University of Southern California scientists have developed a machine-translation technique that treats translation as a cryptographic challenge rather than a matter of analyzing the statistical characteristics of the same text written in two different languages.
Software written by researchers Sujith Ravi and Kevin Knight calculates the likelihood that a foreign word matches an English word based on the frequency it occurs within the text. The scientists employ another piece of software that assesses the quality of the resulting English to guarantee that the translation makes sense, which consequently supplements the probabilities used in the translation software.
Johns Hopkins University researcher Chris Callison-Burch says that Ravi and Knight's technique holds much promise, but has yet to prove itself. Meanwhile, his team is developing translation software that avoids parallel data, crawling online texts and comparing disparate texts from different dialects.
Ravi and Knight also are investigating how monolingual techniques might help translate unknown ciphers or long-lost languages. Such translation also may help soldiers or aid workers in rapidly responding in nations with unfamiliar languages.
From New Scientist
View Full Article