Slashdot Mirror
DARPA Starts Ultimate Language Translation Project
An anonymous reader writes "Defense Advanced Research Projects Agency (DARPA) has launched the ultimate speech translation engine project that would be capable of real-time interpretation of television and radio programs as well as printed or online textual information in order to be summarized, abstracted, and presented to human analysts emphasizing points of particular interest." If combined with the tower of babel project we discussed earlier, it could only lead to awesomeness.
As if the government doesn't already have legions of translators at the ready. Military linguists are trained at Defense Language Institute at the Presidio of Monterey. I studied Chinese there while serving in the Navy, and while most of my fellow enlisted servicemen were likewise studying languages of some clear strategic value, there are also courses in various other languages for officer exchange programs, as well as the occasional course in something really exotic. Combined with the simple possibility of the government paying a native speaker to work for them, this means that the government already has the language skills it needs even without a whizbang translation machine.
All I really want is a free online translator for web pages (ala Babelfish and Google) that aren't terrible at it. Seriously, the quality of Babelfish translations has stayed constant since it came on the scene in the late 90s, even though machine translation in general has made some rather significant advances. I don't really use them enough to justify plopping down $500 on the professional packages, but the current systems are just terrible.
I read the internet for the articles.
Seriously though, I just don't believe it. I've worked on a number of DARPA robot projects, and have heard a lot of their babble. They claim to be funding all these fantastic ideas, but none of them ever work except in a limited capacity. The robot projects I worked on were very lame in that DARPA created these really specific environments for the robots that were light years away from what they were saying they were really going to do. All of the Universities involved failed to accomplish even the simplest tasks. So my experiance with them is that they talk a big talk, and no one ever goes back to check "hey did you really ever do that ?" Now granted, some of their work is supposed to be high risk, but they never emphasize which projects are expected to have a high failure rate. Largely because they don't care. It's really all about funding your academic buddies or whoever is going to be able to scratch you back in some way. It is very much an old boys network, with an emphasise on PR and not much about real science. Much like the MIT media lab. (Just thought I'd get another jab in there....)
One of the problems with using humans is that they are expensive--the other is that they become bored easily. It isn't like the defense establishment isn't using human translators, the NSA is the largest employer of translators in the world. They use humans in every listening post out there, but for the same reasons that humans make lousy airport security sceeners, they make poor translators AND intelligence analylists. This isn't saying that machine translators are a panacea, but they can solve a small section of the problem that we have been trying to solve with a very human capital intensive solution for years now.
This project, along with CMU's Tower of Babel, certainly get props in the coolness category, but the practicality is still lacking. I believe DARPA is barking up the wrong tree for now, or at least biting off more than they can chew.
Speech Recognition is the hardest problem to tackle on the path to recognition, and MUST be addressed before there is a viable real-time (or even delayed) translation engine. Currently, even the best speech recognition software can achieve at best ~80% accuracy when faced with a large vocabulary with no limits on speakers/dialects, and this level of accuracy is typically not achieved in real-time. While this 80% level is actually pretty good when transcribing to text (since the reader can typically decipher what the computer meant), it's downright awful if trying to translate the resulting text to another language.
For example, if I say "I like ice cream" into voice recognition software and 'hears' "I like, I scream", the reader might understand what this means, particularly if they say it in context and aloud. However, let's say we translate each sentence into Spanish ("Tengo gusto del helado" and "Tengo gusto, yo grito" respectively, according to Babel Fish), and the speaker would be completely lost as the out of context phrases don't sound anything alike. In a natural language translation, even under relatively accurate recognition scenarios, would be frought with misunderstandings.
Once speech recognition is tackled, it's just a matter of translation then voice synthesis. Fortunately these problems aren't nearly as difficult, and current solutions would suffice (with the only pitfall being poor grammer in the destination language, and a robotic sounding voice).
Crack - Free with every butt and set of boobs
I attend one of the many Universities where DOD research is currently being conducted. Portions of our graduate student body and faculty are working on the powered armor concept in conjunction with UC-Berkley (they're doing the frame and kinematics, we're doing the control theory/system and power supply). We're actually making quite a bit of progress in the field of alternative batteries (the current iteration is a peroxide-fueled hydraulic hybrid-type system widget) and mechanical interface control theory application. So, while God knows we won't see cap' troopers in 'suits any time soon, we are at least progressing towards that end while developing widely applicable technologies along the way (this is, if I may remind you, the way many technologies we love dearly were spun off from the space program et. al).