First announced last November, Google’s neural machine translation systems enable Translate to process entire sentences at once. Previously, Translate operated on a per-phrase level, falling back to per-word for more obscure terminology. This is why digital translators tend to fall short of humans as the word is separated from the context of the rest of the sentence.
By considering the entire sentence, Google can calculate a much more accurate translation with a greater probability of matching the original. In its original research document on the technology, Google said it had observed 55% – 85% reductions in translation errors on major language pairs sourced from Wikipedia articles.
“Neural translation is a lot better than our previous technology, because we translate whole sentences at a time, instead of pieces of a sentence,” said Google. “This makes for translations that are usually more accurate and sound closer to the way people speak the language.”
After the success of its first trial with eight languages, Google is now ready to expand it to “many more.” From today, Hindi, Russian and Vietnamese translations will benefit from neural network processing. Additional languages are being prepared and will launch in the coming weeks.
Neural network processing brings substantial improvements to Google Translate. However, Google warned it’s not a magic solution that enables Translate to deliver flawless results. The new system is still capable of making “significant” errors that lead to loss of meaning. Because the matching system is based on probability and weight distribution, words can be lost or mistranslated along the way.
“GNMT can still make significant errors that a human translator would never make, like dropping words and mistranslating proper names or rare terms, and translating sentences in isolation rather than considering the context of the paragraph or page,” explained Google.
On the whole, the use of machine learning makes Translate a more useful language aid. When working with common phrases and words, the neural network should make light work of most translations. It’ll begin to fall down with more obscure vocabulary though, particularly if an unknown proper name or rare term is encountered.
Google said it will continue to work on the technology by rolling out the current iteration across Translate and developing more advanced versions for later use. The company noted that words in some longer translations require the context of the paragraph or entire page to be considered before they can be translated properly. Potentially, it could widen the AI’s reach in the future, giving it scope to understand the broader meaning of a complete text.
