ABSTRACT
Pat Suppes used to say 20 years ago tha t he would pay attention to AI and natural language processing (NLP) only when it could “do something of book length” . T hat day is now pretty close, though old-fashioned machine translation systems like SYSTRAN [Toma, 1977] met his criterion at a low level many years ago. The present could seem like a rerun (in NLP) of the struggles within machine vision in the late 1970’s: the high-level, top-down paradigm of scene analysis (Guzman, Brady, Waltz, etc.) was crumbling in the face of low-level, bottom-up, arguments posed by Marr. NLP has been told for some time tha t it should be more scientific “like machine vision,” so perhaps the current emergence of connectionist, statistical and associationist techniques in NLP is a form of progress by virtue of tha t fact alone.
