Google's Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation

Melvin Johnson
Google

Neural Machine Translation (NMT) has been a big success story in the deep learning revolution. It has grown out of academic labs to large-scale adoption in a short period of time. Recently, we at Google announced that we are now providing neural translations to our users. In this talk, I will briefly review the history of machine translation and explain our GNMT (Google's NMT) system. I will talk about our approach to Multilingual NMT which aims to translate between multiple languages at the same time. This opens many interesting avenues for further research. Most notably, it enables us to perform Zero-Shot translation - the ability to translate between languages the model has never seen before. Further analysis of this phenomenon hints at a presence of an interlingua - a language independent representation.

This is joint work with many members from the Google Brain and Google Translate teams.

Videos:

• Join the live presentation. Wednesday 4:30-5:45.  Requires Microsoft Windows Media player.
• View the Stanford Archive Lectures
• View video by lecture sequence in HTML5. Available after 8PM Pacific on the day of the lecture.
• View Video on YouTube. The YouTube video is uploaded the day following the presentation. You may need to disable your pop-up blocker for this link to work. Or use the direct YouTube link , which will be active a day or so after presentation.

About the speaker:

Melvin Johnson received his M.S in Computer Science from Stanford University in 2015. At Stanford, he was part of Chris Manning's Natural Language Processing group where he worked on Knowledge Base extraction and related tasks. He is now at Google, where he works on research in Machine Translation and NLP.

Contact information:

Melvin Johnson
Google