Translation used to be a major component of a business’s marketing hierarchy, especially if advertising would be seen by many ethnicities speaking more than one language. Even intra-company communications across borders may have gone through a professional translator before machine translation became reliable. The introduction of Neural Machine Translation (NMT) brought hope that computers would, one day, be able to flawlessly translate between any language necessary.
Statistical Machine Translation
Pre-2016, computer translation was processed in a mostly phrase-based fashion; any aspiring linguist knows that this is a more effective method than ‘word-for-word’ translation, but it still had far too many errors to be businessworthy. SMT’s translation algorithm was complemented with the knowledge of a particular language’s popular terms and sentence order.
Unfortunately, SMT led to translation outputs for which the entire meaning of the sentence had changed. “For example, real-world training sets may override translations of, say, proper nouns. An example would be that ‘I took the train to Berlin’ gets mistranslated as ‘I took the train to Paris’ due to an abundance of ‘train to Paris’ in the training set.” 
For these reasons, it was more prudent to hire a professional for business-related translations – a costly and time-consuming compromise. It was, of course, of the utmost importance that marketing and communications content not be misinterpreted or confused.
Neural Machine Translation – Fast, Cheap, Accurate
In November of 2016, most large companies hosting machine translation were using neural-based systems. While consumers had little idea what it had taken to improve their translation output, linguists and computer science technicians had been hard at work for years building a translation system that could learn and grow.
In fact, the basic concept had first been conceived and contemplated in the 1990s’ Speech Recognition (SR) boom – SR’s ability to learn a user’s voice and intonation would be applied to numerous software programs to come. NMT, like SR, is a fluid program that learns based on the inputs it receives over time, as well as that which is ‘learned’ from the developer directly. This allows the translation service to expand not only its vocabulary but also its general command of a language’s grammar and syntax.
The first appearance of NMT in a peer-reviewed research article appears in 2014, after which point there is an explosion of interest within the linguistic and computer science community. By 2015, NMT systems were being used in machine translation competitions hosted by NIST to support machine translation research and development.  A year later, NMT systems accounted for 90% of the competition’s winners. 
How Does NMT Differ From SMT?
Two of the world’s largest tech companies, Google and Microsoft, use NMT-based algorithms for their translation systems – this alone suggests that there is some inherent benefit to one over the other. What makes NMT so far superior to its predecessor, SMT?
According to a study done by Tilde, a technology company focusing on supporting translation for less-supported languages,
- “NMT systems are up to five times better at handling word ordering and morphology, syntax and agreements (including long-distance agreements) than the SMT systems.
- Translations from NMT systems are more fluent and also more precise than SMT translations. 
NMT requires far more physical resources from the server; however, this makes sense when considering how much more knowledge is being taken into account per translation. For example, the average NMT system takes longer to ‘train’ than an SMT system; weeks rather than days, usually. This is a much larger range of data points to incorporate and access at any given moment during a translation sequence.
This larger, easily-curated, linguistic data set that NMT continues to expand also allows for “zero-shot translation”. Previous translation methods would have had to translate twice for an uncommon leap, such as Chinese to Russian . A third language would have been used in between the starting and destination languages, such as English. An uncanny example of the results of these translations would be the child-hood game telephone. Even two hops between language barriers can yield wildly inaccurate translations.
Another benefit of NMT that supersedes SMT’s abilities is that neural systems are able to identify and properly translate similar words into the proper parallel in the destination language. One example of this feature could be three English synonyms translating into the same German word.
Who Can Benefit From NMT Systems?
From freelancers to corporate CEOs, NMT has applications for pretty much every business venture. Marketing will always be a massive sphere for NMT’s wide array of advantages, however, they surely go beyond this. Small businesses faced with local inquiries in other languages could run texts through an NMT-based system, college students can seek help with translation, a non-native speaking immigrant could have a conversation with their in-laws, the list goes on.
Furthermore, this amazing advance in language translation goes to show what an impact neural-like learning is going to have in the world of computer science. Artificial intelligence (AI) can be a profound tool in learning, managing, and developing not only for businesses but also for individuals. We live in an exciting time, and technology will only get more complex and impressive from here.