Deep Learning Drives Great Progress in Online Translations

Recent translation from Italian to English on Facebook:

The Flea market of square has no longer existed for a few years in its original place in square square. It has been moved against the will and protests of the owners of the stands, in in, cemetificato, less central and assolatissimo. It seemed that the city couldn’t wait for a minute so much was the rush that had to “retrain the area” (and I still didn’t understand what to do next). About two years after eviction (month plus, month less), space remains boarded, covered with weeds and holes, with a digger in between. A space like this is waiting for him in Mosul. Not in the middle of Florence. He follows pictures, as soon as the sun sets and me me on the outside.

Facebook recently switched its backend translation systems entirely to neural networks, which handle more than 2,000 translation directions and 4.5 billion translations every day. They say that these translations are more accurate than Facebook’s previous system, which used phrase-based machine translation models.

Sources: SiliconAngle, Facebook

Imperial College London professor Erol Gelenbe says artificial neural networks can ease language translation by executing a three-step process. The process includes word translations, syntax mapping, and contextual translation, which Gelenbe, recipient of the 2008 ACM SIGMETRICS Achievement Award, says the neural networks can achieve by storing and matching patterns. A key element of the translation process is long short-term memories (LSTMs), which support machine learning and can learn from experience. Swiss Dalle Molle Institute for Artificial Intelligence president Jurgen Schmidhuber expects LSTM recurrent neural networks to eventually enable “end-to-end video-based speech recognition and translation, including lip-reading and face animation.” Meanwhile, Google Brain recently announced its researchers are using neural networks to improve speech-to-text translation. Microsoft Research’s Rick Rashid says the creation and deployment of deep-learning neural networks by his company’s researchers has significantly reduced word error rates in transcribed translations, which he notes could be useful to international business dealings, and have a major effect on cross-industry learning.

Source: ITpro