On September 27, 2016, Google announced that it had replaced the statistical engine behind Google Translate with a deep neural network, called the Google Neural Machine Translation system, or GNMT. The blog post describing it was accompanied by a technical report, “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation” (arXiv 1609.08144).
GNMT treated translation as an end-to-end learning problem. Instead of breaking a sentence into phrases and stitching them back together, it read the whole input sentence and generated the whole output sentence using a deep network with, in Google’s description, “8 encoder and 8 decoder layers using attention and residual connections.” This was a production-scale realization of the encoder-decoder-with-attention approach that researchers had introduced for translation a few years earlier.
The quality jump was large and immediate. Google reported that GNMT “reduces translation errors by more than 55%-85% on several major language pairs,” measured by human side-by-side comparison. The system went live first for Chinese-to-English, which Google said accounted for about 18 million translations per day, and rolled out to more language pairs over the following months. Google acknowledged it could still make errors a human never would, such as dropping words or mistranslating names.
For business readers, the GNMT launch is one of the most visible demonstrations of deep learning’s practical payoff. A widely used consumer product got noticeably better almost overnight because its underlying method changed, foreshadowing the even larger leaps that transformer-based models would soon bring across many AI products.