DARPA: TRANSLATE THIS

It’s not Kirk and Spock’s uni­ver­sal trans­la­tor. Not quite. But the Pentagon is look­ing to for researchers to build a soft­ware set “with the goal of elim­i­nat­ing the need for lin­guists and ana­lysts and auto­mat­i­cally pro­vid­ing rel­e­vant, dis­tilled, action­able infor­ma­tion.“
Global Autonomous Language Exploitation, or GALE, is a project of — who else? — Defense Department mad sci­ence divi­sion Darpa. And the idea, accord­ing to Darpa’s call for pro­pos­als, is to “develop and apply com­puter soft­ware tech­nolo­gies to absorb, ana­lyze and inter­pret huge vol­umes of speech and text in mul­ti­ple lan­guages.“
The result won’t nec­es­sar­ily be a “nat­ural lan­guage” dia­logue between man and inter­pret­ing machine. But, if GALE works as planned, it will deliver “con­sol­i­dated infor­ma­tion in easy-​​to-​​understand forms to mil­i­tary per­son­nel and mono­lin­gual English-​​speaking ana­lysts in response to direct or implicit requests.“
The American mil­i­tary is still strug­gling to fill its ranks with Arabic speak­ers, three-​​and-​​a-​​half years after 9/​11. Language train­ing for enlisted men and junior offi­cers is min­i­mal. And the tech­no­log­i­cal solu­tions to the prob­lem — like the hand-​​held Phraselator and Interact sys­tems — really only work for the most mono­syl­labic of con­ver­sa­tions.
What Darpa wants instead are a trio of soft­ware tools for sol­diers and spooks:

A tran­scrip­tion engine that pro­duces English tran­scripts [from for­eign speech] with 95% accu­racy
A trans­la­tion engine pro­duc­ing English text [from for­eign prose] with 95% accu­racy
A dis­til­la­tion engine able to fill knowl­edge bases with key facts and to deliver use­ful infor­ma­tion as pro­fi­ciently as humans can. 

And Darpa’s not talk­ing about just trans­lat­ing a cou­ple of news­pa­pers in Baghdad. GALE researchers have to be ready to have their algo­rithms inter­pret “all the fol­low­ing types:”

Broadcast news (radio, tele­vi­sion)
Talk shows (stu­dio, call-​​in)
Newswire
Newsgroups
Weblogs
Telephone con­ver­sa­tions
The source lan­guages will be English, Chinese and Arabic plus sur­prise lan­guages to be announced later.

March 29th, 2005 | Cloak and Dagger | Comments Off Both comments and pings are currently closed.