Camellia japonica: How best to add the remaining 6,994 known living languages, Jeff - https://scott-macleod.blogspot.com/2017/01/zebu-cows-breakthrough-googles-zero.html - and as what I'm calling as a new method "ethno-wiki-virtual-world-graphy" ?, "Jeff Dean via Google+1 month ago - Shared publicly Zero-shot translations: translate Korean->Japanese without ever seeing Korean->Japanese training data," re the World University and School Universal Translator with Google Translate / Zero Shot, Stanford Law Codex presentation and Language / Translation strategies?, Harvard Law JuryX course with Professor Charles Nesson, Stanford Law China Guiding Cases Project, In what ways could World University bring together legal translation with Stanford Law Codex re a high bar for generating excellent text versions in other languages, and from HarvardX video too?

Zero-shot translations: translate Korean->Japanese without ever seeing Korean->Japanese training data.

One of the very interesting aspects of our recent work on neural machine translation is that we can train a single, multi-lingual model that is able to translate between many language pairs (say, English->Japanese, English->Korean, Japanese->English, and Korean->English). This often is able to improve quality for many of the language pairs, and also simplifies many production aspects of the system. However, a really interesting property is that a system trained on these language pairs is actually able to do a reasonable job of translating between language pairs where it has never seen training data before (say, Korean->Japanese).

This blog post gives the high level details:

https://research.googleblog.com/2016/11/zero-shot-translation-with-googles.html

Our Arxiv paper has details about this approach:

https://arxiv.org/abs/1611.04558