With the vast developments of Artificial Intelligence (AI) in recent years, a wide array of industries –ranging from factory work and manufacturing to even finance and healthcare – have begun to change and adapt accordingly, and the translation industry is no different. Translators and the industry alike have already benefitted greatly from the developments in technology, specifically machine translation. With AI at the helm, machine translation looks set to be taken to the next level. But just how far will it go?
Neural Machine Translation
It was only at the end of 2016 that Google first introduced their version of Neural Machine Translation (NMT), a technology that uses computing systems inspired by biological neural networks to increase the fluency and accuracy of machine translations. Rather than translating a sentence piece by piece, NMT looks at a sentence as a whole. The system then considers the broader context of the words in the sentence and thus applies the most relevant translation, before rearranging and adjusting the sentence using proper grammar. By applying an example based machine translation method, the system learns from millions of previous examples to arrive at the most accurate translation it can. This means that over time, as more examples are logged in the system, translations become more natural and effectively the machine learns from its previous mistakes.
So while NMT is an amazing self-learning technology that offers intelligent, instant translations we have never seen before, the systems still need to be trained initially before it can translate between language pairs. This had the linguistic technology experts wondering, “Is it possible to create a system that can translate between languages without going through the initial training process?” As of now, yes. This very impressive phenomenon, known as “Zero-Shot” translation, is set to further change the industry as we know it. To understand why this is important, it is crucial to understand how machine translators worked prior to this breakthrough.
Before Zero-Shot Translation, if someone wanted to translate Traditional Chinese into Russian with no Chinese to Russian training data, they would have to translate the Chinese into a shared third language (let’s say English) and use this language as a ‘pivot’ to translate into the target language. They would translate the Chinese into English and then the English into Russian. While eventually reaching their target language, this process led to a slightly distorted translation due to the extra set of linguistic and cultural differences faced.
Zero-Shot translation negates this extra set of steps and feeds all of the training data into one engine, allowing it to build connections across multiple languages rather than individually between them. If no training data existed between Chinese and Russian, the system would assess correlations between other language pairs before producing an output. Of course, this output still isn’t as accurate as it is when language pairs exist, but it is more accurate than using a pivot. As systems learn zero-shot translation, improvements in the process are constantly being made.
Just how accurate is it?
This is a massive development for the translation industry and certainly further bridges the gap between human and machine translation, but is it as accurate as a human translator? On the most part, no. The technology is still relatively new and isn’t as accurate as Google’s previous NMT when they had trained language pairs in place. Yet this technology does show the unparalleled level of skill of artificial intelligence. Human translators study and train themselves for years before they can be as competent as translation systems, and Google has produced something that can translate without any training. While still behind the humans in terms of quality, with advances like this, the machines are quickly gaining ground in the battle of man vs. machine.