The University of Maryland's Bonnie Dorr was recently featured in a Slate.com article for her work on the Broad Operational Language Translation (BOLT) program.
Researchers at the Defense Advanced Research Projects Agency (DARPA) have been working on developing a "speech to speech" computer translation system as part of a five-year program named TransTac. TransTac software was intended to listen to a human speaking one language, translate it and then speak in another language. At the conclusion of the program, TransTac reached around eighty percent accuracy.
A method to improve accuracy is statistical machine translation (SMT) that will infer language rules from existing databases of translated texts. SMT seeks to build models that properly align sentences, demonstrate how languages work independently and identify similarities between each one. A metric to score machine translation programs, named BLEU, was developed by IBM. BLEU compares the computer translation to a human translation of the same sentence. However, this is not entirely effective becuase the scores do not hold significant meaning.
Bonnie Dorr, a professor at the University of Maryland College of Computer, Mathematical and Natural Sciences, is the program manager for the BOLT program. Dorr says that DARPA is "very focused on moving beyond statistical models" because as more data is used to create algorithms, it begins to output "diminishing returns. The payoff gets smaller, and you start to plateau with your results even if you increase the volume of training data.”
DARPA continues to try new and innnovative methods to enhance computer translation software, such as semantic tagging.
University of Maryland researchers engaged in work related to the BOLT program include: Doug Oard, Philip Resnik, Jimmy Lin, and Hal Daumé III.
May 11, 2012