Deepening wisdom or compromised principles?-the hybridization of statistical and symbolic MT systems

摘要

Departing from your principles is hard. The frequency and volume with which the principles of statistical and symbolic MT were repeated during the Statistics Wars means impressive-sounding arguments will continue for a while. But it is perhaps more instructive to consider what happens in practice. Baldly stated, if you want to build a non-toy MT system-a system with more than approximately 5,000 lexical items-that handles previously unseen input robustly, you always end up including some statistics-based modules or knowledge. And if you want a reasonable level of output quality, you always end up including some symbolic/linguistically motivated knowledge or modularization.