Philipp Koehn

A college textbook on the newfangled statistical approach to automatic translation between natural languages. In the pure form, the basic idea is to have translator programs train themselves automatically by digesting huge parallel text corpora consisting of many mega-, giga-, or terawords, with little attention paid to linguistic theory. The book gets into the approach's many messy (IMHO) subproblems, such as "alignment", "reordering", and "decoding", all modeled with rigorous but not-very-elegant mathematics. A lot of progress seems to have been made, but I would hope that the ultimate realization of machine translation, if there is ever to be such, will involve a return to the general interlingua idea envisaged in the old-fashioned approaches to the problem. (In other words, the book did not make a convert out of me.) Www.statmt.org.… (mer)