GNMT: neural translation in Google Translate
On 26 September 2016 Google posted GNMT, a neural translation system with eight LSTM layers in the encoder and eight in the decoder, and on 27 September moved 100 per cent of Chinese-to-English translations in Google Translate to it, about 18 million a day. In human evaluation it reduced errors against the phrase-based statistical system by 60 per cent on average.
Why it matters
Neural translation, until then mostly a research result, went into a product with millions of users and replaced there, for Chinese to English, the phrase-based statistical system Google Translate had launched with ten years earlier, by the post. It took wordpieces, quantisation and TPUs to do it.
In the first version of the preprint: WMT'14 English to French, 38.95 BLEU for a single model; 36 million sentence pairs; the base model trained for about six days on 96 NVIDIA K80 GPUs; TPU decoding 3.4 times faster than CPU. Table 10: on six pairs of 500 sentences from Wikipedia and news, scored 0 to 6, relative improvement from 58 to 87 per cent. Google's post of the same day gives '55%-85%', so two Google documents disagree. The record gives no English-German result: in version 1 it is inconsistent (24.17 in the introduction and for WPM-32K in the table, 24.61 in the text). The authors note that isolated simple sentences were evaluated.