By Amaani Lyle
American Forces Press Service
WASHINGTON, Jan. 9, 2014 – Defense Advanced Research Projects Agency scientists will build on language processing technologies with improved speed and accuracy –- offering an advantage to analysts in a variety of military and non-military scenarios, a program manager said today at the DARPA Congressional Tech Showcase here.
Dr. Bonnie Dorr, DARPA Human Language Technologies, demonstrated Raytheon BBN Technologies’ “Byblos,” one of several speech recognition systems that represent the state of the art in trainable, large-vocabulary, speaker-independent speech recognition.
“What’s of interest here is gleaning information from the huge volumes that come through to us in foreign languages,” Dorr said. “So it’s really [addressing] the big data problem.”
The natural language processing technologies can locate, identify, and organize information from a variety of sources and in at least 15 languages.
English-speaking analysts once saddled with sifting through a barrage of information in foreign languages can now use real-time filters to pinpoint information in audio and video broadcasts.
“The system goes into the video, pulls out the audio, separates it into sentences, renders it as text, and translates it into English so that the human, who speaks only English, can then read what this Arabic broadcast news is about,” Dorr explained.
She added that despite a three-minute delay from a live broadcast, the real-time feed of identifying and aggregating individual pieces of information from raw data is remarkable.
The next chapter, Dorr said, involves developing what the translation output does to enhance information analytics.
“In the future, we want to be able to read through language to meaning because people don’t always explicitly state all the assumptions that are underlying what they’re saying,” she said.