The problem with gathering people around the world on a social media platform is the fact that they’ll struggle to communicate due to language and culture differences. Although the latter are harder to overcome, Facebook believes it has found the way to deal with the first. The company is using neural networks to make more accurate translations #machinemagic
The 4.5 billion automatic translations Facebook tackles every day used to be managed by simple, phrase-based machine translation models. But in this case simple didn’t mean better, so Mark Zuckerberg’s company switched to a more complex but effective system – machine learning: “We need to account for context, slang, typos, abbreviations, and intent simultaneously.”
So, Facebook has started processing those translations with the help of neural networks. If the previous system translated individual words or short phrases in sentences and then put them together, the new one considers whole sentences before translation. Therefore, the translation is more fluid and contextual. LSTM or long short-term memory network is the component that makes this magic possible.
Another reason why neural networks are preferred to the old system is their ability to generate an alternative word when it doesn’t have a direct corresponding in another language. It can also turn abbreviations into full words.
“With the new system, we saw an average relative increase of 11 percent in BLEU — a widely used metric for judging the accuracy of machine translation — across all languages compared with the phrase-based systems,” the company said.
Follow TechTheLead on Google News to get the news first.
