The neural machine translation has been in development for a decade but has only recently been applied mass-market and many of the organisations who could benefit the most from it still don’t know much about it.
Without going into too much history, the problem with previous machine translation has always been that it was based on pre-translated data, algorithms, and a limited context/grammar understanding. Essentially, the machine would take what you were trying to say, attempt to identify it based on the aforementioned criteria, and give you (according to its data) what you were statistically likely trying to say. As could be imagined, this worked well on small, simple sentences; however, for more complex, detail-oriented sentences, the translation received would be exceeding difficult to understand, if it made any sense at all.
Comparatively, neural machine translation is a great leap forward. Instead of operating within the rigid parameters of statistical machine translation, by using deep learning technology and representation learning and by deploying Big Data and Artificial Intelligence, it attempts to imitate how the human brain learns and understands concepts. It does this by taking an already known understanding of a language and adapting it based on how we use language differently depending on factors such as situation and context. Not only can it give a better contextual translation by looking at how a word fits into an entire sentence instead of the few words surrounding it, but the ability to learn and self-adjust based on the structure of an individual language as well as usage variations gives it a better understanding and flexibility about how one language properly translates into another. Adding all of this together, neural machine translation’s increased accuracy has the potential to reduce post-editing time by up to 25%.
While that is a definite benefit, at the same time, we shouldn’t be too quick to give way to the rise of the machines, as there are still significant flaws. First and foremost, while they can translate words in the context of a sentence, they cannot translate words in the context of an entire document, especially if the terms differ based on a specific industry or company. Secondly, though they are, after all, language translators, machines have not yet been specifically trained to understand localised colloquialisms, so the idea of localising marketing or communication is still unavailable. Thirdly, the implementation requires a significant initial investment in hardware and neural networks, not to mention the subsequent training and maintenance costs needed for operating this technology. Finally, confirming what we all learned from the 1999 film, The Matrix, the biggest limitation for neural machine translation is that it can only work within its resources, meaning that, in order to truly expand its capability, it requires more training data.
And while the expected technological advancements, along with the predicted increase of data volumes by 40 percent year over year will continue the explosion of Big Data and facilitate the accuracy and efficiency of neural machine translations and the early adopters of the technology will most likely benefit from new effective ways to utilize data and reduce the costs of their international go-to-market strategies, there will always be a need for the human touch and we should not expect that technology will replace professional human translation services anytime soon.
EVS Translations sees neural machine translation as a technology to open new linguistic service options for both language service providers and buyers.