The ML-series: Neural machine Translation vs. Statistics
Douglas Adams idea of translation is perfectly depicted through the Babel Fish and in the Matrix Neo downloaded language skills in the blink of an eye. The dream of computer translation is nearer than ever, seeing as earpods and gadgets have already enabled us a sneak-peak at hitchhiker’s guide to the known galaxy by babbling in your ear at the push of a button.
As a result of growing interest fro the subject we’ve decided to focus more on machine translation (MT short) and its societal impacts as well as future forecasts and industry insights.
Firstly, let’s clear out the air and define the difference between NMT and statistical ML. Larry Wasserman, professor in both the Department of Statistics and in the Machine Learning Department at Carnegie Mellon, one of the ML luminaries once stated that the difference between the two approaches is essentially zilch:
“The short answer is: None. They are … concerned with the same question: how do we learn from data?”
Yeah ok, that is not entirely true and you deserve a more nuanced answer.
Stat MT
Stats are based on probabilities, meaning that for each initial text there is an X-number of possible targets. They are thus programmed know how to select the segment with the highest probability. It’s also the reason why stats can be so buggy and deliver some pretty s****y translations when it comes to creative texts. Still, stats are a breeze of fresh wind when comparing the resources its little brother, NMT, eats up to spew out cool translations.
The upside of Stat MT are that the system is self-learning - duh!- and relatively free of human curation and thus resource. Once prepped up, all it needs is a chunk of inputting commands and to be fed continuously to which it can relate its translations. Works wonders with tech translation and in fields which present rigour and have a set of predefined rules, such as in example Law. Key is to build a decent training corpus on which the machine can fall back on. Stat MT analyse zillions of texts or sentences in two languages, breaking them down into tiny word groups of 2–3, called n-grams. The system then looks at patterns between the texts and makes up rules of translation from these n-grams. Finally, Stats see which cluster of English letters could correlate to other languages and spews out the nearest mumbo-jumbo :).
But there is always a catch. In this case, the major flaw of Stat MT stems from one teeny-tiny problem, its fallback corpus (database). So, if you have a database of fluent translations, Stat MT should perform within the norm. Incomes the problem, the internet is flooded with really bad auto-translated materials. Such as bad websites and so on. So bad and so many in fact, that they take up a huge chunk of the internet. Thus, becoming the norm. Stats can do only as much as their corpus provides them and if the database is infected with a shitty translation virus, well…
NMT
Neural Machine translation or NMT for short, are built to mimic our interconnected system, pretty much like neurons lighting up in the brain and absorbing info to recognize and assess a bigger picture and act upon it. Each individual connection it makes is rated and then it calculates the strength of attachment between these. Information is inputted via the input receptors through processing networks. Finally, it compares the sections with what was expected to come out. The major difference is it rates and learns, becoming more performant with each analysed data input.
The resources NMT eats up are immense. You could write entire chapters for each resolution and output and there is a huuuge body of work on the subject as is.
As we wrote in the past article, the future of NMT is very exciting. Right now, teams are focusing on text fluidity and defeating the human better than bot bias. Improvements come daily, in form of idiom analysis, teaching AI the transcreation process. The open source philosophy also offers a democratic teaching system to young programmers, interested in the learning process. Google and Facebook are spearheading their NMT platforms via Torch and TensorFlow, offering ML frameworks for everyone. Hardware is also advancing at hallucinating speeds through so called smart chips. These AI based chips should counterbalance the magnitude of programming required for NMT. Chips are custom-tailored for NMT to do faster and more efficient work than regular processors. Examples of such chips have been announced by Google and Amazon, to power their own enterprises.
And the future shines even brighter, especially when looking at what Google and Co. are up to these days. Samsung acquired Systran, an MT giant, with the sole purpose of retaining cutting edge translations for its own products. Google is trying to cover all fields, by for one building Noto, which provides a universal typeface family that supports every natural language. Google is also keeping really busy with instantaneous voice translation which it hopes to integrate in all Android phones. At Microsoft, programmers have presented a cutting-edge demo of the Skype Translator, which runs real-time trans of video calls in 45 languages.
So let us be real, the data shows that MT is anything but a niche and quite frankly here to stay, whether it comes in form of text-to-text, text-to-speech or other combinations. And when looking at the next 10 years, MT will change the way we perceive languages and translation altogether.
Lastly, the gap between traditional human translation and MT will close — and we’re intent on finding out just how far this can go.
About Beluga
Beluga helps fast-moving companies to translate their digital contents. With more than a decade of experience, professional linguists in all major markets and the latest translation technology at use, Beluga is a stable partner of many of the most thriving enterprises in the technology sector. The business goal: To help fast-growing companies offer their international audiences an excellent and engaging user experience.
