« Back to all posts

Google vs. Facebook Machine Translation – What is the difference?

Google vs. Facebook NMT – What is the difference?

In this era of multi-mediality and media communication, we want to be able to communicate with people all over the world without a little language difference getting in the way. When this happens online, machine translation steps in to help make languages understandable across the world.

Two leaders of online global communication, Google and Facebook, are focused on improving their own machine translation systems in order to allow users to share their lives with ease and have conversations across the world. According to the Facebook Artificial Intelligence Research (FAIR) “language translation is important to Facebook’s mission of making the world more open and connected, enabling everyone to consume posts or videos in their preferred language — all at the highest possible accuracy and speed.”  Since about half of Facebook’s users don’t speak English, we want to be able to write to friends in Portugal or China and have our social media do the translation work for us.

The ability to localize, rather than simply translate, is an important part of this process: Facebook and Google’s users want their translated text to sound like the language they use everyday. This means that machine translation systems need to understand idioms and the everyday evolutions of language. For example, French teenagers are creating new variations of the English word “wow,” like “uau.” Facebook’s algorithms picked up on the trend and can now translate these phrases.

Right now, two major types of machine translation are being tested by Google and Facebook. Each type has their own benefits, but Facebook’s recent decision to use convolutional neural networks (CNNs) over the more common recurrent neural networks (RNNs) is showing promise in its ability to produce translations that more closely resemble localized text.

Typically, computers translate text by reading a sentence in one language and predicting a sequence of words in another language with the same meaning. RNNs operate on this principle, translating text one word at a time in a strict left-to-right or right-to-left order. The most commonly used type of RNNs are long short-term memories (LSTMs): given a sequence of words they predict the probability of each word given the previous words. Google Translate and many other text applications use RNNs to search through a database of texts and uses statistical analysis to suggest the most likely translation to the user.

Facebook recently tested a form of machine translation based on a CNN approach, which has typically been used for image recognition tasks. Unlike RNNs, which process information linearly and methodically, CNNs can process information in a hierarchy, which allows them to look for non-linear relationships in data. When it comes to translation, this means that a CNN can more easily grasp contextual meaning and translate accordingly.

The biggest strength of Facebook’s CNN, however, is its multi-hop attention capability, which mimics the way humans translate a sentence. When we break down a sentence, we don’t just look at it once and then translate completely. Instead, humans return to the sentence multiple times to check and double-check meaning. Multi-hop attention means a CNN can replicate this process: the network looks repeatedly at the sentence and makes choices about what to translate first. A first “look” might point the CNN to a verb; a second “look” to an associated verb or a subject. By translating in multi-hops, a CNN can figure out relationships in the text that might emerge contextually or hierarchically. Although RNNs are valued for their accuracy, Facebook’s CNN can translate an entire sentence nine times faster than the strongest RNN (which is what Google uses).

And Facebook has an additional advantage: it studies the real-time language use of its users, so its CNNs can translate real sentences and figure out how Facebook users are actually using language. So when you use Facebook, you’re teaching its algorithm how to translate effectively.

So what does this mean for translators? It can seem like these new technological developments are slated to replace the hard work of translators in the not-too-distant future. However, there’s still hope: although Google’s Neural Machine Translator and Facebook’s CNN brain can produce approximate meanings of text, much of the nuance of meaning still gets lost in translation. By observing how humans speak, Facebook’s CNN can learn to produce something like localized text, although it can’t know the cultural reasons for a change. What this tells us is that someday, CNNs may help assist translators to begin the translation and localization process. It’s unlikely that the machines will take over translation jobs completely. Rather, translators of the future will probably be working hand-in-hand with artificial intelligence.

Localization in an agile environment