Slashdot Mirror

← Back to Stories (view on slashdot.org)

DARPA Project Babylon: Universal Translator

Posted by michael on from the unleash-the-kitty dept.

silance writes "Take a look at this project from DARPA (Defense Advanced Research Projects Agency)! This time the boys are trying to hammer out a portable, two-way, real-time, multi-lingual audible speech translator proposed to be run on everything from PDA's to wearable military hardware to workstations (to replace their PRE-EXISTING ONE-WAY real-time hand-held audible translators, of course!). The site contains descriptions of technical approaches, a technical milestones timeline, and a nifty Power Point presentation for the executive-types ;) They should give William Shatner a beta model out of pure respect... Here's a link to Google's cached HTML version of the Power Point presentation just in case. (P.S. - get a load of that logo at the bottom of the page!)"

1 of 324 comments (clear)

  1. /.ed already... by Anonymous Coward · · Score: -1, Redundant

    Babylon (this program has moved to IAO)
    LTC. James D. Bass

    Technical Approach:
    The Babylon technology holds great promise for improving communication and information collection for operations in foreign countries by providing rapid two-way speech-to-speech translation capabilities. A typical proposed system today uses a voice/speech recognizer, machine translation and text-to-speech (TTS) output, all individual components developed by different groups with each component contributing errors in both function and integration domains. Creating a truly effective two-way speech translator has proven to be more complicated than simple integration of these components. Research is required in developing an accurate, reliable and modifiable method of language translation. By using a knowledge-based translation system called Interlingua, meaning is applied to phrases being translated making it more accurate than straight word-to-word translation. Methods such as this will focus on exploring spoken language interpretation (understanding meaning and then paraphrasing in second language), which is done by most skilled human translators, rather than spoken language translation (word-to-word translation). The program strategy for Babylon is to organize the work into several primary task areas plus a data collection and evaluation effort:

    1. Tech-base: Core research developing better parsing techniques, semantic knowledge representation, morphological engines, and Interlingua-based models. Tech-base research will also include research efforts on the code optimization, component integration, validation, data repository for resource collection and sharing, and other aspects of speech translation. Performers for this effort will be selected by a conventional BAA process.

    2. DARPA 1+1: The 1+1 technology presents two radically different approaches to creating a highly constrained translation system supporting communication in both directions. The teams will compete with the best (fully functional) system being selected for limited production. 1+1 represents a stopgap technology to meet the needs for emergency translation support while the more stable and robust two-way systems are being developed (see Babylon Teams).

    3. Babylon Teams: These are the primary multidisciplinary teams that will build two-way prototypes systems to create and demonstrate new speech-to-speech translation technologies developed over the past 12-24 months. These teams will attack the most difficult factors causing poor speech translation performance. These teams will not engage in simple evolutionary advancement in speech translation performance. The task goal is to produce ten working two-way prototypes from each of four teams by the end of 18-months. The languages that will be translated are Farsi, Dari, Arabic, Pashto, Mandarin, and Uzbeki. The teams must have a diversity of members to overcome such ambitious performance goal objectives. The success of the prototypes will determine the support of the development teams after the 18-month mark. The funding for the surviving teams must support not only continued refinement and technology insertions, but also obtain perceived user acceptance of the developed technology from the user evaluation teams. Because of the compressed schedule and extreme risk, these teams will be chosen from the most elite organizations in the field and supported via competed contracts already in existence from our agent teams.

    4. Other Tasks: The additional efforts represent formal data collection efforts, repository activities, and leveraging of other approved research programs. These activities will be monitored by the PM and details of progress will be presented at the Babylon annual reviews for DIRO approval/endorsement. MITRE was selected as the primary performer for data collection and to act as ombudsmen between the user evaluation teams and the PI's. The Language Data Consortium (LDC) will perform data transcription and alignment for the development, training, and evaluation data sets.