Is Neural Machine Translation (NMT) a game changer? I last wrote about the unreliability of computer-based translation a couple of years ago. You might be forgiven for thinking things are very different now. The arrival of NMT means that software can now translate text as well as human translators, it is said. But is this really true?
What is NMT?
NMT is a form of artificial intelligence (AI). It can work reasonably well because it ingests massive amounts of translation data. You may have noticed that Google Translate’s fluency has improved recently – that’s a result of NMT.
NMT’s neural networks are, in simple terms, trained to recognise language patterns, rather than just words. This allows them to translate – hopefully – a sentence in, say Spanish, into its English equivalent. But that improvement in fluency is not perfect. It’s just less clunky and might just be mistaken for something produced by a human.
Two NMT Problems: Reliability and Context
Reliability: NMT translations are frequently unreliable. They often leave out entire words and/or phrases and have real problems with negatives. So, if you wanted to say something as simple as ‘Do not switch this off’– this could easily end up as ‘Do switch this off’. Part of this unreliability is a result of input versus output. If the text is very different from what it ‘knows’ then claptrap will be the result.
Context: NMT tends to be single sentence driven. It ignores what has gone before or comes after. When we read we take context into account, it informs our understanding of what is being said. When context is ignored it can be jarring and/or unintelligible. Because NMT is inconsistently inaccurate, it is difficult to detect mistakes. This is compounded by its translations looking, at first glance, superficially fluent and accurate. This can make mistakes even harder to pick up.
Some day, it might be that NMT will replace human translation. Until then, don’t take the risk. At Accutranslate our translators will get it right for you and for your business.