A Short Introduction to Google’s Neural Machine Translation (GNMT)
Google has recently unveiled the Google Neural Machine Translation system (GNMT), the most advanced machine translation tool developed to date. GNMT goes a step beyond the word-by-word and phrase-by-phrase translation we have come to expect from tools such as Google Translate, looking at sentences as a whole. Powered by math and stats, the new system is ushering in a new age for machine translations.
First we had MT, now what the heck is GNMT?
Like other translation tools, GNMT doesn’t actually understand language in the way we humans do. It cannot completely differentiate between perfect tense and past perfect. Nor does it understand words based on their etymological value. What it does instead is use its computing power to reduce translation errors by as much as 85% on several language pairs, including English to Spanish, English to French, or English to Chinese and vice versa compared to existing machine translation tools.
Existing translation tools, including Google Translate, can carry out fairly effective word and phrase-based translations, using their computational power to take in statistical models and a lot of data – well beyond what a definitive English to Spanish dictionary would include – and use that to differentiate between idioms, phrasal verbs, or similar constructions. A computer needs a lot of data to understand phrases such as “set an example,” “set foot on,” “set in motion,” or “set up housekeeping.”
The Power of Neural Networks
To oversimplify neural networks; a neural network is essentially a cluster of processors arranged to function like neurons in the brain, with each tier receiving output from the tier before it, rather than from the same output source. Tiers are interconnected through processing nodes that can come with their own rules for processing information, and which can adapt their behaviour based on the information that has passed through them.
GNMT relies on the adaptive power of neural networking and the deep learning they make possible to translate at a sentence level. Complex predictive rulesets enable it to handle huge sets of data that are more accurate than any phrase-based translation tools available, nearing human translations in accuracy.
The Future of Machine Translation
GNMT is not as good as a human translator…yet. It is however, a major improvement over existing automated translation tools. For example, even though it may have difficulties translating uncommon words, whose infrequent use makes them harder to recognize by its neural network, GNMT can break these into smaller pieces and associate them with other words and structures of language, resulting in fewer errors even for complex translations.
Google already uses GNMT to translate Chinese to English queries, and in the future we are likely to be able to use it as easily as we now use Google Translate. It is likely to become a valuable tool for translators, speeding up translations and reducing costs.
But advanced as GNMT is, it is important to remember that it is essentially a math-powered tool adapted to work with language through rulesets. It will not eliminate the need for translators or professional translation services for companies that need the best quality translation possible.