In this paper, we present a method of avoiding the combinatorial explosion encountered in Factored Models during the construction of translation options caused by the large number of possible combinations of target language lemmas and morpho-syntactic factors. We automatically extract factor templates froma word-aligned annotated bilingual corpus and use them to distinguish which morpho-syntactic factors should be translated separately from lemmas and in doing so avoid the large number of translation options otherwise considered for generation. Besides Phrase-Based SMT, FactoredModels can be applied to SMT via deep syntactic transfer, which is the focus of our work. We therefore include an experimental evaluation of our method for a SMT via deep syntactic transfer system, comparing the baseline standard Factored Model with one that uses factor templates for translating morpho-syntactic factors, resulting in a large increase in BLEU score.
Cite as: Graham, Y., Genabith, J.v. (2010) Factor templates for factored machine translation models. Proc. International Workshop on Spoken Language Translation (IWSLT 2010), 275-282
@inproceedings{graham10_iwslt,
author={Yvette Graham and Josef van Genabith},
title={{Factor templates for factored machine translation models}},
year=2010,
booktitle={Proc. International Workshop on Spoken Language Translation (IWSLT 2010)},
pages={275--282}
}