Best of Frenemies? Machine and Human Translation

It was all so simple in theory. With enough computing power, a handful of smart linguistics boffins, a few computer scientists and enough time, human translators would be as obsolete as punch card programming. No matter the text, no matter the subject, Machine Translation (MT) would give you a word perfect version every time in whatever language you wished.

As with most utopian dreams, it would all end up a bit more complicated than anyone could have predicted. Language, you see, is a funny beast. Just as you think you have it all described and analysed, along comes some creative addition or reuse. Tired of “footprint” just referring to the remnants of last night’s walk along the beach? Strap the word “carbon” onto the front and you have a chic environmental term. Bored with “surfing” only involving foamy waves and sleek hair? Mix things up a bit and add some fishing terminology and you get “surfing the net,” which, oddly enough has everything to do with technology and nothing much to do with the sea.

Even traditional songs take on a whole new slant once people start playing with language. “Don we now our gay apparel” used to mean getting dressed up for a Christmas meal. Now, on the other hand, it has a completely different meaning altogether!

All this was enough to play havoc with translation engines. How on earth could anyone write software based on nice, stable rules when word use and meaning changed all the time? What a mess!

No worries, thought the MT gurus, we have a plan. That plan was to beat creativity at its own game. Instead of trying stable, unchangeable rules, MT software started using existing language as a template for its versions. Give your software a big enough database of language and some nifty statistical algorithms and you could get reliable results, or so it was thought.

Actually, this works fairly well. The much maligned rule-based Babelfish has been taken over by the stats-based Google Translate, which has the bonus of a huge database of UN language to work from. Feed in the right kind of language to Google Translate and you do get passable results. On the other hand, feed in certain kinds of language, translate backwards and forwards a bit and Rick Astley’s “Never gonna give you up” becomes “You’ll never leave,” which is funny and ironic but wrong.

So, for the moment, it looks like MT isn’t quite good enough on its own, especially since it can only ever be as smart as the data humans give it. In fact, it is this partnership between human common sense and machine processing that seems to be the way forward. Now, instead of MT taking over, it has been coupled with its somewhat more respectable cousin Computer Aided Translation (CAT) into flexible translation toolkits. Today’s translators, it seems, are happy to take whatever help they can get and are smart enough to be able to use a variety of tools at once.

More importantly, this partnership suggests that translators’ jobs are safe, at least in the near future. Rather than Machine vs. Human being a battle to the death, it has become more of a growing partnership.

Jonathan Downie