a metric like BLEU or chrF.
  • I3. Measuring delay at two levels. A global measure (AP) and a local measure (CW) capture different user complaints: overall lag versus long silences. Evaluating streaming systems on both gives a more honest picture than a single latency number.
  • I4. Configurable target delay. Training toward an explicit delay target, rather than simply minimizing delay, lets one framework serve applications with different latency budgets. This is useful when the same model must work for both lectures and conversations.
  • I5. Diagnosing word-order problems through alignment heatmaps. Plotting attention over the growing input reveals where the policy stalls or guesses, such as verb-final clauses. This analysis could help identify which language pairs, including Ukrainian ones with flexible word order, need extra handling.

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

109 papers of our library cite it (of the 140884 whose bibliographies we read); 7 of them have a page here, 102 not yet.

Cite

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

@inproceedings{gu2017learning,
  title = {{Learning to Translate in Real-time with Neural Machine Translation}},
  author = {Gu, Jiatao and Neubig, Graham and Cho, Kyunghyun and Li, Victor O.K.},
  booktitle = {Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers},
  year = {2017},
  doi = {10.18653/v1/e17-1099},
  url = {https://doi.org/10.18653/v1/e17-1099}
}

References

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

  1. Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2014. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473. find our card
  2. Srinivas Bangalore, Vivek Kumar Rangarajan Sridhar, Prakash Kolan, Ladan Golipour, and Aura Jimenez. 2012. Real-time incremental speech-to-speech translation of dialogs. In Proceedings of the 2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langua… find
  3. Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? arXiv preprint arXiv:1606.02012. — [arXiv:1606.02012](https://arxiv.org/abs/1606.02012) find
  4. Christian Fügen, Alex Waibel, and Muntsin Kolss. 2007. Simultaneous translation of lectures and speeches. Machine Translation, 21(4):209–252. find
  5. Tomoki Fujita, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2013. Simple, lexicalized choice of translation timing for simultaneous speech translation. In INTERSPEECH. find
  6. Alvin Grissom II, He He, Jordan Boyd-Graber, John Morgan, and Hal Daumé III. 2014. Dont until the final verb wait: Reinforcement learning for simultaneous machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1342– 1352, Doha,… find
  7. Andriy Mnih and Karol Gregor. 2014. Neural variational inference and learning in belief networks. arXiv preprint arXiv:1402.0030. — [arXiv:1402.0030](https://arxiv.org/abs/1402.0030) find
  8. Yusuke Oda, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2014. Optimizing segmentation strategies for simultaneous speech translation. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 551–556, Baltimore, M… find
  9. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting on association for computational linguistics, pages 311–318. Association for Computational Linguistics. find our card
  10. Vivek Kumar Rangarajan Sridhar, John Chen, Srinivas Bangalore, Andrej Ljolje, and Rathinavelu Chengalvarayan. 2013. Segmentation strategies for streaming speech translation. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human La… find
  11. Harsh Satija and Joelle Pineau. 2016. Simultaneous machine translation using deep reinforcement learning. Abstraction in Reinforcement Learning Workshop, ICML2016. find
  12. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909. find
  13. Navdeep Jaitly, Quoc V Le, Oriol Vinyals, Ilya Sutskeyver, and Samy Bengio. 2015. An online sequence-to-sequence model using partial conditioning. arXiv preprint arXiv:1511.04868. — [arXiv:1511.04868](https://arxiv.org/abs/1511.04868) find
  14. Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural networks. In Advances in neural information processing systems, pages 3104–3112. find
  15. Diederik Kingma and Jimmy Ba. 2014. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980. find
  16. Ronald J Williams. 1992. Simple statistical gradient-following algorithms for connectionist reinforcement learning. Machine learning, 8(3-4):229–256. find
  17. Chin-Yew Lin and Franz Josef Och. 2004. Automatic evaluation of machine translation quality using longest common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, page 605. Association for Computational Linguistics. find our card
  18. Mahsa Yarmohammadi, Vivek Kumar Rangarajan Sridhar, Srinivas Bangalore, and Baskaran Sankaran. 2013. Incremental segmentation and decoding strategies for simultaneous translation. In IJCNLP, pages 1032–1036. find
  19. Yuping Luo, Chung-Cheng Chiu, Navdeep Jaitly, and Ilya Sutskever. 2016. Learning online alignments with continuous rewards policy gradient. arXiv preprint arXiv:1608.01281. — [arXiv:1608.01281](https://arxiv.org/abs/1608.01281) find
  20. Lei Yu, Jan Buys, and Phil Blunsom. 2016. Online segment to segment neural transduction. arXiv preprint arXiv:1609.08194. find
  21. Takashi Mieno, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2015. Speed or accuracy? a study in evaluation of simultaneous speech translation. In INTERSPEECH. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/gu-2017-learning-translate-real-time-neural/