Artificial intelligence (AI) has been at the forefront of technology in recent years, with massive companies like Facebook and Google investing heavily in its development. Both those companies have been utilizing artificial intelligence in improving the quality of their translations.
Currently, Google and Facebook can offer a rough, but far from flawless, version of original text in several other languages. I was curious to understand how artificial intelligence manages to translate anything at all, as imperfect as it is at the task currently and for the foreseeable future.
It turns out there’s been multiple schools of thought when it comes to artificial intelligence. One approach works on the basis of providing symbolic logic-based rules for the AI to operate under. One simple illustrative example of a logic-based rule would be: “if A, then B.” When computers weren’t as powerful as they are now, this approach was believed to be the best way to replicate human thinking. It provided some limited results, but with the processing power at our disposal in the past 10-15 years, an AI approach called deep learning is getting all the attention.
Deep learning uses artificial neural networks, which essentially emulate how the human brain operates. A network of artificial neurons is provided massive quantities of data, and from that data, connections and a form of understand can arise. They key difference between symbolic AI and neural networks is that while the former type of AI is given instructions, the latter can understand things without human instructions. The more data provided and the more layers in the network, the more complex the understanding can be. And now we have computers with the processing power to test this hypothesis out.
In 2012, Google X, a secretive research and development company, worked on a project that exposed a massive neural network to millions of YouTube video thumbnails, and without being given concepts like cat or human faces, it identified both from a list of 20,000 items it was given.
Similar techniques to this kind of image recognition is how AI is being used to translate. As I’ve talked about recently, translation is an art one must use finesse with. We learn how language works and how concepts and phrases translate by the wealth of our experience, by our understanding of the nuance of context. This is what is missing in current AI translation. Having the correct word-for-word translation will provide something you can basically understand, but nothing close to what a human is capable of doing. Our brains can do so much, but AI is catching up the more powerful computers get and the more data they can process.