A Pendulum Swung Too Far

Today's students might be faced with a very different set of challenges from those of the 1990s in the not-too-distant future. What should they do when most of the low hanging fruit has been pretty much picked over? In the particular case of Machine Translation, the revival of statistical approaches (e.g., Brown et al. (1993)) started out with finite-state methods for pragmatic reasons, but gradually over time, researchers have become more and more receptive to the use of syntax to capture long-distance dependences, especially when there isn't very much parallel corpora, and for language pairs with very different word orders (e.g., translating between a subject-verb-object (SVO) language like English and a verb final language like Japanese). Going forward, we should expect Machine Translation research to make more and more use of richer and richer linguistic representations. So too, there will soon be a day when stress will become important for speech recognition. Since it isn't possible for textbooks in computational linguistics to cover all of these topics, we should work with colleagues in other departments to make sure that students receive an education that is broad enough to prepare them for all possible futures, or at least all probable futures.