German has a richer system of inflectional morphology than English, which causes problems for current approaches to statistical word alignment. Using Giza++ as a reference implementation of the IBM Model 1, an HMM-based alignment and IBM Model 4, we measure the impact of normalizing inflectional morphology on German-English statistical word alignment. We demonstrate that normalizing inflectional morphology improves the perplexity of models and reduces alignment errors.

PDF file

Publisher  Association for Machine Translation in the Americas
Copyright Springer-Verlag Berlin Heidelberg 2004

Details