Statement — Jason Eisner

Much of my research in computational linguistics has been bound up with two emerging strategies of the 1990’s: statistical optimality and phonological optimality. Statistics properly used can translate and enrich linguistic theories—rather than replacing them—and can connect those theories to processing models. That is, the statistical worldview need not be tied to simplistic engineering tricks: it is a plausible framework for rethinking linguistic competence, variation, and acquisition (see e.g. Abney 1996). Statistical techniques emerged in the computational community as a way of coping with real-world, ambiguity-riddled, sometimes ungrammatical language. I have worked to incorporate more linguistic sophistication into the statistical models—obtaining both linguistic and computational payoffs. Optimality Theory, by contrast, has developed in the world of pure linguistics. It has been used primarily as a framework for thinking about phonologies, but has some computational trappings. I have worked to add more computational sophistication to OT—proposing constraints on constraints, a representational formalism, and an efficient generation algorithm that can handle infinite candidate sets. Again, the benefits are interdisciplinary: well-specified theories are more congenial for both linguists and computers.