Neural Machine Translation: a Little More Like Human Thought

Machine Translation (MT) has been with us for several decades. The first effective form developed was Rule-Based MT (RBMT), initiated in the 1950’s. RBMT became obsolete when Statistical MT (SMT) was refined in the 1990s, and one form of SMT – Phrase-Based MT – still defines major online translation services.


Since 2014, Neural Machine Translation (NMT) has moved into the language services arena, opening the door for a potential paradigm shift. This is because the way NMT operates is fundamentally different, so the form it will take as it is used and evolves is not easy to predict. It’s described as “mysterious” in a Systran blog that attempts to breakdown in detail how it works,* and part of the reason it is so complex and difficult to explain is because the NMT seeks patterns on its own without being told exactly what to look for, and in the layers of processing its hard to detect how it makes its decisions.

SMT basically works by comparing source text ‘n-grams’ – 6-word groupings of words – to target language match possibilities. NMT, on the other hand, builds its data and methods through ‘deep learning’ processes which, as NMT’s name indicates, somewhat resemble the biological neural networks of animal brains. Rather than following task-specific programming, NMT systems approach problems by seeking connections from examples.

NMT systems run on Graphical Processing Units (GPUs), which are powerful and require a fraction of the memory that the Central Processing Units (CPUs) that SMT need. However, the training that the systems require is “computationally expensive,” Google says.**

Other drawbacks are that NMT doesn’t handle rare words well and this has hindered its efficiency. But with isolated, simple sentences, Google’s NMT “reduces translation errors by an average of 60% compared to Google’s phrase-based production system,” according to a Google abstract.**

Four NMT systems are currently available: Google translate, Microsoft Translator, Systran Pure Neural Machine Translation, and an open source NMT called OpenNMT from the Harvard NLP group. As the more sophisticated MT systems that Language Service Providers utilize incorporate NMT, you can be sure that Skrivanek will keep you informed of any new capabilities the technology might make available to you.

* How does Neural Machine Translation work? October 17, 2016

**2016 Google abstract

 Contact us!

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s