Cutting through the hype of Neural Machine Translation | Professor Andy Way

Understand with Unbabel is a series where we dive deep into the issues, topics and challenges faced as we accelerate towards a world without language barriers.
In this episode we sit down with Professor Andy Way of Dublin City University, known by many as one of the fathers of machine translation. He gives us a rundown of the history of MT, its latest developments and how he believes there will always be a role for human input and oversight of the translation process, no matter how advanced technologies become.
Available in 6 languages through Unbabel's transcription and translation platform.

[@tomaspinho9813]

Incredibly informative! It's important to know more about the history of your field, if not to only to avoid committing past mistakes.

[@esteveslisboeta]

These talks are amazing !

[@TheExpert8204]

I still don't see how machine translation is "not a threat to human translators". If machine translations do most of the work, and you are left pretty much with post-editing, the volume of work decreases, perhaps 3 or 4-fold (I base this on my personal experience of using MT engines in legal translation). So, of course, MT will not replace translators "completely", but there will simply be less work to be done. It's like saying to factory workers some 100 years ago. Not to worry, the machine will not replace you, because someone will have to operate the machine. Well, yes, that someone is 5 or 10 per cent of the previous factory workforce. The rest are gone. Not to say that they were necessarily starving or unemployed, but they had to look for a job elsewhere. That's what happens with the machine. It never leaves the workforce intact. That is the purpose of it: to increase efficiency and reduce labour costs.

[@ivymuse]

The quality of machine translation is heavily dependent on a lot of different factors, such as language pair, type of text, complexity of the text, domain, whether the text has been optimized for machine translation (using unambiguous language, shorter sentences, consistent terminology, etc.) and whether the MT system has been trained sufficiently to specifically translate this type of text. Additionally, MT systems need insane amounts of data to be trained (especially neural MT systems), which just isn't available in some domains or language pairs. None of that is really an issue to a qualified, trained human translator. Plus, the volume of texts to be translated is only increasing dramatically in the future. Of course MT translation will only improve as time goes on, but it definitely isn't at a point yet where I see the job of human translator disappearing completely. The job profile might change, that's for sure, but I think that on a lot of fronts MT just can't deliver yet. Even if it's properly trained the texts need post-editing, and I think for anything outside of a super-technical, standardized text (think books or movies etc.) it will never be on par with a human. It's not like humans always interpret all language correctly, so just think about how likely it is that a machine will.

[@robertderidder]

The problem is, that if a MT has to be corrected at all by a human, the machine part isn't of much help at all. The human still has to know what the entire correct text would be, so even if the MT has part of that correct, it is of no importance.

[@unbabel3053]

Studies have shown that post-editing MT is faster than pure human translation. That is also our experience here at Unbabel.

[@TheSnekkerShow]

Human translations also need to be checked and corrected by humans. Humans often get corrected by spellchecker when writing in their native languages.

[@HistoryandTutorials]

@@unbabel3053 It really depends. I use MT on daily basis, but sometimes for opinion articles (like Forbes) it is all rubbish. I have to delete everything and retranslate myself. This is true with standartized texts, yes, but when it comes to little bit of creativity and human expression, it collapses.