Improving Lexical Alignment Using Hybrid Discriminative and Post-Processing Techniques

Automatic lexical alignment is a vital step for empirical machine translation, and although good results can be obtained with existent models (e.g. Giza++), more precise alignment is still needed for successfully handling complex constructions such as multiword expressions. In this paper we propose an approach for lexical alignment combining statistical and linguistic information. We describe the development of a baseline discriminative aligner and a set of language dependent post-processing functions that allow the inclusion of shallow linguistic knowledge. The post-processing functions were designed to significantly improve word alignment mainly on verb-particle constructs both over our baseline and over Giza++.