Google: Neural Machine Translation

Wednesday, 23 November, 2016

Hello Google Essentially, Google’s “Neural Machine Translation” system converts whole sentences, rather than just word by word. It has been activated for eight language pairs to and from English and French, German, Spanish, Portuguese, Chinese, Japanese, Korean and Turkish. “These represent the native languages of around one-third of the world’s population,” writes Barak Turovsky in a piece titled “Found in translation: More accurate, fluent sentences in Google Translate.”

Note: The system behind Neural Machine Translation is being made available for all businesses through the Google Cloud Translation API.

Comments (1)

Trackback URL | Comments RSS Feed

  1. Henry Barth says:

    Google bought reCaptcha a few years ago. ReCaptcha has been used to make OCRI work better.

    But no amount of AI and machine translation will ever be able to convey the thouands of meanings of ‘fuck’ and its variations. Even the meaning of FCUK will remain obscure to non-English speakers.


    The reCAPTCHA Project lies at the intersection of cyber security, technology, and the ongoing effort by several commercial and non-profit organizations to digitize printed information. In the process of scanning and digitizing books and other print materials, approximately 20% of words are found to be inherently distorted and unreadable by Optical Character Recognition (OCR) programs. reCAPTCHA takes advantage of this glitch in the digitization process. Each reCAPTCHA image is constructed using one of these inherently distorted words as a starting point. The image is then further distorted by adding random warping and lines. Finally, reCAPTCHA pairs this image with another distorted word image also taken from printed material. The end result is a CAPTCHA that, while still readable and decipherable by humans, is more complex and more difficult for computers to read.