Statistical translation at Google
In April 2006 Google opened its statistical Arabic-English translation system to everyone, as Franz Och wrote on 28 April. The system was trained on billions of words of monolingual text and aligned human translations; in NIST's 2005 evaluation it had the highest BLEU of all participants for Arabic and Chinese.
Why it matters
Translation built by statistical learning on large data moved out of research evaluations into a product open to all; in NIST's evaluation a year earlier the same system had led both the research statistical systems and rule-based SYSTRAN.
NIST MT-05, official results of 1 August 2005, BLEU-4, large data track: Arabic to English Google 0.5131, ISI 0.4657, IBM 0.4646, SYSTRAN 0.1079; Chinese to English Google 0.3531, ISI 0.3073, SYSTRAN 0.1471. 100 news articles from AFP and Xinhua per language, each with four human translations. NIST cautions that research systems, not commercial products, were scored. The post's link to these results now leads to NIST's home page, so an archived copy of the same file was read. The post gives no day of launch. The record does not claim when Google moved its other language pairs to its own system: no source on that was read.