Elamur Research
☰
home / papers / 2019 / STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework

STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/P19-1289.pdf

Summary

Simultaneous machine translation must produce output while the speaker is still talking. Word-order mismatches make this hard: when translating an SOV language such as German into an SVO language such as English, the system may need to wait for a verb at the end of the sentence. Earlier systems, such as the READ/WRITE approach of Gu et al. (2017), still relied on models trained on whole sentences. They also could not guarantee a specific delay and required complex pipelines, including reinforcement learning.

Ma et al. instead train a model from scratch on a prefix-to-prefix basis, so each target word is predicted from only part of the source. Their main instance is the wait-k policy. It reads k source words first and then emits one target word per new source word, staying k words behind until the source ends. Because training contains many partial sentence pairs, the model learns to anticipate implicitly, for example guessing an English verb before the Chinese verb appears. The authors also propose a latency metric, Average Lagging.

On Transformer models for zh↔en and de↔en, wait-k keeps latency low with reasonable quality relative to full-sentence translation. This makes controllable-latency translation practical with only minor changes to standard seq-to-seq models.

Key points

  • The latency function g(t) is a monotonic function that sets how many source words are visible when producing target word t. Under wait-k, the source ends at decoding step |x| − k + 1, after which the remaining "tail" can be decoded with beam search.
  • A baseline called test-time wait-k applies the wait-k schedule to a full-sentence model. It cannot anticipate and performs much worse at small k, though both approaches converge toward the full-sentence result as k grows.
  • In human evaluation, the share of anticipated words was 1.4%, 0.1% and 0% for en→de and 4.5%, 1.5% and 0.6% for de→en, at k = 3, 5 and 7. This confirms that SOV→SVO translation requires more prediction.
  • en→zh proved hardest because English sentence-final temporal clauses must move much earlier in Chinese. This exposes a limit of any fixed wait-k policy.
  • Experiments use the Transformer, and earlier RNN results were similarly strong. Code, data and demos are available at simultrans-demo.github.io (Baidu Research / Oregon State University).

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

Cited by

168 papers of our library cite it (of the 140673 whose bibliographies we read); 1 of them have a page here, 167 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/p19-1289

@inproceedings{ma2019stacl,
  title = {{STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency using Prefix-to-Prefix Framework}},
  author = {Ma, Mingbo and Huang, Liang and Xiong, Hao and Zheng, Renjie and Liu, Kaibo and Zheng, Baigong and Zhang, Chuanqiang and He, Zhongjun and Liu, Hairong and Li, Xing and Wu, Hua and Wang, Haifeng},
  booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
  doi = {10.18653/v1/p19-1289},
  url = {https://doi.org/10.18653/v1/p19-1289}
}

References

22 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ashkan Alinejad, Maryam Siahbani, and Anoop Sarkar. 2018. Prediction improves simultaneous neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. find
  2. 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
  3. Srinivas Bangalore, Vivek Kumar Rangarajan Sridhar, Prakash Kolan, Ladan Golipour, and Aura Jimenez. 2012. Real-time incremental speech-to-speech translation of dialogs. In Proc. of NAACL-HLT. find
  4. Kyunghyun Cho and Masha Esipova. 2016. Can neural machine translation do simultaneous translation? volume abs/1606.02012. http://arxiv.org/abs/1606.02012. — [arXiv:1606.02012](https://arxiv.org/abs/1606.02012) ↗
  5. T Fujita, Graham Neubig, Sakriani Sakti, T Toda, and S Nakamura. 2013. Simple, lexicalized choice of translation timing for simultaneous speech translation. Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH . find
  6. Alvin Grissom II, He He, Jordan Boyd-Graber, John Morgan, and Hal Daumé III. 2014. Don’t 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. find
  7. Alvin Grissom II, Naho Orita, and Jordan Boyd-Graber. 2016. Incremental prediction of sentence-final verbs: Humans versus machines. In Proceedings of The 20th SIGNLL Conference on Computational Natural Language Learning. find
  8. Jiatao Gu, Graham Neubig, Kyunghyun Cho, and Victor O. K. Li. 2017. Learning to translate in real-time with neural machine translation. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2017, Valencia, Spain, April 3-7, 2017, Volume… find
  9. He He, Jordan Boyd-Graber, and Hal Daumé III. 2016. Interpretese vs. translationese: The uniqueness of human strategies in simultaneous interpretation. In North American Association for Computational Linguistics. find
  10. He He, Alvin Grissom II, Jordan Boyd-Graber, and Hal Daumé III. 2015. Syntax-based rewriting for simultaneous machine translation. In Empirical Methods in Natural Language Processing. find
  11. Liang Huang, Kai Zhao, and Mingbo Ma. 2017. When to finish? optimal beam search for neural text generation (modulo beam size). In EMNLP. find
  12. Navdeep Jaitly, David Sussillo, Quoc V Le, Oriol Vinyals, Ilya Sutskever, and Samy Bengio. 2016. An online sequence-to-sequence model using partial conditioning. In Advances in Neural Information Processing Systems. pages 5067–5075. find
  13. G. Klein, Y. Kim, Y. Deng, J. Senellart, and A. M. Rush. 2017. OpenNMT: Open-Source Toolkit for Neural Machine Translation. ArXiv e-prints . find our card
  14. S Matsubarayx, K Iwashimaz, N Kawaguchizx, K Toyama, and Yasuyoshi Inagaki. 2000. Simultaneous japanese-english interpretation based on early prediction of english verb . find
  15. Barbara Moser-Mercer, Alexander Künzli, and Marina Korac. 1998. Prolonged turns in interpreting: Effects on quality, physiological and psychological stress (pilot study). Interpreting 3(1):47–64. find
  16. 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). find
  17. Yusuke Oda, Graham Neubig, Sakriani Sakti, Tomoki Toda, and Satoshi Nakamura. 2015. Syntax-based simultaneous translation through prediction of unseen syntactic constituents. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Co… find
  18. Ofir Press and Noah A. Smith. 2018. You may not need attention. https://arxiv.org/abs/1810.13409. — [arXiv:1810.13409](https://arxiv.org/abs/1810.13409) ↗
  19. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Neural machine translation of rare words with subword units. arXiv preprint arXiv:1508.07909 . find
  20. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems 30. find
  21. Yilin Yang, Liang Huang, and Mingbo Ma. 2018. Breaking the beam search curse: A study of (re-) scoring methods and stopping criteria for neural machine translation. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. find
  22. Baigong Zheng, Renjie Zheng, Mingbo Ma, and Liang Huang. 2019. Simultaneous translation with flexible policy via restricted imitation learning. In ACL. find