Machine Learning in Microsoft's Translation A.I. Right Out of 'Star Trek'
Microsoft has taken some inspiration from Star Trek. The company announced Tuesday a program called Microsoft Translator that listens to someone, transcribes what they’ve said, and then shares the translated text with their audience.
The system is kind of like the universal translator from the most (or second-most, depending on who you ask) influential sci-fi series. It even speaks Klingon — Microsoft Translate supports 60 languages for speech-to-text translation, and the one used by an alien species is apparently one of them.
Microsoft said that its translator can work with up to 100 people at once, even if they’re split between multiple languages, just as well as it can facilitate communications between two people. That versatility is courtesy of new machine learning techniques at the core of the service.
Translations involve more than just figuring out the meanings of individual words. Machines have to learn how the words come together to learn the difference between “bloody rag” and “bloody hell,” for example, or to differentiate between the name Rainy and the adjective used to describe precipitation.
Microsoft used A.I. to teach its new app how to handle these distinctions. The company said it moved from statistical machine translation to deep neural network-based translations with Microsoft Translate. This is supposed to make it more efficient and more human-like.
This requires Microsoft Translator to represent every word with a 500-dimensional vector (set of 500 numbers) and to assess words in a sentence using a 1,000-dimensional vector. The app then models all of this information and translates that model, not each individual word, into other languages all at once.
“Neural networks are inspired by people’s theories about how the pattern-recognition process that occurs in the brains of multilingual humans works,” Microsoft said, “leading to more natural-sounding translations.” It’s all about translating an idea that someone was trying to convey, not the words they used.
Google did something similar with Google Translate. The app was updated to use neural networks (check) to translate whole sentences (check) to “be more like a human speaking with proper grammar” (check) in November. The result: smarter A.I. capable of more accurate translations.
Google then went a step further and taught A.I. to translate between two languages even if it had never seen them paired before. This is called zero-shot learning, and it essentially allows machines to solve problems without being trained on thousands or millions of examples.
Microsoft Translator doesn’t yet offer zero-shot learning; the company said its neural networks will have to be trained on professionally translated documents. But using A.I. does mean the app can improve over time, and Microsoft is likely to improve the service with new machine learning techniques.