Researchers at Google announced that the company’s recently introduced AI implementation, called Google Neural Machine Translation (GNMT) has come up with its very own language translation system.
GNMT has gone live in September 2016, less than 3 months ago. Prior to this new translation tool, Google used a model called Neural Machine Translation (NMT) – capable of translating only one pair of languages, for instance, translating Portuguese to English and vice versa; or translating Spanish to English and vice versa. Google’s old translation tool, however, cannot translate Portuguese to Spanish or vice versa on its own. Prior to GNMT, Google had to use a number of models to translate languages.
In a study published at Arxiv, Google researchers led by Melvin Johnson demonstrated that Google’s new translation tool GNMT is capable of translating languages on its own – a phenomenon called by the researchers as “zero-shot translation.” The researchers wrote that GNMT “implicitly learns to translate between language pairs it has never seen (zero-shot translation) – a working example of transfer learning within neural translation models.”
Without external input, GNMT is capable of translating Portuguese to Spanish and vice versa without using English as a bridge language. GNMT is similarly capable of translating Korean to Japanese and vice versa without using English as a bridge language. GNMT’s capacity to translate language pairs on its own is the first time “a form of true transfer learning has been shown to work for machine translation,” the researchers said.
The researchers added that GNMT simplifies production deployment as it cuts down the total number of models necessary when translating languages. The researchers, however, disclosed that the translation quality of GNMT’s zero-shot language pairs was “improved with little additional data of the language pair in question.”
GNMT is currently used by Google to translate over 100 languages, and over 140 billion words each day.