Treatment of Meaning in MT Systems

Fully-automated, high-quality MT systems can be developed only if a realistic way to treat meaning in natural languages is found. This paper analyzes a set of positions on the treatment of meaning in machine translation, which we consider the crucial point of difference among MT approaches. We present and criticize a chain of claims that is skeptical with respect to the prospects for meaning-based machine translation. Viewed as a meta-stance, this chain of opinions argues — wrongly, we suggest — against the feasibility, the need for and even the possibility of meaning analysis as the basis for machine translation technology. We set out to demonstrate that these viewpoints, however popular, are flawed. In fact, we suggest that the methodological debate between the interlingua and the transfer approaches to MT is, in fact, misplaced. The meaning-based approach to MT can in principle be explored using either of these architectures or — even more promisingly — some combination of them.