Practical spoken language translation using compiled feature structure grammars

Practical work on spoken language translation must pursue two types of efficiency: computational efficiency, and “language engineering” efficiency. This paper describes the design, implementation, and evaluation of the GPL-based framework for spoken language translation that addresses both of these goals. In this framework, computational grammars are written in GPL, an easy-to-use imperative programming language that allows the direct expression of linguistic algorithms in terms of rewrite-grammars with feature structure tests and manipulations. Computational efficiency is achieved with the GPL compiler, which converts GPL grammars into efficient C routines, and with the GPL runtime environment, which provides services for linguistic representations, manipulation, and memory management. An evaluation of an English-Japanese spoken language translation system based on GPL shows that it is linguistically powerful, yet only requires reasonable computational resources.