The Future of MT in the New Millennium

As we noted at the beginning of this book, MT is having a revival in the US and to a lesser extent in Europe drive – n like so much else by the World Wide Web – and to be somewhat depressed in Japan compared with ten years ago. Ways forward from here seem to come down to either a better attempt at matching the existing technology to market needs, which are real and expanding, and in improving the basic technology by getting above the 65–70% level that customers, of the best systems on a good day, will now accept. What does this latter mean in concrete terms and how does it relate to the arguments about principle that have raged through MT’s past, and have represented in most chapters of this book? There are some things everyone, from every research school can agree with, whether one still advocates knowledge-based methods, is a believer in linguistic purity, in interlinguas, in statistical methods or (more sceptically) in no more than vastly improving editing tools and interfaces for users. There are now much better decomposable modules for linguistic analysis available: part of speech taggers, automatically and semi-automatically derived lexicons and concordances, effective grammars and parsers far closer to corpora than before. Yet their effect is not apparent in systems in the market – researchers who help build marketable systems still often throw away all their theoretical beliefs and their successes when going to market, as if they themselves do not believe in effectiveness of their own published work. We noted at the beginning of the book that Martin Kay once argued that, even if all the problems of syntax, morphology, computational semantics had been individually solved, it might not improve MT. One may not be quite sure what he meant but it is an important thought; if one asks what he thought might still be missing one might list research in