ML Newsnew | past | best | rss | submit about
Multilingual Neural Machine Translation in the Wild (arxiv.org)
25 points by hardmaru on Jul 13, 2019 | hide | past | pdf | 4 comments on HN

In plain words: A single translation system handles 103 languages, learning from all of them at once so it can translate between any pair. It boosted quality for languages with little training data while matching the usual one-language-pair systems on major languages.

Abstract · Massively Multilingual Neural Machine Translation in the Wild: Findings and Challenges

We introduce our efforts towards building a universal neural machine translation (NMT) system capable of translating between any language pair. We set a milestone towards this goal by building a single massively multilingual NMT model handling 103 languages trained on over 25 billion examples. Our system demonstrates effective transfer learning ability, significantly improving translation quality of low-resource languages, while keeping high-resource language translation quality on-par with competitive bilingual baselines. We provide in-depth analysis of various aspects of model building that are crucial to achieving quality and practicality in universal NMT. While we prototype a high-quality universal translation system, our extensive empirical analysis exposes issues that need to be further addressed, and we suggest directions for future research.

Naveen Arivazhagan, Ankur Bapna, Orhan Firat, Dmitry Lepikhin, Melvin Johnson, Maxim Krikun, Mia Xu Chen, Yuan Cao, George Foster, Colin Cherry, Wolfgang Macherey, Zhifeng Chen, et al.
arXiv:1907.05019 · cs.CL, cs.LG · submitted Jul 11, 2019
abstract · pdf · html

add comment on HN

m0zg on Jul 14, 2019 [–]
People working on machine traslation need to figure out what is it that https://www.deepl.com/translator is doing, and then do the same thing. All sorts of papers get presented claiming "state of the art" results on translation, and meanwhile Deepl is sitting there in the corner, smiling, and offering translation which is so much better than the purported "state of the art" (and rather obviously so) it's not even funny.
reply
junipertea on Jul 15, 2019 | parent [–]
"People working on machine translation" also release their findings and sometimes even code for the benefit of everyone.
reply
m0zg on Jul 15, 2019 | parent [–]
Yes, but this is one of those rare areas where the purported academic "state of the art" is far below the _actual_ state of the art in the industry. Don't get me wrong, I'm grateful that they're releasing results at all, and NLP has gotten really exciting in the past couple of years thanks to those publications. It's just that there seems to be a bit of a research bubble going on and people are veering into ever more exotic and obscure problems before they've achieved _actual_ SOTA result on even some of the most common language pairs.
reply
trevyn on Jul 14, 2019 [–]
(This is from the Google AI team.)
reply
About | RSS | RSS (all) | HN arXiv