Basing life and death decisions on automatic translation is a bad idea

Automatic translation tools are terrible. There’s nothing wrong at all with looking up a word at an electronic dictionary, but letting one translate a sentences always ranges between pathetic and hilarious.

Even the best of translation tools does a terrible job of discerning context. And when each word can have several different translations, and often holds several different meanings, context is everything. You get that wrong, and some of the words wrong, and the translated result is pure gibberish.

Even worse, if the translation program tries to decipher some context and fails, as will often happen, the end result may not appear to be total gibberish at first glance. But we’re still far away from tools at even this level.

And all that is even when you translate between languages which are similar, and have a relatively recent common ancestor.

You can even check for yourself, just for fun. Go to one of the available translation tools on-line, enter a paragraph of text from someplace, and try to translate it first into a different language, and then back. See what a sordid mess you’ll get.

So, to change the subject completely (OK, not really), what new toys are the Americans up to in order to assist their soldiers in Iraq?

The risky business of battle-zone translation could get a technological boost, however, as researchers prepare to test a system that instantly translates spoken conversations to and from English and Iraqi Arabic.

Funded by Darpa, the system would allow troops to communicate in Arabic through a laptop computer equipped with voice recognition and translation software. Troops could speak in English and have their words instantly translated into Iraqi Arabic, “spoken” by a computerized man’s voice. The program also translates Arabic into English.

Automatic translation between Arabic and English. Two languages which are not even close to being remotely similar. At least those people haven’t lost their sense of humour. Though they may just do that after some bad incidents of horrible mistranslation that will cost human lives.

And they’re thinking of translating voice, even, not text. Voice recognition these days is still pretty bad. The best voice recognition programs these days don’t really work under non-ideal conditions, or without a lot of time dedicated to studying each individual voice they’ll be expected to deal with.

So all they have to do here is take a badly working voice recognition, drive the output from that one through a badly working translation engine, and then synthesize the output of that one to voice. Sure. No problems at all, none whatsoever. It’s going to go perfectly smooth. Nothing at all in there that they can’t get working within the year…

The only thing that seems plausible given the time frame is the voice synthesis at the end. And, well, even voice synthesis of text isn’t too hot yet these days. But really, that’s the least of their concerns.

Leave a Reply

You must be logged in to post a comment.