Improved Discriminative ITG Alignment using Hierarchical Phrase Pairs and Semi-supervised Training
Shujie Liu, Chi-Ho Li, Ming Zhou · 2010
While ITG has many desirable properties for word alignment, it still suffers from the limitation of one-to-one matching. While existing approaches relax this li-mitation using phrase pairs, we propose a ITG formalism, which even handles units of non-contiguous words, using both simple and hierarchical phrase pairs. We also propose a parameter estimation me-thod, which combines the merits of both supervised and unsupervised learning, for the ITG formalism. The ITG align-ment system achieves significant im-provement in both word alignment quali-ty and translation performance. 1