ABSTRACT

LFG-DOT (Way, 2001) has been developed as a novel, hybrid model for Machine Translation (MT) based on LFG-DOP (Bod & Kaplan, 1998). We shall demonstrate that LFG-DOT can cope with translational phenomena which prove problematic for other LFG-based models of translation. Furthermore, LFG-DOT improves the robustness of LFG-MT (Kaplan et al.,1989), and the probability models of LFG-DOT give a `level of correctness' figure to alternative translations. We shall also show that LFG-DOT improves upon the DOT (Poutsma, 1998; 2000) MT systems. DOT1 is not guaranteed to produce the correct translation when this is non-compositional and considerably less probable than the default, compositional alternative. DOT2 solves most of the problems of DOT1 and seems to be able to cope with the translational phenomena which other LFG-based systems find problematic. Nevertheless, DOT2 allows the formation of certain ill-formed structures which are prevented in LFG-DOT models by recourse to monolingual f-structure information.

We shall present four models of translation based on LFG-DOT. The problematic translation data will be presented and discussed, and results will be provided showing how DOT, LFG-MT and LFG-DOT attempt to cope with such phenomena, with differing degrees of success.