Session 3: Machine Translation

In 1965 the United States Academy of Science cemmissoned a study of he state of the art in Machine Translation, whose findings were published the following year and become popularly known as the ALPAC report. In essence, ALPAC argued that there was insufficient scientific basis in natural langauge processing to perform reliable machine t~anslatinn, and the large expensive computers of the time would make NIT eeonornically infeasible. Both situations have since changed drastically, invaLidating the ALPAC conciusions. In fact, DARPA has played a major role in fostering the development of the NLP scientific infrsstmcture in the post-ALPAC years. IBM, in which the direct-transfer paradigm is still king and translation is viewed as transduction between two character (or word) s t reams--essent ia l ly two encodings of the same message. However, the direct transfer rules are totally learned by statistical analysis of large bi-lingual corpora, rather than laboriously and incompletely hand-coded. A drawback of the statistical approach, of course, is that it carmot guarantee the accuracy of any textual passage being translated, but rather strives to minimize the total number of errors over time.