More than 35% of the requests that Google Translate process will now be handled via advanced neural networks, boosting the 10-year-old tool's ability to offer translations that aren't embarrassingly wrong.
The technology, which translates whole sentences at a time, rather than reading phrase by phrase, first debuted for translations between Chinese and English in September 2016. It will roll-out to Spanish, French, German, and a handful of other languages starting today, according to what Google Product Lead, Barak Turovsky, announced at a press event in San Francisco.
Expanding the use of neural networks allows Google Translate to base translations on a sentence's context. For example, the tool currently translates an original German quote from Albert Einstein as: "No problem can be solved from the same consciousness that they have arisen." With neural machine translation, the final result will instead be: "Problems can never be solved with the same way of thinking that caused them."
Based on unannounced Beta Testing for Turkish translations earlier this week, Turovsky shared how users immediately recognised the improvement. "People just notice the difference," he said, noting an overwhelmingly positive response on Twitter.
When Google Translate launched a decade ago, its phrase-based algorithms simply looked up individual words, replaced them with the equivalent word in another language, and pieced them together. That often resulted in incoherent sentences. It later added the ability for users to suggest translation improvements, a feature Turovsky said won't be going away as Google transitions to neural networks.
The company will also expand neural machine translation to its cloud API, a paid service that third-party developers can use to integrate translation into their websites and apps.
Tidak ada komentar:
Posting Komentar