MT and NMT – what’s the state of play?
Sophisticated machine translation (MT) and neural machine translation (NMT) technologies are emerging in the translation industry at an ever-increasing pace. Hardly a month goes by without a tech giant announcing a new advancement or releasing a device that purports to solve the world’s translation woes.
But what is the difference between MT and NMT, and what is the state of play with each?
Statistical machine translation (SMT) has been the most widely used technology over the last decade. It uses statistical algorithms to process large volumes of data from existing translations (translation memories). Rather than just learning individual words in isolation and attempting to string them together in a meaningful way, it learns by processing so-called “n-grams”: chains of up to six linguistically linked words (rarely above four). Where it fails is when linguistic rules occur outside or at the juncture of these word chains. Basically, SMT cannot capture nuance, humour, artistic expression, cultural and political references, which are crucial to understanding.
Ask any professional translator today and they will agree that although machine translation can be impressive, it has a long way to go before we can expect it to consistently deliver acceptable results. It is therefore most widely used for what’s known as “gisting”, where the reader only expects to be able to understand the gist of the message and isn’t concerned with linguistic quality (correct grammar, semantics, idiomatic quality).
In the professional translation sector, where publication quality is expected, SMT plays a less widespread but still significant role as a productivity-enhancing tool. Rather than having to think up the full translation from scratch for each sentence, the translator instead proofreads and edits the raw machine translation output to, ideally, make it indistinguishable from a fully human-produced translation.
But now there’s a relative newcomer to the race: machine translation’s more advanced and sophisticated cousin, Neural Machine Translation (NMT).
NMT also relies on existing translation memories and statistical algorithms, but combines them with “deep learning” to emulate a human brain’s neural network. It operates on the assumption that there is a finite set of variants available for any translatable text and it calculates all the possible combinations until it identifies the best one. Finding the best one is always a work in progress and the machine, theoretically, will learn from its mistakes and improve its results with each try.
NMT is generally considered a form of artificial intelligence (AI), and a study earlier this year by Oxford University suggests that within 120 years, AI will have replaced all human jobs. However, a much more startling and impending prediction is that by 2024 technology will be able to perform translation as well as a human who is fluent in both languages but unskilled at translation, for most types of text and for most popular languages. Another prediction is that tech will be able to write a novel or short story good enough to make it to the New York Times best-seller list by 2049.
If true, this is both terrifying and awe inspiring in equal measure.
Google, Amazon, Facebook, Baidu and Microsoft are among some of the big hitters who are developing NMT systems, and it would appear that this is where the translation industry is inexorably headed. A recent report on the translation industry in 2022, by TAUS, predicted a drastically changing landscape for the coming five years as the industry adopts more and more automation in translation production and management functionalities through machine learning.
STP founder and executive chairman, Jesper Sandberg, believes the industry cannot afford to ignore the writing on the wall and must adapt and prepare to embrace the coming changes.
‘’Mine and STP’s journey with SMT started in January 2010. It would be an understatement to say I was shocked to learn back then that using machine translation was becoming a realistic prospect for professional translators, and I was apprehensive about informing our many in-house translators that STP would start using the technology as soon as possible,’’ he admitted.
Fast-forward nearly eight years, and STP now has a mature, company-wide deployment of more than 60 ever-improving SMT engines for our core languages, which our project managers and in-house translators use to aid them in their work.
Jesper said it had been a long and bumpy road with several failures and reboots along the way and with no sign of a net return on investment until the last few years.
“But these are the risks early adopters must take,” he said.
‘’Our investments in technical capacity and human skills relating to MT and post-editing have kept us in the game and may even have strengthened our market position. But I know for sure they have helped prepare us for the next big step in MT – neural.”