Getting Started with Watson Neural Machine Translation

Exciting changes are coming to the IBM Watson Language Translator service. Gain early access via our API header to the new Neural Machine Translation (NMT) technology for the following 13 initial language pairs:

  • English to and from: Arabic, Chinese, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese (Brazilian), Russian, Spanish, and Turkish.
  • French to and from: German, Spanish
  • German to and from: Italian

NMT is a new machine translation method based on Deep Learning that has led to improvements in translation fluency, achieving higher human evaluations compared to previous technologies. With NMT, developers will see more natural translations between languages at improved speeds and quality.

When calling an API request, specify the header X-Watson-Technology-Preview:2017-07-01 along with the character codes for the source and target languages you want to use.

The following example shows how to translate English to Spanish with an NMT preview model:


Currently, NMT early access does not support corpus customization and only supports forced glossary customization.

Learn More About Watson Language Translator

Watson Language Translator Website | Watson Developer Slack Community | Watson LT Release Notes

2 comments on"Early Access | Watson Neural Machine Translation"

  1. Interesting to try this and compare results against SMT for English -> French and English -> Italian. Although just a small sample size, my feeling is at the moment in general it is not quite up to the SMT standard, though for some short and very short pieces of text (e.g. involving titles, but not of known works), it produces better results. (Apparently Yandex uses both SMT and NMT then somehow decides which was the best translation, which seems a reasonable approach apart from the issue of how to decide which…)

    Also worth noting that the same HTTP header can be sent to the `models` API to get the list of models supported by NMT.

  2. @Jake – this is strange since NMT is superior to SMT in almost all aspects. So, this means that the model needs to get improved. I would be curious to know how it is built – i.e. if it uses multi-language model with one encoder and multiple decoders or they have built a separate model for each pair. I think IBM should also experiment with ensemble models (SMT + NMT) to have more professional outputs for different domains, or experiment with including syntax in the NMT model.

Join The Discussion

Your email address will not be published. Required fields are marked *