{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/googles-multilingual-neural-machine","title":"Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation","arxiv_id":"1611.04558","date":"2016-11-14","proceeding":"TACL 2017 1","authors":["Melvin Johnson","Mike Schuster","Quoc V. Le","Maxim Krikun","Yonghui Wu","Zhifeng Chen","Nikhil Thorat","Fernanda Viégas","Martin Wattenberg","Greg Corrado","Macduff Hughes","Jeffrey Dean"],"abstract":"We propose a simple solution to use a single Neural Machine Translation (NMT)\nmodel to translate between multiple languages. Our solution requires no change\nin the model architecture from our base system but instead introduces an\nartificial token at the beginning of the input sentence to specify the required\ntarget language. The rest of the model, which includes encoder, decoder and\nattention, remains unchanged and is shared across all languages. Using a shared\nwordpiece vocabulary, our approach enables Multilingual NMT using a single\nmodel without any increase in parameters, which is significantly simpler than\nprevious proposals for Multilingual NMT. Our method often improves the\ntranslation quality of all involved language pairs, even while keeping the\ntotal number of model parameters constant. On the WMT'14 benchmarks, a single\nmultilingual model achieves comparable performance for\nEnglish$\\rightarrow$French and surpasses state-of-the-art results for\nEnglish$\\rightarrow$German. Similarly, a single multilingual model surpasses\nstate-of-the-art results for French$\\rightarrow$English and\nGerman$\\rightarrow$English on WMT'14 and WMT'15 benchmarks respectively. On\nproduction corpora, multilingual models of up to twelve language pairs allow\nfor better translation of many individual pairs. In addition to improving the\ntranslation quality of language pairs that the model was trained with, our\nmodels can also learn to perform implicit bridging between language pairs never\nseen explicitly during training, showing that transfer learning and zero-shot\ntranslation is possible for neural translation. Finally, we show analyses that\nhints at a universal interlingua representation in our models and show some\ninteresting examples when mixing languages.","url_abs":"http://arxiv.org/abs/1611.04558v2","url_pdf":"http://arxiv.org/pdf/1611.04558v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"googles-multilingual-neural-machine","repo_url":"https://github.com/tilde-nlp/multilingual-nmt-data-prep","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"machine-translation","task_name":"Machine Translation"},{"task_slug":"nmt","task_name":"NMT"},{"task_slug":"sentence","task_name":"Sentence"},{"task_slug":"transfer-learning","task_name":"Transfer Learning"},{"task_slug":"translation","task_name":"Translation"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1611.04558","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}